Real Time Location using Ultra-Wideband (UWB)

Ultra-Wideband ranging modules

I’ve been looking for a system that can provide fast & accurate position measurement within a defined area; this is generally known as an RTLS (Real Time Location System).

My previous experiments have involved optical measurements, which have good accuracy, but are constrained by line-of-sight and range issues. So why not use wireless, and measure the time it takes for a radio pulse to travel from transmitter to receiver? Given the speed of light is roughly 300 mm (or 1 foot) per nanosecond, it may seem impossible to get an accurate position that way, but Decawave claim that their DW1000 time-of-flight chip gives measurements around 100 mm (4 inch) accuracy, using an Ultra Wideband (UWB) radio.

The DWM1000 module is bottom-left in the photo above, and consists of the DW1000 chip with a crystal reference, voltage regulator and on-board antenna. The other two boards use the same wireless chip, in different form-factors, but with additional embedded CPUs. I wanted to experiment with the DW1000 chip in a wide range of scenarios, and fully understand its low-level operation, so decided to use the simpler DWM1000 module, driven by my own Python code.

Decawave provide a large amount of documentation on their chip, and several software packages, but these are quite large: for example, their DecaRanging application has over 20,000 lines of C and C++ source code, which is a bit intimidating if you’re a newcomer to sub-nanosecond radio timing.

So I’ve created a Python test framework from scratch; at under 1000 lines of code, it can’t compete with the Decawave packages, but hopefully it’ll give an insight as to how the chip works, and enable you to experiment with this interesting technology.

Ultra Wideband

You may not have seen this RF technology before, but it has been around a while; the IEEE standard 802.15.4a is dated 2007. Just because it is part of the 802.15.4 family, you may think it is similar to Zigbee or 6LoWPAN, but that is not true. The RF operation is completely different: instead of transmitting on a single frequency, it covers a wide spectrum. This makes it much more resistant to single-frequency interferers, but of course raises the prospect that the UWB transmitter could interfere with other radio systems nearby.

For this reason, there are some quite complex rules about which frequencies can be used, the permissible power-levels, and the transmit repetition-rate. So it is possible that the RF emissions generated by my software are not permitted in your locality. If in doubt, consult a suitably-qualified RF engineer before doing any UWB testing.

Tags and Anchors

Real Time Location System

The standard Location System consists of several ‘anchors’, which are positioned at known locations, and ‘tags’ which move around a defined area; they determine their position by measuring the transit-time of the signals from several anchors.

This scheme works well for, say, locating people within a shopping mall; their mobile phones can be the tags, displaying the location within that mall – and non-coincidentally, the Apple iPhone 11 does have an UWB capability.

An alternative scenario is where the tags are fitted to vehicles in a warehouse, allowing a management system to track their whereabouts. There are two ways this can be achieved; either a tag just transmits a simple beacon message, and the anchors share their time-measurements to establish its position, or alternatively the tag measures its distance from the nearest anchors, and transmits the result for forwarding to the management system.

Implicitly, a tag is a battery powered device that only transmits occasionally, but in reality there are many other ways to configure a location network, depending on the overall requirements.

This flexibility comes from the fact that the ranging messages can also carry data (up to 127 bytes as standard), so there are numerous ways the RTLS can be structured. In this first post, I’m ignoring all that complexity, and just focusing on the distance measurement between two systems, which could be tags, anchors, or anything else you decide.


Ranging is the process whereby two UWB radio systems can measure the distance between themselves. Simplistically, one might think that it is just necessary for the transmitter to note the time of a message transmission, and the receiver to note the time it is received: subtract the two and you get a time-difference, which is directly proportional to the distance between them.

However, it isn’t quite that simple, because:

  1. The measurement has to be very accurate; a radio wave travels at around 300,000,000 metres per second (1 foot per nanosecond) so to achieve any degree of accuracy, we need a time measurement in picoseconds (10-12 seconds).
  2. In this method, the time-clocks of the transmitter and receiver have to be very accurately synchronised, and that isn’t easy.
  3. To keep hardware costs down, each of the units will have an inexpensive quartz crystal as the timing reference, and we have to accommodate variations in the crystal frequency due to its tolerance, temperature drift, etc.
  4. The radio wave that arrives at the receiver won’t be an accurate copy of what is transmitted; there will be distortions due to the radio circuitry, and reflections from nearby objects.

Fortunately, these problems can be addressed by using a technique called ‘Asymmetric Two Way Ranging’:

  1. Use very fast, high-resolution timers; the sampling clock on the the DW1000 runs at 63.8976 GHz, and feeds a 40-bit counter.
  2. Don’t synchronise the clocks in the transmitter and receiver; let them just free-run.
  3. Measure the difference in crystal frequency, and compensate for it.
  4. Use Ultra-Wideband (UWB) which is more resilient than conventional radio systems.
Asymetric Two-Way Ranging

In the diagram above, there are 3 messages passing between two units; unit A transmits messages 1 and 3, unit B sends message 2. Each unit records a timestamp when the message was sent or received, so we have a total of 6 timestamps, from which to determine the transit time, and hence the distance.

Simplistically the transit time can be measured by comparing Rx2 – Tx1 with Tx2 – Rx1, but you’ll see that the time clocks for units A and B are running at different speeds. In reality they’ll only differ by a few parts per million (the difference has been greatly magnified for the illustration) but a small difference creates in a large position error, so we need a method to compensate for it. This is done by getting the two units to make the same measurement, and comparing the result; the obvious candidate is the time between the two transmissions (Tx3 – Tx1) and the time between the two receptions (Rx3 – Rx1). These should be equal, so the ratio of the times will be the ratio of their clock frequencies.

The final formula for the transit time (taken from Decwave’s APS013 application note) is:

rnd1 = Rx2 - Tx1 # Round-trip 1 to 2
rep1 = Tx2 - Rx1 # Reply time 1 to 2
rnd2 = Rx3 - Tx2 # Round-trip 2 to 3
rep2 = Tx3 - Rx2 # Reply time 2 to 3 
time = (rnd1 × rnd2 - rep1 * rep2) / (rnd1 + rnd2 + rep1 + rep2) 


DWM1000 module on carrier board

The Decawave DW1000 chip can be purchased from electronic distributors, but unless you’re into microwave PCB design, you’ll want to buy a pre-packaged module. The simplest of these is the DWM1000, which includes the necessary power circuitry and ceramic chip antenna. It has no on-board CPU, so is driven by an external processor over a 4-wire Serial Peripheral Interface (SPI).

You could solder wires direct to the package, but I used an adaptor board that brings out the connections to a breadboard-friendly 0.1″ pitch. The adaptor is the “DWM1000 Breakout-01”, available from OSH Park.

Aside from the SPI interface (CLK, MISO, MOSI and CS) you only have to provide 3.3V power and ground, though I also connected reset (RST) and interrupt (IRQ) signals. Reset is very useful to clear down the chip before programming, and the interrupt saves the chip from frenetic polling during transmission or reception (which can disrupt the RF section of some chips).

Which CPU to drive the module? Any microcontroller would do, but I’d like to control the modules using a single Python program on a PC; this is much easier than updating multiple copies of the software on different CPUs. So I’m attaching each module to a Raspberry Pi, to act as a relatively dumb network-to-SPI converter; I can then send streams of SPI commands from the PC program over a WiFi network to 2 or more UWB modules, without having to reprogram their CPUs.

Pi ZeroW and DWM1000 module

SPI port 0 or 1 can be used on the RPi, so long as it is enabled in /boot/config/txt. The pin numbers are:

# Connector pin numbers:
#       SPI0        SPI1
# GND   25          34
# CS    24 (CE0)    36 (CE2)
# MOSI  19          38
# MISO  21          35
# CLK   23          40
# IRQ   18          32
# RESET 22 (BCM25)  37 (BCM26)
# NRST  16 (BCM23)  31 (BCM6) 
# 3.3V  17

I have provided a positive-going reset signal (RESET) and negative-going (NRST). This is because my early hardware had a transistor inverter in the reset line, so needed a positive-going signal. If you are connecting the RPi pin direct to the module, use the NRST signal. [And in case you’re wondering, I realise that the RPi mustn’t drive the module reset line high; my software does not do this, it drives the line low to reset, or lets it float.]

A useful extra is to fit an LED indicator to the module interrupt line (with a current-limiting resistor of a few hundred ohms to ground). This will flash in a recognisable pattern when ranging is working correctly, which is very useful when testing the module’s operational limits.

The module with a Pi ZeroW and USB power pack is a neat package; I had some concerns about taking 3.3V power from the RPi, due to possible electrical noise issues, but it seems to work fine, providing you keep the cable to the module short – I’d suggest a maximum of 100 mm (4 inches) if you want to avoid problems.

Raspberry Pi Software

Ranging test system

We need a simple way of sending commands to the network nodes from the PC; since each command is a small data block, and we have to wait for the command to be executed before sending the next, the logical choice is User Datagram Protocol (UDP). This is an ‘unreliable’ protocol, as it has no mechanism for retrying any lost transmissions, or eliminating any duplicates, so I’ve added a lightweight client-server error-handling layer. Each data block (‘datagram’ in UDP parlance) has a 1-byte sequence number, a 1-byte length, and a payload of up to 255 bytes. The client (PC system) increments the sequence number with each new transmission; the server (Raspberry Pi) checks whether that sequence number has already been received. If so, the data is ignored, and the previous response is just resent; if not, the data is sent to the UWB module over the SPI interface, and the response is returned to the client.

Network Server

The code on the Raspberry Pi has been kept simple; it is single-threaded by using the ‘select’ mechanism to poll the socket for incoming data, with a timeout that allows the interrupt indicator to be polled:

import socket, select

# Simple UDP server
class Server(object):
    def __init__(self):
        self.rxdata, self.txdata = [], []
        self.sock = self.addr = None

    # Open socket
    def open(self, portnum):
        self.sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
        self.sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
        self.sock.bind(('', portnum))
        return self.sock

    # Receive incoming data with timeout
    def recv(self, maxlen=MAXDATA, timeout=SOCK_TIMEOUT):
        rxdata = []
        socks = [self.sock]
        rd, wr, ex =, [], [], timeout)
        for s in rd:
            rxdata, self.addr = s.recvfrom(maxlen)
        return rxdata

    # Receive incoming request, return iterator for data blocks
    # If sequence number is unchanged, resend last transmission
    def receive(self):
        self.rxdata = bytearray(self.recv(MAXDATA))
        if len(self.rxdata) > SEQLEN:
            if len(self.txdata)>SEQLEN and self.rxdata[0]==self.txdata[0]:
                self.txdata = [self.rxdata[0]]
                rxd = self.rxdata[SEQLEN-1:]
                while len(rxd)>1 and len(rxd)>rxd[0]:
                    n = rxd[0] + 1
                    rxd = rxd[n:]

    # Add response data to list
    def send(self, data):
        self.txdata += [len(data)] + data

    # Transmit responses
    def xmit(self, txdata, suffix=''):
        if self.addr and len(txdata)>SEQLEN:
            txd = bytearray(txdata)
            self.sock.sendto(txd, self.addr)

SPI interface

This consists of a clock line, data from the RPi to the module (MOSI: Master Out Slave In), data from the module (MISO: Master In Slave Out) and a Chip Select (CS) line that frames each transmission.

For protocol details, see the Decawave DW1000 datasheet. The most significant bit of the first byte indicates a read or write cycle; a read cycle returns one or more garbage bytes (depending on the addressing mode) followed by the actual data; my software returns all of these bytes back to the PC. A write-cycle returns no useful data (it is generally all-ones) but this is still passed back to the PC, as an acknowledgement that the write command has been received.

import spidev, RPi.GPIO as GPIO

# Open SPI interface
spif = 0,0
spi = spidev.SpiDev()*spif)
spi.max_speed_hz = 2000000
spi.mode = 0

# Set up board I/O
rst_pin, irq_pin = 22, 18
GPIO.setup(rst_pin, GPIO.OUT)
GPIO.setup(irq_pin, GPIO.IN)
GPIO.add_event_detect(irq_pin, GPIO.RISING, callback=irq_handler)

The clock speed of 2 MHz is well within the specified limits for the module. The interrupt (IRQ) line is high when asserted, so positive-edge-detection is used; the callback just sets a global variable that is polled in the main loop.

Running code on startup

It is convenient for the SPI server code to automatically run when the RPi boots; there are various ways to do this, which are beyond the scope of this blog. I used systemd as follows:

sudo systemctl edit --force --full spi_server.service

# Add the following to spi_server.service..
   Description=SPI server

# Enable the service using:
sudo systemctl enable spi_server.service
sudo systemctl start spi_server.service  # ..or 'stop' to stop it

# To check if service is running..
systemctl status spi_server

Main Program

This Python program ( runs on a PC, feeding command strings over the network to the Raspberry Pi UDP-to-SPI adaptors.

Device Initialisation

The bulk of the main program is involved in device initialisation, as the DW1000 has a remarkably large number of registers – my software defines 106, and that isn’t all of them. To add to the complexity, they vary in size between 1 and 14 bytes, have multiple bitfields within them, and are accessed by a multi-level addressing scheme.

By any measure, this is a complex chip, and is a very easy for the software to degenerate into endless sequences of ANDing SHIFTing and ORing to insert new data into a register. To avoid this, the C language has bitfields, and the equivalent in Python is ‘ctypes’, indeed this library was created to allow Python to access DLLs written in C.

I’ve used ctypes in a novel way to give a clean way of reading & writing one or more fields of a register, without any cumbersome logic operations.

To give a simple example, DW1000 register 0 is 32 bits wide, containing a 4-bit revision number in the least significant bits, then a 4-bit version, 8-bit model, and a 16-bit tag number.

I have defined this as:

DEV_ID = 0x0, 4, None, (("REV", U32, 4), ("VER",   U32, 4),
                        ("MODEL",U32, 8), ("RIDTAG",U32,16))

This data is passed to a Python class:

from ctypes import LittleEndianStructure as Structure, Union
from ctypes import c_uint as U32, c_ulonglong as U64

# DW1000 register class
class Reg(object):
    def __init__(self, regdef, val=0):, self.value = regdef, val, self.len, self.sub, self.fields = globals()[regdef]
        class struct(Structure):
            _fields_ = self.fields
        class union(Union):
            _fields_ = [("reg", struct), ("value", U64)]
        self.u = union()
        self.u.value = val
        self.reg = self.u.reg

    # Read register value
    def read(self, spi):
        # [Do SPI read cycle]
        return self

    # Write register value
    def write(self, spi):
        # [Do SPI write cycle]
        return self

# Set a field within a register
    def set(self, field, val):
        if hasattr(self.reg, field):
            setattr(self.reg, field, val)
            print("Unknown attribute: '%s'" % field)
        self.value = self.u.value
        return self

The union overlays an array of bytes on top of the register value; this provides a byte data-stream to be used by the SPI read & write functions.

Instantiating the class gives us a local copy of the DW1000 register, and the ‘read’ method populates the copy with values from the remote register, e.g.

r = Reg('DEV_ID')
print("%X" % r.reg.RIDTAG)

Note that the class methods return ‘self’, so can be chained; for example, here is a read-modify-write cycle that sets the transmit frame length, which is in the bottom 7 bits of the 40-bit register 8:

TX_FCTRL  = 0x8, 5, None,(("TFLEN", U64, 7), ("TFLE", U64, 3), ("R", U64, 3),
                          ("TXBR",  U64, 2), ("TR",   U64, 1), ("TXPRF", U64, 2),
                          ("TXPSR", U64, 2), ("PE",   U64, 2), ("TXBOFFS", U64, 10),
                          ("IFSDELAY",  U64, 8))

Reg('TX_FCTRL').read(spi).set('TFLEN', txlen).write(spi)

‘spi’ in these examples is a class instance that contains the code to read or write the SPI interface; in my test framework, this is actually a network interface that sends the data to a Raspberry Pi, and obtains the response. This is necessary because I have one Python program controlling two (or more) DW1000 modules, so I need a class instance for each SPI interface, giving an IP address and UDP port number, e.g.

# Class for an SPI interface
class Spi(object):
    def __init__(self, spif, ident='1'):
        self.spif, self.ident = spif, ident
        self.txseq = 0
        self.verbose = self.interrupt = False
        self.sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
        if self.sock:
            print("Connected to %s:%u" % spif[1:])
            print("Can't open socket")

SPIF = "UDP", "", 1401
print("Connecting to %s port %s:%u" % SPIF)
spi = Spi(SPIF)

Before leaving the subject of device initialisation, it is worth mentioning that I’ve used several Python dictionaries (as lookup tables) to simplify the underlying logic: for example, the transmitter analog setting RF_TXCTRL, which depends on the channel number (1 – 7, excluding 6)

CHAN_RF_TXCTRL  = {1:0x5c40, 2:0x45ca0, 3:0x86cc0, 4:0x45c80, 5:0x1e3fe0, 7:0x1e7de0}
Reg('RF_TXCTRL', CHAN_RF_TXCTRL[chan]).write(self.spi)

The register class is instantiated using a value from the dictionary, then that value is written to the hardware.

There is a drawback to my approach; the maximum size of any register is limited to 64 bits (c_ulonglong). Fortunately there are only 2 registers longer than this (RX_TIME: 14 bytes, TX_TIME: 10 bytes) and these can be split into sections to come within the 8-byte limit.

Frame Format

A transmitted message (‘frame’) consists of a preamble that the receiver will synchronise to, a start-of-frame delimiter (SFD), and the data payload. The preamble & SFD are automatically inserted by the DW1000, and are used by the timing logic to produce an accurate timestamp, so generally the recommended values will be used. The data payload, however, can be anything; if you want to inter-operate with other UWB 802.15.4 devices it can be a maximum of 127 bytes and must have a standardised header; if not, it can be any format up to 1023 bytes.

Normally, the payload would be used to convey timing information from a tag to an anchor, but in my case the main Python program has visibility of all data through the Wifi network, so I don’t need to send any data across UWB. Arbitrarily, I chose to send the data of an 802.15.4 ‘blink’, which is a very short message containing a 1-byte prefix, 1-byte sequence number and 8-byte address.

# Blink frame with IEEE EUI-64 tag ID
BLINK_MSG=(('framectrl',   U8), 
           ('seqnum',      U8),
           ('tagid',       U64))

This is instantiated using a Frame class that is similar to the Reg class described above, allowing us to refer to the fields individually, or collectively as a stream of bytes.

blink1 = Frame(BLINK_MSG)
blink1.values.framectrl = BLINK_FRAME_CTRL
blink1.values.tagid = 0x0101010101010101
txdata =


Having done all the hard work initialising the chip, transmission is just a question of setting the frame data & length, then setting a single bit.

dw1 = DW1000(spi1)

The timing-specific information is handled automatically, so the precise time of the transmission (specifically, the timing of the SFD) can be determined by a single function call:

    # Get Tx timestamp
    def tx_time(self):
        return Reg('TX_TIME1').read(self.spi).reg.TX_STAMP

You can set the hardware to generate an interrupt (IRQ signal) when transmission is complete, but I haven’t found this necessary.


To receive a frame, the preamble, SFD and data must be decoded; the data must pass a CRC check, and the address must match the filtering criteria if these are enabled. Success or failure is signalled by various bits in the SYS_STATUS register; these bits can also be used to signal an interrupt, if enabled in the SYS_MASK register. In my code, the following signals are enabled as interrupts:

  • RXPHE: phy header error
  • RXFCG: receiver frame check good
  • RXFCE: receiver frame check eror
  • RXRFSL: receiver frame sync loss
  • RXRFTO: receiver frame wait timeout
  • RXSFDTO: receive SFD timeout
  • AFFREJ: automatic frame filtering reject

The most important signals are RXFCG / RXFCE to signal a good frame, or an error condition. The Decawave code has complex error handling, to tackle some ways in which the chip might lock up, and stop responding. Since we have one master program controlling both transmission and reception, we can adopt a simplistic approach to error handling, and if a chip fails to receive several consecutive transmissions, it is reset and re-initialised.

Assuming reception is successful, the data and timestamp can be read out; in our case, we’re only really interested in the time:

    # In DW1000 class..
    def get_rxdata(self):
        rxdata = []
        if self.check_interrupt():
            status = Reg('SYS_STATUS').read(self.spi)
            if status.reg.RXDFR:
                rxdata = self.rx_data()
        return rxdata

    # Get Rx timestamp
    def rx_time(self):
        return Reg('RX_TIME1').read(self.spi).reg.RX_STAMP
rxdata = dw1.get_rxdata()
dt1 = dw1.rx_time() - dw1.tx_time()
dt2 = dw2.tx_time() - dw2.rx_time()

Running the test

The Python source files are on Github, they are:

Main PC program:

  • main program to run the test
  • classes describing the UWB chip internals
  • SPI-over-UDP interface

Raspberry Pi:


The files are compatible with Python 2.7 or 3.x

I didn’t get around to providing a neat UI on the main program, so at the top of you have to enter the IP addresses of the two RPi units, e.g.

# Specify SPI interfaces:
#   "UDP", "<IP_ADDR>", <PORT_NUM>
SPIF1       = "UDP", "", 1401
SPIF2       = "UDP", "", 1401

There is an optional verbose ‘-v’ command-line flag that enables the display of all the incoming & outgoing data. This includes a modulo-10-second timestamp with 1 millisecond resolution, which is useful for tracking down timing problems.

The SPI server running on the RPi units has a similar ‘-v’ option for verbose mode, and an optional port number that also changes the SPI interface, so port 1401 can be SPI0, and 1402 can be SPI1.

To run the test, first make sure the SPI servers are running on the 2 RPis; you can run them from ssh consoles, but I have found that this noticeably degrades the UDP response times on a Pi Zero (see the ‘potential problems’ section below) so once you’ve proved they work, I’d recommend running the code on startup, as described above.

Then run the main program; you should see a stream of ranging results, e.g.

Connected to
Connected to
147.136 156.569
146.991 156.616
147.127 156.602
146.053 156.555
144.017 156.555
146.001 156.588 
..and so on..

This is from 2 units 2 metres (6.5 feet) apart. The first column is the distance in metres for simple 2-message ranging with no measurement of the difference in clock frequencies; the second column is for full asymmetric ranging, that uses a total of 3 messages to compensate for clock inaccuracies.

I’ve said that the units are 2 metres apart, so you’d expect a value of 2 to be displayed, not 147 or 156. The reason for this discrepancy is that the RF circuitry adds a very large time-delay to the signal, that has to be subtracted from the final result. The best way to calculate this compensation value is to measure several known distances, and adjust the multiplier and constant values to produce the right answers.

I haven’t done this calibration process yet, so the un-adjusted result is displayed. The main focus of my current test is to see how repeatable the results are, i.e. how much jitter there is in the position value.

Taking 1000 readings, at distances between 2 and 6 metres, (roughly 6.5 to 20 feet) produces the following histogram of the error between the actual distance (as indicated by the average of all the samples) and the reported distance:

You’ll see the error doesn’t get much greater as the distance increases, i.e. it is not a percentage of the distance measured. This shows that (under good-signal conditions) the main error source is the jitter in the capture and measurement of the incoming wave, as discussed in the Decawave literature, and this is relatively constant irrespective of distance.

The above tests are in good line-of-sight conditions, so to degrade the signal I did a 9 metre (30 foot) range test obstructed by a sizeable brick wall (1900-vintage, not a flimsy modern partition) and a few items of furniture. To my surprise, the error histogram didn’t show very much degradation.

It is also worth noting that despite the convoluted control and measurement process (PC talking to Raspberry Pis), around 5 ranging results are returned per second. Using local CPUs (as in many of the Decawave demonstration systems) will produce a major speed improvement, and a simple rolling average would markedly improve the position accuracy.

Potential problems

Here are some issues you may encounter:

  1. Power supply. In my experience, the most common problem is with the power supply. When receiving or transmitting, the Decawave module takes around 160 milliamps, which is more than some simple 3.3V supplies can handle. Also, the module may appear to work, even if the power supply is completely disconnected; the startup current is sufficiently low that the module can power itself from the I/O lines, and return a valid ID across the SPI interface, even though it is unpowered. Of course it will fail as soon as any real operations start, but the initial SPI response may lead you to look for complex bugs in your code, rather than a simple power supply fault.
  2. IRQ. The software does include a check that the interrupt (IRQ) line is operational, by setting it as an output, then toggling it; see the ‘pulse_irq’ method. If this check fails, there is no point proceeding with the tests.
  3. Missed interrupts. After each transmission, the main software waits for 50 milliseconds to get an interrupt from the receiving unit; if this doesn’t arrive, it polls the receiver’s status register, and if an interrupt is pending (i.e. the message has been received), it reports ‘missed interrupt’. This is harmless, and could be disabled; the reason is that the Raspberry Pi networking stack occasionally adds a long delay to UDP transmissions. To see this in action, try sending flood pings from the RPi to a fast server; I’ve seen round-trip times vary between 3 and 80 msec, even if the WiFi network is very lightly loaded, and the RPi is doing nothing else.
  4. Hardware Reset. Whilst not essential for the testing, if the reset line isn’t connected, things can get confusing, since the chip won’t be cleared down between tests, and the software does assume a ‘clean slate’ at the start of the test.
  5. Status display. If reception fails with error flags set, I display the receiver status; this information is useful in formulating an error-handling strategy.
  6. Bursts of failures. Sometimes when seeing a poor signal, the units stop communicating, and rack up continuous errors. If my software detects 10 such errors, it resets the two units, then carries on as normal. This is not the correct approach; if you look at the Decawave source code, they check the status register to look for potential lock-up conditions, and take appropriate action; they don’t wait for multiple failures.
  7. RF performance. Another weakness of my approach is that it doesn’t represent an accurate simulation of the RF performance of the Decawave chips. Radio circuitry needs decent RF design, and putting a module with an adaptor PCB on a breadboard is not good from an RF perspective. It is credit to the Decwave designers that their module tolerates this approach, producing reasonable results – but if they fall short of expectations, don’t be too surprised; to get the best from the chips and modules, proper hardware design is required.

Ultra-Wideband is a remarkable technology, and I’ve only scratched the surface of what it can do; now see what you can make of it…

Copyright (c) Jeremy P Bentham 2019. Please credit this blog if you use the information or software in it.

ARM GCC Lean: programming and debugging the Nordic NRF52

The nRF52832 is an ARM Cortex M4 chip with an impressive range of peripherals, including an on-chip 2.4 GHz wireless transceiver. Nordic supply a comprehensive SDK with plenty of source-code examples; they are fully compatible with the GCC compiler, but there is little information on how to program and debug a target system using open-source tools such as the GDB debugger, or the OpenOCD JTAG/SWD programmer.

This blog will show you how to compile, program and debug some simple examples using the GNU ARM toolchain; the target board is the NRF52832 Breakout from Sparkfun, and the programming is done via a Nordic development board, or OpenOCD on a Raspberry Pi. Compiling & debugging is with GCC and GDB, running on Windows or Linux.

Source files

All the source files are in an ‘nrf_test’ project on GitHub; if you have Git installed, change to a suitable project directory and enter:

git clone

Alternatively you can download a zipfile from github here. You’ll also need the nRF5 15.3.0 SDK from the Nordic web site. Some directories need to be copied from the SDK to the project’s nrf5_sdk subdirectory; you can save disk space by only copying components, external, integration and modules as shown in the graphic above.

Windows PC hardware

Cortex Debug Connection to a Nordic evaluation board.

The standard programming method advocated by Nordic is to use the Segger JLink adaptor that is incorporated in their evaluation boards, and the Windows nRF Command Line Tools (most notably, the nrfjprog utility) that can be downloaded from their Web site.

Connection between the evaluation board and target system can be a bit tricky; the Sparkfun breakout board has provision for a 10-way Cortex Debug Connector, and adding the 0.05″ pitch header does require reasonable soldering skills. However, when that has been done, a simple ribbon cable can be used to connect the two boards, with no need to change any links or settings from their default values.

One quirk of this arrangement is that the programming adaptor detects the 3.3V power from the target board in order to switch the SWD interface from the on-board nRF52 chip to the external device. This has the unfortunate consequence that if you forget to power up the target board, you’ll be programming the wrong device, which can be confusing.

The JLink adaptor isn’t the only programming option for Windows; you can use a Raspberry Pi with OpenOCD installed…

Raspberry Pi hardware

Raspberry Pi SWD interface (pin 1 is top right in this photo)

In a previous blog, I described the use of OpenOCD on the raspberry Pi; it can be used as a Nordic device programmer, with just 3 wires: ground, clock and data – the reset line isn’t necessary. The breakout board needs a 5 volt supply which could be taken from the RPi, but take care: accidentally connecting a 5V signal to a 3.3V input can cause significant damage.

Rasberry Pi SWD connections
NRF52832 breakout SWD connections

Install OpenOCD as described in the previous blog; I’ve included the RPi and SWD configuration files in the project openocd directory, so for the RPi v2+, run the commands:

cd nrf_test
sudo openocd -f openocd/rpi2.cfg -f openocd/nrf52_swd.cfg

The response should be..

BCM2835 GPIO config: tck = 25, tms = 24, tdi = 23, tdo = 22

Info : Listening on port 6666 for tcl connections
Info : Listening on port 4444 for telnet connections
Info : BCM2835 GPIO JTAG/SWD bitbang driver
Info : JTAG and SWD modes enabled
Info : clock speed 1001 kHz
Info : SWD DPIDR 0x2ba01477
Info : nrf52.cpu: hardware has 6 breakpoints, 4 watchpoints
Info : Listening on port 3333 for gdb connections

The DPIDR value of 0x2BA01477 is correct for the nRF52832 chip; if any other value appears, there is a problem: check the wiring.

Windows development tools

The recommended compiler toolset for the SDK files is gcc-arm-none-eabi, version 7-2018-q2-update, available here. This places the tools in the directory

C:\Program Files (x86)\GNU Tools Arm Embedded\7 2018-q2-update\bin

Check that this directory in included in your search path by opening a command window, and typing

arm-none-eabi-gcc  -v

If not found, close the window, add to the PATH environment variable, and retry.

You will also need to install Windows ‘make’ from here. At the time of writing, the version is 3.81, but I suspect most modern versions would work fine. As with GCC, check that it is included in your executable path by opening a new command window, and typing

make -v

Linux development tools

A Raspberry Pi 2+ is quite adequate for compiling and debugging the test programs.

Although RPi Linux already has an ARM compiler installed, the executable programs it creates are heavily dependant on the operating system, so we also need to install a cross-compiler: arm-none-eabi-gcc version 7-2018-q2-update. The easiest way to do this is to click on Add/Remove software in the Preferences menu, then search for arm-none-eabi. The correct version is available on Raspbian ‘Buster’, but probably not on earlier distributions.

The directory structure is the same as for Windows, with the SDK components, external, integration and modules directories copied into the nrf5_sdk subdirectory.

As with Windows, it is worth typing

arm-none-eabi-gcc  -v make sure the GCC executable is installed correctly.


This is in the nrf_test1 directory, and is as simple as you can get; it just flashes the blue LED at 1 Hz.

// Simple LED blink on nRF52832 breakout board, from

#include "nrf_gpio.h"
#include "nrf_delay.h"

// LED definitions
#define LED_PIN      7
#define LED_BIT      (1 << LED_PIN)

int main(void)

    while (1)
        NRF_GPIO->OUT ^= LED_BIT;

// EOF

An unusual feature of this CPU is that the I/O pins aren’t split into individual ports, there is just a single port with a bit number 0 – 31. That number is passed to an SDK function to initialise the LED O/P pin, and I could have used another SDK function to toggle the pin, but instead used an exclusive-or operation on the hardware output register.

The SDK delay function is implemented by performing dummy CPU operations, so isn’t particularly accurate.


For both platforms, the method is the same: change directory to nrf_test1, and type ‘make’; the response should be similar to:

Assembling ../nrf5_sdk/modules/nrfx/mdk/gcc_startup_nrf52.S
 Compiling ../nrf5_sdk/modules/nrfx/mdk/system_nrf52.c
 Compiling nrf_test1.c
 Linking build/nrf_test1.elf
    text    data     bss     dec     hex filename
..for Windows..
    1944     108      28    2080     820 build/nrf_test1.elf
..or for Linux..
    2536     112     172    2820     b04 build/nrf_test1.elf

If your compile-time environment differs from mine, it shouldn’t be difficult to change the Makefile definitions to match, but there are some points to note:

  • The main changeable definitions are towards the top of the file. Resist the temptation to rearrange CFLAGS or LNFLAGS, as this can create a binary image that crashes the target system.
  • You can add files to the SRC_FILES definition, they will be compiled and linked in; the order of the files isn’t significant, but I generally put gcc_startup_nrf52.S first, so Reset_Handler is at the start of the executable code. Similarly, INC_FOLDERS can be expanded to include any other folders with your .h files.
  • The task definitions toward the bottom of the file use the tab character for indentation. This is essential: if replaced with spaces, the build process will fail.
  • ELF, HEX and binary files are produced in the ‘build’ subdirectory; ELF is generally used with GDB, while HEX is required by the JLink flash programmer.
  • I’ve defined the jflash and ocdflash tasks, that do flash programming after the ELF target is built; you can add your own custom programming environment, using a similar syntax.
  • The makefile will re-compile any C source files after they are changed, but will not automatically detect changes to the ‘include’ files, or the makefile itself; when these are edited, it will be necessary to force a re-make using ‘make -B’.
  • If a new image won’t run on the target system, the most common reason is an un-handled exception, and it can be quite difficult to find the cause. So I’d recommend that you expand the code in relatively small steps, making it easier to backtrack if there is a problem.

Device programming

Having built the binary image, we need to program it into Flash memory on the target device. This can be done by:

  • JLink adaptor on an evaluation board (Windows PC only)
  • Directly driving OpenOCD (RPi only)
  • Using the GNU debugger GDB to drive OpenOCD (both platforms)

Device programming using JLink

Set up the hardware and install the Nordic nRF Command Line Tools as described above, then the nrfjflash utility can be used to program the target device with a hex file, e.g.

nrfjprog --program build/nrf_test1.hex --sectorerase
nrfjprog --reset

The second line resets the chip after programming, to start the program running. This is done via the SWD lines, a hardware reset line isn’t required; alternatively you can just power-cycle the target board.

The above commands have been included in the makefile, so if you enter ‘make jflash’, the programming commands will be executed after the binary image is built.

An additional usage of the JLink programmer is to restore the original Arduino bootloader, that was pre-installed on the Sparkfun board. To do this, you need to get hold of the softdevice and DFU files from the Sparkfun repository, combine them using the Nordic merge utility, then program the result using a whole-chip erase:

mergehex -m s132_nrf52_2.0.0_softdevice.hex sfe_nrf52832_dfu.hex -o dfu.hex
nrfjprog --program dfu.hex --chiperase
nrfjprog --reset 

Device programming using OpenOCD

OpenOCD can be used to directly program the target device, providing the image has been built on the Raspberry Pi, or the ELF file has been copied from the development system. Install and test OpenOCD as described in the Raspberry Pi Hardware section above (check the DPIDR value is correct), hit ctrl-C to terminate it, then enter the command:

sudo openocd -f ../openocd/rpi2.cfg -f ../openocd/nrf52_swd.cfg -c "program build/nrf_test1.elf verify reset exit"

The response should be similar to:

 ** Programming Started **
 Info : nRF52832-QFAA(build code: E0) 512kB Flash
 Warn : using fast async flash loader. This is currently supported
 Warn : only with ST-Link and CMSIS-DAP. If you have issues, add
 Warn : "set WORKAREASIZE 0" before sourcing nrf51.cfg/nrf52.cfg to disable it
 ** Programming Finished **
 ** Verify Started **
 ** Verified OK **
 ** Resetting Target **
 shutdown command invoked

Note the warnings: by default, OpenOCD uses a ‘fast async flash loader’ that achieves a significant speed improvement by effectively sending a write-only data stream. Unfortunately the Nordic chip occasionally takes exception to this, and returns a ‘wait’ response, which can’t be handled in fast async mode, so the programming fails – in my tests with small binary images, it does fail occasionally. As recommended in the above text, I’ve tried adding ‘set WORKAREASIZE 0’ to nrf52_swd.cfg (before ‘find target’), but this caused problems when using GDB. By the time you read this, the issue may well have been solved; if not, you might have to do some experimentation to get reliable programming.

The makefile includes the OpenOCD direct programming commands, just run ‘make ocdflash’.

Device programming using GDB and OpenOCD

The primary reason for using GDB is to debug the target program, but it can also serve as a programming front-end for OpenOCD. This method works with PC host, or directly on the RPi, as shown in the following diagram.

GDB OpenOCD debugging

In both cases we are using the GB ‘target remote’ command; on the development PC we have to specify the IP address of the RPi: for example, as shown above. If in doubt as to the address, it is displayed if you hover the cursor over the top-right network icon on the RPi screen. By default, OpenOCD only responds to local GDB requests, so the command ‘bindto’ must be added to the configuration. This means anyone on the network could gain control of OpenOCD, so use with care: consider the security implications.

Alternatively, the Raspberry Pi can host both GDB and OpenOCD, in which case the ‘localhost’ address is used, and there is no need for the additional ‘bindto’.

The commands for the PC-hosted configuration are:

# On the RPi:
  sudo openocd -f ../openocd/rpi2.cfg -f ../openocd/nrf52_swd.cfg -c "bindto"

# On the Windows PC:
arm-none-eabi-gdb -ex="target remote" build\nrf_test1.elf -ex "load" -ex "det" -ex "q"

The PC connects to the OpenOCD GDB remote server on port 3333, loads the file into the target flash memory, detaches from the connection, and exits. The response will be something like:

Loading section .text, size 0x790 lma 0x0
 Loading section .ARM.exidx, size 0x8 lma 0x790
 Loading section .data, size 0x6c lma 0x798
 Start address 0x2b4, load size 2052
 Transfer rate: 4 KB/sec, 684 bytes/write.
 Detaching from program: c:\Projects\nrf_test\nrf_test1\build\nrf_test1.elf, Remote target
 Ending remote debugging.

I have experienced occasional failures with the message “Error finishing flash operation”, in which case the command must be repeated; see my comments on the ‘fast async flash loader’ above.

The Rpi-hosted command sequence is similar:

# On the RPi (first terminal):
sudo openocd -f ../openocd/rpi2.cfg -f ../openocd/nrf52_swd.cfg

# On the RPi (second terminal):
gdb -ex="target remote localhost" build\nrf_test1.elf -ex "load" -ex "det" -ex "q" 

Note that the GDB programming cycle does not include a CPU reset, so to run the new program the target reset button must be pressed, or the board power-cycled.


There are many ways the first test program can be extended, I chose to add serial output (including printf), and also a timeout function based on the ARM systick timer, so the delay function doesn’t hog the CPU. The main loop is:

int main(void)
    uint32_t tix;

    mstimeout(&tix, 0);
    printf("\nNRF52 test\n");
    while (1)
        if (mstimeout(&tix, 500))
            NRF_GPIO->OUT ^= LED_BIT;

I encountered two obstacles; firstly, I ran out of time trying to understand how to create a non-blocking serial transmit routine using the SDK buffering scheme, so implemented a simple circular buffer that is polled for transmit characters in the main program loop.

The second obstacle was that the CPU systick is a 24-bit down-counter clocked at 64 MHz, which means that it wraps around every 262 milliseconds. So we can’t just use the counter value to check when 500 milliseconds has elapsed, it needs some creative coding to measure that length of time; with hindsight, it might have been better to use a conventional hardware timer.

To build the project just change directory to nrf_test2, and use ‘make’ as before. The source code is fairly self explanatory, but the following features are a bit unusual:

  • For printf serial output, the Arduino programming link on the 6-way connector can’t be used, so we have to select an alternative.
  • A remarkable feature of the UART is that we can choose any unused pin for I/O; the serial signals aren’t tied to specific pins. I’ve arbitrarily chosen I/O pin 15 for output, 14 for input.
  • The method of initialising the UART and the printf output is also somewhat unusual, in that it involves a ‘context’ structure with the overall settings, in addition to the configuration structure.

Viewing serial comms

Serial I/O pins used by nrf_test2
Raspberry Pi SWD and serial connections

The serial output from the target system I/O pin 15 is a 3.3V signal, that is compatible with the serial input pin 10 (BCM 15) on the RPi (TxD -> RxD). To enable this input, launch the Raspberry Pi Configuration utility, select ‘interfaces’, enable the serial port, disable the serial console, and reboot.

To view the serial data, you could install a comms program such as ‘cutecom’, or just enter the following command line in a terminal window (ctrl-C to exit):

stty -F /dev/ttyS0 115200 raw; cat /dev/ttyS0


We have already used GDB to program the target system, a similar setup can be used for debugging. Some important points:

  • You’ll be working with 2 binary images; one that is loaded into GDB, and another that has been programmed into the target, and these two images must be identical. If in doubt, you need to reprogram the target.
  • The .elf file that is loaded into GDB contains the binary image and debug symbols, i.e.the names and addresses of your functions & variables. You can load in a .hex file instead, but that has no symbolic information, so debugging will be very difficult.
  • Compiler optimisation is normally enabled (using the -O3 option) as it generates efficient code, but this code is harder to debug, since there isn’t a one-to-one correspondence between a line of source and a block of instructions. Disabling optimisation will make the code larger and slower, but easier to debug; to do this, comment out the OPTIMISE line in the makefile (by placing ‘#’ at the start) and rebuild using ‘make -B’
  • OpenOCD must be running on the Raspberry Pi, configured for SWD mode and the NRF52 processor (files rpi2.cfg and nrf52_swd.cfg). It will be fully remote-controlled from GDB, so won’t require any other files on the RPi.
  • GDB must be invoked in remote mode, with “target remote ADDR:3333” where ADDR is the IP address of the Raspberry Pi, or localhost if GDB and OpenOCD are running on the same machine.
  • GDB commands can be abbreviated providing there is no ambiguity, so ‘print’ can be shortened to ‘p’. Some commands can be repeated by hitting the Enter key, so if the last command was ‘step’, just hit Enter to do another step.

Here is a sample debugging session (user commands in bold):

# On the RPi:
sudo openocd -f ../openocd/rpi2.cfg -f ../openocd/nrf52_swd.cfg -c "bindto"

# On the PC, if RPi is at
arm-none-eabi-gdb -ex="target remote" build/nrf_test2.elf
Target system halts, current source line is shown

# Program binary image into target system
Loading section .text, size 0x215c lma 0x0
Loading section .log_const_data, size 0x10 lma 0x215c
..and so on..

# Print Program Counter (should be at reset handler)
p $pc
$1 = (void (*)()) 0x2b4 <Reset_Handler>

# Execute program (continue)

# Halt program: hit ctrl-C, target reports current location
Program received signal SIGINT, Interrupt.
 main () at nrf_test2.c:72
 72              poll_serial();

# Print millisecond tick count
p msticks
$3 = 78504

# Print O/P port value in hex
$4 = 0x8080

# Toggle LED pin on O/P port
set NRF_GPIO->OUT ^= 1<<7

# Restart the program from scratch, with breakpoint
set $pc=Reset_Handler
b putch
Breakpoint 1, putch (c=13) at nrf_test2.c:149
 149         int in=ser_txin+1;

# Single-step, and print a local variable
151         if (in >= SER_TX_BUFFLEN)
p in
$5 = 46

# Detach from remote, and exit

Next step

I guess the next step is to get wireless communications working, watch this space…

Copyright (c) Jeremy P Bentham 2019. Please credit if you use the information or software in here.

Raspberry Pi position detection using fiducial tags


What is a fiducial?

You may not have heard the word ‘fiducial’ before; outside the world of robotics (or electronics manufacture) it is little known. It refers to an easily-detected optical marker that is added to an object, so its position can be determined by an image-processing system.

It is similar to a 2-dimensional QR barcode, but has a much simpler structure, so can be detected at a distance; the tags in the photo above are only 12 mm (0.5 inch) in size, but I’ve successfully detected them in an HD image at a distance of 1.6 metres (over 5 feet).

The image analysis returns the x,y position of the tag centre, and the coordinates of its 4 corners, which can be used to highlight the tag outline in the camera image display; there is also a ‘goodness factor’ that indicates how well the tag has been matched; this can be used to filter out some spurious detections.

There isn’t just one type of fiducial; several organisations have developed their own formats. The type directly supported by OpenCV is known as ArUco, but I’ve opted for a rival system developed by the University of Michigan, called AprilTag. They have a full set of open-source software to generate & decode the tags; the decoder is written in C, with Python bindings, so can easily be integrated into a Raspberry Pi image processing system.

The AprilTag package has several tag ‘families’, that are characterised by two numbers; the number of data bits in a square, and the hamming distance between adjacent tags, e.g. 16h5 is a 4-by-4 data square, with a hamming distance of 5. The hamming distance is used to remove similar-looking tags that might easily be confused for each other, including rotations, so although 16h5 has 16 data bits, there are only 30 unique tags in that family.

I’m using 3 of the simpler families: 16h5, 25h9 and 36h11. Here are the tag values of 0 to 2 for each of them:


Generating Apriltag images

The original Apriltag generator here is written in Java, with the option of auto-generating C code. For simplicity, I’ve completely rewritten it in Python, with the option of outputting a bitmap (PNG/JPEG) or vector (SVG) file. The vector format allows us to generate tags with specific dimensions, that can accurately be reproduced by a low-cost laser printer.

To generate the tags, we need some ‘magic numbers’ that indicate which bits are set for a given tag. I got these numbers from the original Java code, for example has the lines:

public class Tag16h5 extends TagFamily
  public Tag16h5()
    super(16, 5, new long[] { 0x231bL, 0x2ea5L, 0x346aL etc..

I’ve copied the first 10 data entries from Tag16h5, 25h9 and 36h11 Java files:

tag16h5 =  16, 5,(0x231b,0x2ea5,0x346a,0x45b9,0x79a6,
tag25h9  = 25, 9,(0x155cbf1,0x1e4d1b6,0x17b0b68,0x1eac9cd,0x12e14ce,
tag36h11 = 36,11,(0xd5d628584,0xd97f18b49,0xdd280910e,0xe479e9c98,0xebcbca822,

If you need more than 10 different tags of a given family, just copy more data values.

In my code, a tag is created as a 2-dimensional Numpy array, where ‘0’ is a black square, and ‘1’ is white. The source data is a right-justified bit-stream, for example the above value of 231b hex is decoded as follows:

There is a 1-bit solid black frame around the data bits, and an (invisible) 1-bit white frame round that. The encoder steps are:

  • Calculate the number of data bits per row by taking the square root of the area
  • Load the data for the required tag as an 8-byte big-endian value, convert it to a linear array of byte values
  • Convert the byte array into bits, discard the unused left-most bits, and reshape into a square array
  • Add a black (0) frame around the array
  • Add a white (1) frame around the black frame
# Generate a tag with the given value, return a numpy array
def gen_tag(tag, val):
    area, minham, codes = tag
    side = int(math.sqrt(area))
    d = np.frombuffer(np.array(codes[val], ">i8"), np.uint8)
    bits = np.unpackbits(d)[-area:].reshape((-1,side))
    bits = np.pad(bits, 1, 'constant', constant_values=0)
    return np.pad(bits, 2, 'constant', constant_values=1)

We now have a numpy array with the desired binary pattern, that needs to be turned into a graphic.

Bitmap output

The extension on the output filename (.png, .jpg, .pgm, or .svg) determines the output file format. If a bitmap is required, Python Imaging Library (PIL, or the fork ‘pillow’) is used to convert the list of tag arrays into graphic objects. The binary bits only need to be multiplied by 255 to provide the full monochrome value, then are copied into the image. This creates one-pixel squares that are invisible without zooming, so the whole image is scaled up to a reasonable size.

# Save numpy arrays as a bitmap
def save_bitmap(fname, arrays):
    img ='L', (IMG_WD,IMG_HT), WHITE)
    for i,a in enumerate(arrays):
        t = Image.fromarray(a * WHITE)
        img.paste(t, (i*TAG_PITCH,0))
    img = img.resize((IMG_WD*SCALE, IMG_HT*SCALE)), FTYPE)

PGM output is an old uncompressed binary format, that is rarely encountered nowadays: it can be useful here because it is compatible with the standard apriltag_demo application, which I’ll be describing later.

Vector output

The vector (SVG) version uses the ‘svgwrite’ library, that can be installed using pip or pip3 as usual. The tag size is specified by setting the document and viewport sizes:

    SCALE     = 2
    DWG_SIZE  = "%umm"%(IMG_WD*SCALE),"%umm"%(IMG_HT*SCALE)
    VIEW_BOX  = "0 0 %u %s" % (IMG_WD, IMG_HT)

This means each square in the tag will be 2 x 2 mm, so 4 x 4 data bits plus a 1-bit black frame makes the visible tag size 12 x 12 mm.

The background is defined as white, so only the black squares need to be drawn; the numpy ‘where’ operator is used to return a list of bits that are zero.

# Save numpy arrays as a vector file
def save_vector(fname, arrays):
    dwg = svgwrite.Drawing(fname, DWG_SIZE, viewBox=VIEW_BOX, debug=False)
    for i,a in enumerate(arrays):
        g = dwg.g(stroke='none', fill='black')
        for dy,dx in np.column_stack(np.where(a == 0)):
            g.add(dwg.rect((i*TAG_PITCH + dx, dy), (1, 1)))

Each tag is defined as a separate SVG group, which is convenient if it has to be copy-and-pasted into another image. If you are unfamiliar with SVG, take a look at my blog on the subject.

Source code for Apriltag generator

The source code ( is compatible with Python 2.7 or 3.x, and can run on Windows or Linux. It requires numpy, svgwrite, and PIL/pillow to be installed using pip or pip3 as usual:

# Apriltag generator, from

import sys, math, numpy as np, svgwrite
from PIL import Image

filename  = 'test.svg'  # Default filename (.svg, .png, .jpeg or .pgm)
family    = 'tag16h5'   # Default tag family (see tag_families)
NTAGS     = 10          # Number of tags to create
TAG_PITCH = 10          # Spacing of tags
WHITE     = 255         # White colour (0 is black)

# First 10 values of 3 tag families
tag16h5 =  16, 5,(0x231b,0x2ea5,0x346a,0x45b9,0x79a6,
tag25h9  = 25, 9,(0x155cbf1,0x1e4d1b6,0x17b0b68,0x1eac9cd,0x12e14ce,
tag36h11 = 36,11,(0xd5d628584,0xd97f18b49,0xdd280910e,0xe479e9c98,0xebcbca822,
tag_families = {"tag16h5":tag16h5, "tag25h9":tag25h9, "tag36h11":tag36h11}

# Set up the graphics file, given filename and tag family
def set_graphics(fname, family):
    FTYPE = fname.split('.')[-1].upper()
    FTYPE = FTYPE.replace("PGM", "PPM").replace("JPG", "JPEG")
    IMG_HT = int(math.sqrt(family[0])) + 6

    # Vector definitions
    if FTYPE == "SVG":
        SCALE     = 2
        DWG_SIZE  = "%umm"%(IMG_WD*SCALE),"%umm"%(IMG_HT*SCALE)
        VIEW_BOX  = "0 0 %u %s" % (IMG_WD, IMG_HT)

    # Bitmap definitions
        SCALE = 10

# Generate a tag with the given value, return a numpy array
def gen_tag(tag, val):
    area, minham, codes = tag
    dim = int(math.sqrt(area))
    d = np.frombuffer(np.array(codes[val], ">i8"), np.uint8)
    bits = np.unpackbits(d)[-area:].reshape((-1,dim))
    bits = np.pad(bits, 1, 'constant', constant_values=0)
    return np.pad(bits, 2, 'constant', constant_values=1)

# Save numpy arrays as a bitmap
def save_bitmap(fname, arrays):
    img ='L', (IMG_WD,IMG_HT), WHITE)
    for i,a in enumerate(arrays):
        t = Image.fromarray(a * WHITE)
        img.paste(t, (i*TAG_PITCH,0))
    img = img.resize((IMG_WD*SCALE, IMG_HT*SCALE)), FTYPE)

# Save numpy arrays as a vector file
def save_vector(fname, arrays):
    dwg = svgwrite.Drawing(fname, DWG_SIZE, viewBox=VIEW_BOX, debug=False)
    for i,a in enumerate(arrays):
        g = dwg.g(stroke='none', fill='black')
        for dy,dx in np.column_stack(np.where(a == 0)):
            g.add(dwg.rect((i*TAG_PITCH + dx, dy), (1, 1)))

if __name__ == '__main__':
    opt = None
    for arg in sys.argv[1:]:    # Process command-line arguments..
        if arg[0]=="-":
            opt = arg.lower()
            if opt == '-f':     # '-f family': tag family
                family = arg
                filename = arg  # 'filename': graphics file  
            opt = None
    if family not in tag_families:
        print("Unknown tag family: '%s'" % family)
    tagdata = tag_families[family]
    set_graphics(filename, tagdata)
    print("Creating %s, file %s" % (family, filename))
    tags = [gen_tag(tagdata, n) for n in range(0, NTAGS)]
    if FTYPE == "SVG":
        save_vector(filename, tags)
        save_bitmap(filename, tags)

Decoding Apriltags

For the decoder, I’m using the standard Apriltag ‘C’ code, which includes a Python library, so no knowledge of the C programming language is required. The code is Linux-specific, so will run on the Raspberry Pi, but not on Windows unless you install the Microsoft ‘Windows Subsystem for Linux’, which can compile & run the text-based decoder, but sadly not the graphical display.

On the raspberry pi, I’m using the Raspbian Buster distribution; the Apriltag build process may not be compatible with older distributions. I’ve had no success building on a Pi Zero, due to the RAM size being too small, so had to compile on a larger board, and transfer the files across.

The commands to fetch and compile the code are:

sudo apt install cmake
cd ~
git clone
cd apriltag
cmake .
sudo make install
make apriltag_demo

The installation command returns an error with the Python library, but succeeds in installing the other application files.

You can now run my Python tag encoder, and feed the output into the demonstration decoder supplied in the Apriltag package, for example:

python3 -f tag16h5 test.jpg
apriltag_demo -f tag16h5 test.jpg

You should be rewarded with a swathe of text, such as:

loading test.jpg
 detection   0: id (16x 5)-0   , hamming 0, margin  203.350
 detection   1: id (16x 5)-1   , hamming 0, margin  246.072
 detection   2: id (16x 5)-2   , hamming 0, margin  235.426
 ..and so on..

The -0, -1, -2 sequence shows the decoded tag numbers, and the large ‘margin’ value indicates there is a high degree of confidence that the decode is correct. The time taken by the various decoder components is also displayed, which is useful if you’re trying to optimise the code.

If the decode fails, check that you’ve entered the tag family & filename correctly; the decoder application doesn’t accept JPEG files with a .jpeg extension, it has to be .jpg.

Python tag decoder

To use the Python library interface, you have to tell Python where to find the library file, for example at the command prompt:

export PYTHONPATH=${PYTHONPATH}:${HOME}/apriltag
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${HOME}/apriltag/lib

This can be a bit of a nuisance; a quick (but rather inefficient) alternative is to copy the ‘.so’ library file from the compiled Apriltag package into the current directory. For my current build, the command would be:

cp ~/apriltag/ .

You can now run a simple Python console program to exercise the library. It uses Python OpenCV, which needs to be installed using ‘apt’; see this blog for more information. File

# Simple test of Apriltag decoding from

import cv2
from apriltag import apriltag

fname = 'test.jpg'
image = cv2.imread(fname, cv2.IMREAD_GRAYSCALE)
detector = apriltag("tag16h5")
dets = detector.detect(image)
for det in dets:
    print("%s: %6.1f,%6.1f" % (det["id"], det["center"][0], det["center"][1]))

You will need to run this under python3, as the Apriltag library isn’t compatible with Python 2.x. The output is somewhat uninspiring, just showing the tag value, and the x & y positions of its centre, but is sufficient to show the decoder is working:

0:   49.9,  49.9
1:  149.9,  49.8
2:  249.9,  49.9
..and so on..

Graphical display of detected tags

A better test is to take video from the Raspberry Pi camera, detect the value and position of the tags, and overlay that information onto the display. Here is the source code (

# Detect Apriltag fiducials in Raspbery Pi camera image
# From

import cv2
from apriltag import apriltag

TITLE      = "apriltag_view"  # Window title
TAG        = "tag16h5"        # Tag family
MIN_MARGIN = 10               # Filter value for tag detection
FONT       = cv2.FONT_HERSHEY_SIMPLEX  # Font for ID value
RED        = 0,0,255          # Colour of ident & frame (BGR)

if __name__ == '__main__':
    cam = cv2.VideoCapture(0)
    detector = apriltag(TAG)
    while cv2.waitKey(1) != 0x1b:
        ret, img =
        greys = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        dets = detector.detect(greys)
        for det in dets:
            if det["margin"] >= MIN_MARGIN:
                rect = det["lb-rb-rt-lt"].astype(int).reshape((-1,1,2))
                cv2.polylines(img, [rect], True, RED, 2)
                ident = str(det["id"])
                pos = det["center"].astype(int) + (-10,10)
                cv2.putText(img, ident, tuple(pos), FONT, 1, RED, 2)
        cv2.imshow(TITLE, img)

To test the code, create a tag16h5 file in SVG format:

 python3 -f tag16h5 test.svg

This vector file can be printed out using Inkscape, to provide an accurately-sized set of paper tags, or just displayed on the Raspberry Pi screen, by double-clicking in File Manager. Then run apriltag_view:


With the camera pointed at the screen, you can position the decoded images and original tags so they are both in view. Note that the camera doesn’t need to be at right-angles to the screen, the decoder can handle oblique images. The MIN_MARGIN value may need to be adjusted; it can be increased to suppress erroneous detections, but then some distorted tags may be missed.

To terminate the application, press the ESC key while the decoder display has focus.

The application is a bit slower than I’d like, with a noticeable lag on the image display, so the code needs to be optimised.

Copyright (c) Jeremy P Bentham 2019. Please credit this blog if you use the information or software in it.

Accurate position measurement using low-cost cameras and OpenCV

There are many ways to sense the position of an object, and they’re generally either expensive or low-resolution. Laser interferometers are incredibly accurate, but the complex optics & electronics make the price very high. Hand-held laser measures are quite cheap, but they use a time-of-flight measurement method which limits their resolution, as light travels at roughly 1 foot (300 mm) per nanosecond, and making sub-nanosecond measurements isn’t easy (but do check out my post on Ultra Wideband ranging, which does use lightspeed measurements). Lidar (light-based radar) is currently quite expensive, and has similar constraints. Ultrasonic methods benefit from the fact that sound waves travel at a much slower speed; they work well in constrained environments, such as measuring the height of liquid in a tank, but multipath reflections are a problem if there is more than one object in view.

Thanks to the smartphone boom, high-resolution camera modules are quite cheap, and I’ve been wondering whether they could be used to sense the position of an object to a reasonable accuracy for everyday measurements (at least 0.5 mm or 0.02 inches).

To test the idea I’ve set up 2 low-cost webcams at right-angles, to sense the X and Y position of an LED. To give a reproducible setup, I’ve engraved a baseboard with 1 cm squares, and laser-cut a LED support, so I can accurately position the LED and see the result.

The webcams are Logitech C270, that can provide an HD video resolution of 720p (i.e. 1280 x 720 pixels). For image analysis I’ll be using Python OpenCV; it has a wide range of sophisticated software tools, that allow you to experiment with some highly advanced methods, but for now I’ll only be using a few basic functions.

The techniques I’m using are equally applicable to single-camera measurements, e.g. tracking the position of the sun in the sky.

Camera input

My camera display application uses PyQt and OpenCV to display camera images, and it is strongly recommended that you start with this, to prove that your cameras will work with the OpenCV drivers. It contains code that can be re-used for this application, so is imported as a module.

Since we’re dealing with multiple cameras and displays, we need a storage class to house the data.

import sys, time, threading, cv2, numpy as np
import cam_display as camdisp

IMG_SIZE    = 1280,720          # 640,480 or 1280,720 or 1920,1080
DISP_SCALE  = 2                 # Scaling factor for display image
DISP_MSEC   = 50                # Delay between display cycles
CAP_API     = cv2.CAP_ANY       # API: CAP_ANY or CAP_DSHOW etc...

# Class to hold capture & display data for a camera
class CamCap(object):
    def __init__(self, cam_num, label, disp):
        self.cam_num, self.label, self.display = cam_num, label, disp
        self.imageq = camdisp.Queue.Queue()
        self.pos = 0
        self.cap = cv2.VideoCapture(self.cam_num-1 + CAP_API)
        self.cap.set(cv2.CAP_PROP_FRAME_WIDTH, IMG_SIZE[0])
        self.cap.set(cv2.CAP_PROP_FRAME_HEIGHT, IMG_SIZE[1])

The main window of the GUI is subclassed from cam_display, with the addition of a second display area, and storage for the camera capture data:

# Main window
class MyWindow(camdisp.MyWindow):
    def __init__(self, parent=None):
        camdisp.MyWindow.__init__(self, parent)
        self.camcaps = []
        self.disp2 = camdisp.ImageWidget(self)
        self.capturing = True

On startup, 2 cameras are added to the window:

if __name__ == '__main__':
    app = camdisp.QApplication(sys.argv)
    win = MyWindow()
    win.camcaps.append(CamCap(2, 'x', win.disp))
    win.camcaps.append(CamCap(1, 'y', win.disp2))

As with cam_display, a separate thread is used to fetch data from the cameras:

    # Grab camera images (separate thread)
    def grab_images(self):
        while self.capturing:
            for cam in self.camcaps:
                if cam.cap.grab():
                    retval, image = cam.cap.retrieve(0)
                    if image is not None and cam.imageq.qsize() < 2:
                        time.sleep(DISP_MSEC / 1000.0)
                    print("Error: can't grab camera image")
                    self.capturing = False
        for cam in self.camcaps:

Image display

A timer event is used to fetch the image from the queue, convert it to RGB, do the image processing, and display the result.

    # Fetch & display camera images
    def show_images(self):
        for cam in self.camcaps:
            if not cam.imageq.empty():
                image = cam.imageq.get()
                if image is not None and len(image) > 0:
                    img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
                    cam.pos = colour_detect(img)
                    self.display_image(img, cam.display, DISP_SCALE)
    # Show position values given by cameras
    def show_positions(self, s=""):
        for cam in self.camcaps:
            s += "%s=%-5.1f " % (cam.label, cam.pos)

Image processing

We need to measure the horizontal (left-to-right) position of the LED for each camera. If the LED is brighter than the surroundings, this isn’t difficult; first we create a mask that isolates the LED from the background, then extract the ‘contour’ of the object with the background masked off. The contour is a continuous curve that marks the boundary between the object and the background; for the illuminated LED this will approximate to a circle. To find an exact position, the contour is converted to a true circle, which is drawn in yellow, and the horizontal position of the circle centre is returned.

LOWER_DET   = np.array([240,  0,  0])       # Colour limits for detection
UPPER_DET   = np.array([255,200,200])

# Do colour detection on image
def colour_detect(img):
    mask = cv2.inRange(img, LOWER_DET, UPPER_DET)
    ctrs = cv2.findContours(mask, cv2.RETR_TREE,
    if len(ctrs) > 0:
        (x,y),radius = cv2.minEnclosingCircle(ctrs[0])
        radius = int(radius), (int(x),int(y)), radius, (255,255,0), 2)
        return x
    return 0

This code is remarkably brief, and if you’re thinking that I may have taken a few short-cuts, you’d be right:

Colour detection: I’ve specified the upper and lower RGB values that are acceptable; because this is a red LED, the red value is higher than the rest, being between 240 and 255 (the maximum is 255). I don’t want to trigger on a pure white background so I’ve set the green and blue values between 0 and 200, so a pure white (255,255,255) will be rejected. This approach is a bit too simplistic; if the LED is too bright it can saturate the sensor and appear completely white, and conversely another bright light source can cause the camera’s auto-exposure to automatically reduce the image intensity, such that the LED falls below the required level. The normal defence against this is to use manual camera exposure, which can be adjusted to your specific environment. Also it might be worth changing the RGB colourspace to HSV for image matching; I haven’t yet tried this.

Multiple contours: the findContours function returns a list of contours, and I’m always taking the first of these. In a real application, there may be several contours in the list, and it will be necessary to check them all, to find the most likely – for example, the size of the circle to see if it is within an acceptable range.

However, the measurement method does show some very positive aspects:

Complex background: as you can see from the image at the top of this blog, it works well in a normal office environment – no need for a special plain-colour background.

No focussing: most optical applications require the camera to be focussed, but in this case there is no need. I’ve deliberately chosen a target distance of approximately 4 inches (100 mm) that results in a blurred image, but OpenCV is still able to produce an accurate position indication.

Sub-pixel accuracy: with regard to measurement accuracy, the main rule for the camera is obviously “the more pixels, the better”, but also OpenCV can compute the position to within a fraction of a pixel. My application displays the position (in pixels) to one decimal place; at 4 inches (100 mm) distance, the Logitech cameras’ field of view is about 3.6 inches (90 mm), so if the position can be measured within, say, 0.2 of a pixel, this would be a resolution of 0.0006 inch (0.015 mm).

Of course these figures are purely theoretical, and the resolution will be much reduced in a real-world application, but all the same, it does suggest the technique may be capable of achieving quite good accuracy, at relatively low cost.

Single camera

With minor modifications, the code can be used in a single-camera application, e.g. tracking the position of the sun in the sky.

The code scans all the cameras in the ‘camcaps’ list, so will automatically adapt if there is only one.

The colour_detect function currently returns the horizontal position only; this can be changed to return the vertical as well. The show_positions method can be changed to display both of the returned values from the single camera.

Then you just need a wide-angle lens, and a suitable filter to stop the image sensor being overloaded. Sundial, anyone?

Source code

The ‘campos’ source code is available here, and is compatible with Windows and Linux, Python 2.7 and 3.x, PyQt v4 and v5. It imports my cam_display application, and I strongly recommended that you start by running that on its own, to check compatibility. If it fails, read the Image Capture section of that blog, which contains some pointers that might be of help.

Copyright (c) Jeremy P Bentham 2019. Please credit this blog if you use the information or software in it.

PC / RPi camera display using PyQt and OpenCV

OpenCV is an incredibly powerful image-processing tool, but it can be difficult to know where to start – how do you grab an image from a camera, and display it in a user-friendly GUI? This post describes such an application, that runs unmodified on a PC or Raspberry Pi, Windows or Linux, Python 2.7 or 3.x, and PyQt v4 or v5.


On Windows, the OpenCV and PyQt5 libraries can be installed using pip:

pip install numpy opencv-python PyQt5

If pip isn’t available, you should be able to run the module from the command line by invoking Python, e.g. for Python 3:

py -3 -m pip install numpy opencv-python PyQt5

Installing on a Raspberry Pi is potentially a lot more complicated; it is generally recommended to install from source, and for opencv-python, this is a bit convoluted. Fortunately there is a simpler option, if you don’t mind using versions that are a few years old, namely to load the binary image from the standard repository, e.g.

sudo apt update
sudo apt install python3-opencv python3-pyqt5 

At the time of writing, the most recent version of Raspbian Linux is ‘buster’, and that has OpenCV 3.2, which is quite usable. The previous ‘stretch’ distribution has python-opencv version 2.4, which is a bit too old: my code isn’t compatible with it.

With regard to cameras, all the USB Webcams I’ve tried have worked fine on Windows without needing to have any extra driver software installed; they also work on the Raspberry Pi, as well as the standard Pi camera with the ribbon-cable interface.

PyQt main window

Being compatible with PyQt version 4 and 5 requires some boilerplate code to handle the way some functions have been moved between libraries:

import sys, time, threading, cv2
    from PyQt5.QtCore import Qt
    pyqt5 = True
    pyqt5 = False
if pyqt5:
    from PyQt5.QtCore import QTimer, QPoint, pyqtSignal
    from PyQt5.QtWidgets import QApplication, QMainWindow, QTextEdit, QLabel
    from PyQt5.QtWidgets import QWidget, QAction, QVBoxLayout, QHBoxLayout
    from PyQt5.QtGui import QFont, QPainter, QImage, QTextCursor
    from PyQt4.QtCore import Qt, pyqtSignal, QTimer, QPoint
    from PyQt4.QtGui import QApplication, QMainWindow, QTextEdit, QLabel
    from PyQt4.QtGui import QWidget, QAction, QVBoxLayout, QHBoxLayout
    from PyQt4.QtGui import QFont, QPainter, QImage, QTextCursor
    import Queue as Queue
    import queue as Queue

The main window is subclassed from PyQt, with a simple arrangement of a menu bar, video image, and text box:

class MyWindow(QMainWindow):
    text_update = pyqtSignal(str)

    # Create main window
    def __init__(self, parent=None):
        QMainWindow.__init__(self, parent)

        self.central = QWidget(self)
        self.textbox = QTextEdit(self.central)
        self.textbox.setMinimumSize(300, 100)
        sys.stdout = self
        print("Camera number %u" % camera_num)
        print("Image size %u x %u" % IMG_SIZE)
        if DISP_SCALE > 1:
            print("Display scale %u:1" % DISP_SCALE)

        self.vlayout = QVBoxLayout()        # Window layout
        self.displays = QHBoxLayout()
        self.disp = ImageWidget(self)    
        self.label = QLabel(self)

        self.mainMenu = self.menuBar()      # Menu bar
        exitAction = QAction('&Exit', self)
        self.fileMenu = self.mainMenu.addMenu('&File')

There is a horizontal box layout called ‘displays’, that seems to be unnecessary as it only has one display widget in it. This is intentional, since much of my OpenCV experimentation requires additional displays to show the image processing in action; this can easily be done by creating more ImageWidgets, and adding them to the ‘displays’ layout.

Similarly, there is a redundant QLabel below the displays, which isn’t currently used, but is handy for displaying static text below the images.

Text display

It is convenient to redirect the ‘print’ output to the text box, rather than appearing on the Python console. This is done using the ‘text_update’ signal that was defined above:

    # Handle sys.stdout.write: update text display
    def write(self, text):
    def flush(self):

    # Append to text display
    def append_text(self, text):
        cur = self.textbox.textCursor()     # Move cursor to end of text
        s = str(text)
        while s:
            head,sep,s = s.partition("\n")  # Split line at LF
            cur.insertText(head)            # Insert text at cursor
            if sep:                         # New line if LF
        self.textbox.setTextCursor(cur)     # Update visible cursor

The use of a signal means that print() calls can be scattered about the code, without having to worry about which thread they’re in.

Image capture

A separate thread is used to capture the camera images, and put them in a queue to be displayed. The camera may produce images faster than they can be displayed, so it is necessary to check how many images are already in the queue; if more than 1, the new image is discarded. This prevents a buildup of unwanted images.

IMG_SIZE    = 1280,720          # 640,480 or 1280,720 or 1920,1080
CAP_API     = cv2.CAP_ANY       # or cv2.CAP_DSHOW, etc...
EXPOSURE    = 0                 # Non-zero for fixed exposure

# Grab images from the camera (separate thread)
def grab_images(cam_num, queue):
    cap = cv2.VideoCapture(cam_num-1 + CAP_API)
    cap.set(cv2.CAP_PROP_FRAME_WIDTH, IMG_SIZE[0])
    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, IMG_SIZE[1])
    if EXPOSURE:
        cap.set(cv2.CAP_PROP_AUTO_EXPOSURE, 0)
        cap.set(cv2.CAP_PROP_EXPOSURE, EXPOSURE)
        cap.set(cv2.CAP_PROP_AUTO_EXPOSURE, 1)
    while capturing:
        if cap.grab():
            retval, image = cap.retrieve(0)
            if image is not None and queue.qsize() < 2:
                time.sleep(DISP_MSEC / 1000.0)
            print("Error: can't grab camera image")

The choice of image size will depend on the camera used; all cameras support VGA size (640 x 480 pixels), more modern versions the high-definition standards of 720p (1280 x 720) or 1080p (1920 x 1080).

The camera number refers to the position in the list of cameras collected by the operating system; I’ve defined the first camera as number 1, but the OpenCV call defines the first as 0, so the number has to be adjusted.

The same parameter is also used to define the capture API setting; by default this is ‘any’, which usually works well; my Windows 10 system defaults to the MSMF (Microsoft Media Foundation) backend, while the Raspberry Pi defaults to Video for Linux (V4L). Sometimes you may need to force a particular API to be used, for example, I have a Logitech C270 webcam that works fine on Windows 7, but fails on Windows 10 with an ‘MSMF grab error’. Forcing the software to use the DirectShow API (using the cv2.CAP_DSHOW option) fixes the problem.

If you want to check which backend is being used, try:

print("Backend '%s'" % cap.getBackendName())

Unfortunately this only works on the later revisions of OpenCV.

Manual exposure setting can be a bit hit-and-miss, depending on the camera and API you are using; the default is automatic operation, and setting EXPOSURE non-zero (e.g. to a value of -3) generally works, however it can be difficult to set a webcam back to automatic operation: sometimes I’ve had to use another application to do this. So it is suggested that you keep auto-exposure enabled if possible.

[Supplementary note: it seems that these parameter values aren’t standardised across the backends. For example, the CAP_PROP_AUTO_EXPOSURE value in my source code is correct for the MSMF backend; a value of 1 enables automatic exposure, 0 disables it. However, the V4L backend on the Raspberry Pi uses the opposite values: automatic is 0, and manual is 1. So it looks like my code is incorrect for Linux. I haven’t yet found any detailed documentation for this, so had to fall back on reading the source code, namely the OpenCV videoio ‘cap’ files such as cap_msmf.cpp and cap_v4l.cpp.]

Image display

The camera image is displayed in a custom widget:

# Image widget
class ImageWidget(QWidget):
    def __init__(self, parent=None):
        super(ImageWidget, self).__init__(parent)
        self.image = None

    def setImage(self, image):
        self.image = image

    def paintEvent(self, event):
        qp = QPainter()
        if self.image:
            qp.drawImage(QPoint(0, 0), self.image)

A timer event is used to trigger a scan of the image queue. This contains images in the camera format, which must be converted into the PyQt display format:

DISP_SCALE  = 2                 # Scaling factor for display image

    # Start image capture & display
    def start(self):
        self.timer = QTimer(self)           # Timer to trigger display
                    self.show_image(image_queue, self.disp, DISP_SCALE))
        self.capture_thread = threading.Thread(target=grab_images, 
                    args=(camera_num, image_queue))
        self.capture_thread.start()         # Thread to grab images

    # Fetch camera image from queue, and display it
    def show_image(self, imageq, display, scale):
        if not imageq.empty():
            image = imageq.get()
            if image is not None and len(image) > 0:
                img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
                self.display_image(img, display, scale)

    # Display an image, reduce size if required
    def display_image(self, img, display, scale=1):
        disp_size = img.shape[1]//scale, img.shape[0]//scale
        disp_bpl = disp_size[0] * 3
        if scale > 1:
            img = cv2.resize(img, disp_size, 
        qimg = QImage(, disp_size[0], disp_size[1], 
                      disp_bpl, IMG_FORMAT)

This demonstrates the power of OpenCV; with one function call we convert the image from BGR to RGB format, then another is used to resize the image using cubic interpolation. Finally a PyQt function is used to convert from OpenCV to PyQt format.

Running the application

Make sure you’re using the Python version that has the OpenCV and PyQt installed, e.g. for the Raspberry Pi:


There is an optional argument that can be used if there are multiple cameras; the default first camera is number 1.

On Linux, some USB Webcams cause a constant stream of JPEG format errors to be printed on the console, complaining about extraneous bytes in the data. There is some discussion online as to the cause of the error, and the cure seems to involve rebuilding the libraries from source; I’m keen to avoid that, so used the simple workaround of suppressing the errors by redirecting STDERR to null:

python3 2> /dev/null

Fortunately this workaround is only needed with some USB cameras; the standard Raspberry Pi camera with the CSI ribbon-cable interface works fine.

Source code

Full source code is available here.

For a more significant OpenCV application, take a look at this post.

Copyright (c) Jeremy P Bentham 2019. Please credit this blog if you use the information or software in it.

Raspberry Pi and OpenOCD

In previous blog posts I used an FTDI module and pure Python code to access the internals of an ARM CPU using the SWD interface. I want to expand this technique to provide a more comprehensive real-time display of the CPU status, but the FTDI interface is quite limiting; what I need is an fast intelligent SWD/JTAG adaptor, with a network interface so I can do both local and remote diagnosis.

Enter the raspberry Pi: a lot of computing power at very low cost, either using the built-in HDMI display output, or running ‘headless’ over a wireless network, providing diagnostic data to a remote display.

Connecting the Pi to the target system could hardly be simpler; 3 wires (clock, data & ground) are sufficient to access data from most CPUs with an SWD interface.

Raspberry Pi 3 SWD interface to STMicro ARM CPU

Raspberry Pi ZeroW JTAG interface to Atmel ARM CPU

Software-wise, OpenOCD has all the SWD/JTAG features you’ll ever need, accessed through a network interface; installation may be a bit intimidating if you’re not an experienced Linux user, but is really quite easy, as this blog will (hopefully) demonstrate.

What you end up with is a really powerful local/remote debugger for very little money; around $10 US, in the case of the Pi Zero W.

Installing OpenOCD

You need any Raspberry Pi (RPi), versions 0 to 3. The slower boards will have longer boot times & build times but are otherwise fully functional. The OS version I used was ‘Raspbian Stretch with desktop’; the ‘recommended software’ add-ons are not necessary. The total image size on SD card is around 3 GB.

A convenient way to avoid re-typing the instructions below is to enable the Secure Shell (SSH) protocol using the ‘Raspberry Pi Configuration utility’, then run a remote ssh client (e.g. ‘putty’ on Windows) to access the RPi over the network; you can then cut and paste a command line into the ssh window without re-typing. If in doubt as to the IP address of your RPi, hover the cursor over the network icon in the top-right corner, and the address will be shown, e.g. (or run ‘ifconfig’ if in text mode).

It is best to install OpenOCD from source, as the pre-built images often lack important functionality. Installation instructions can be found on many Web sites, for example Adafruit “Programming Microcontrollers using OpenOCD on a Raspberry Pi”. In summary, the steps are:

cd ~
sudo apt-get update
sudo apt-get install git autoconf libtool make pkg-config libusb-1.0-0 libusb-1.0-0-dev
git clone
cd openocd
./configure --enable-sysfsgpio --enable-bcm2835gpio
sudo make install

The ‘make’ step takes approximately an hour on the slower boards, or 15 minutes on the faster.

Configuration files

OpenOCD has a wide variety of options, so generally needs more than one configuration file, to define:

  • Debug adaptor (in our case, the RPi)
  • Communication method (SWD or JTAG)
  • Target CPU.

There are a large number of files in /usr/local/share/openocd/scripts, most notably the ‘interface’ and ‘target’ sub-directories, however there are so many permutations that it is unlikely you’ll find everything you need, so we need to think about creating our own files.

The most important first step is to work out how the RPi will be connected to the target system…

RPi I/O connections

At the time of writing, there are 3 versions of the RPi I/O connector, and 3 different pin-numbering schemes, so it is easy to get confused. The older boards may be considered obsolete, but are still more than adequate for running OpenOCD, so mustn’t be excluded.

The numbering schemes are:

  1. Connector pin numbers: sequential 1 – 26 or 1 – 40
  2. GPIO bit numbers (also known as Broadcom or BCM numbers) 0 – 27
  3. WiringPi numbers, as used in the Python library

I’ll only be using the first 2 of these. The older boards have 26 pins, the newer 40.

RPi 26-way connector with GPIO numbers

Pins 3 & 5 were initially GPIO 0 and 1, but later became GPIO 2 and 3; they are best avoided.

RPi 40-way connector with GPIO numbers

On the 40-way connector GPIO21 has become 27, so should also be avoided. The choice of ground pin is arbitrary; any of them can be used, but I avoid pin 6, as any mis-connection to the supply pins can result in significant damage.


The SWD connections given in the OpenOCD configuration file ‘raspberrypi2-native.cfg’ are:

raspberrypi2-native SWD connections

The relevant lines in the configuration file are:

# SWD                 swclk swdio
# Header pin numbers: 22    18
bcm2835gpio_swd_nums  25    24

bcm2835gpio_srst_num  18
reset_config srst_only srst_push_pull

In many applications the reset signal is unnecessary – and undesirable, if the objective is to perform non-intrusive monitoring of a running system.


JTAG is an older (and more widely available) standard for debugging, that requires 4 wires in the place of 2 for SWD. There is a standard mapping between them (SWCLK is TCK, SWDIO is TMS), but the JTAG connections in the standard OpenOCD configuration file ‘raspberrypi-native.cfg’ use completely different pins:

raspberrypi-native JTAG connections

The relevant lines in the configuration file are:

# JTAG                tck tms tdi tdo
# Header pin numbers: 23  22  19  21
bcm2835gpio_jtag_nums 11  25  10  9

# bcm2835gpio_srst_num 24
# reset_config srst_only srst_push_pull

As standard, the reset definition is commented out.

I’m not a fan of this pinout scheme; I’d like a single setup that covers both SWD and JTAG.

Other pin functions

You might wish to use the RPi for other diagnostic functions, such as monitoring a serial link, so these pins have to be kept free. The following diagram shows the alternative pin functions.


You can use any of the blue or yellow pins for the SWD/JTAG interface, it is just a question as to which other functionality you may be needing.

Combining SWD and JTAG

The compromise I’ve adopted is to preserve the existing SWD arrangement, but move the JTAG pins so one set of connections can serve both SWD & JTAG on either the 26-way or the 40-way connectors – and I’ve also avoided using any of the predefined pins, so there are no conflicts with other functionality.

rpi_swd_jtag SWD and JTAG connections

The relevant section of the configuration file is:

# SWD                swclk swdio
# Header pin numbers 22    18
bcm2835gpio_swd_nums 25    24

# JTAG                tck tms tdi tdo
# Header pin numbers  22  18  16  15 
bcm2835gpio_jtag_nums 25  24  23  22

Target system connections

The connection points on the target system will vary from board to board; for a previous demonstration I used a ‘blue pill’ STM32F103 board that has ground, SWD clock & data conveniently on some separate header pins, but the most common standard for JTAG & SWD connections is a 20-way 2-row header, as follows:

JTAG     SWD     20-way pin
Ground   Ground  4, 6, 8, 10, 14, 16, 18, 20
TRST             3
TDI              5
TMS      SWDIO   7
TCK      SWCLK   9
TDO              13
RESET            15

There is generally a keyway on the odd-numbered side of the connector.


Two reset signals are defined: TRST is ‘tap reset’, that is intended to just reset the diagnostic port; the other signal marked RESET (which OpenOCD refers to as SRST or ‘system reset’) should reset all devices, as if a reset button has been pressed. In the experimentation I’ve done, the reset lines haven’t been needed, but this is very processor-specific; sometimes the RESET line has to be used to gain control of the target system.

It is convenient to use ribbon cable for wiring up the interface, especially if the wires follow the resistor colour code:

RPi pin  Colour  20-way pin  JTAG/SWD
9        Brown   20          Ground
12       Red     15          Reset
16       Orange  5           TDI
15       Yellow  13          TDO
18       Green   7           TMS/SWDIO
22       Blue    9           TCK/SWCLK

Or in graphical form…


Interface configuration file

The above examples show how the SWD/JTAG connections are handled, but some more data is needed to fully configure the RPi interface, most notably the I/O base address and clock scaling; this tells OpenOCD where to find the I/O interface, and how to compute its speed.

There are 2 possible values for the I/O base address: the RPi zero and v1 use 0x20000000, and v2+ use 0x3F000000. If you are unsure which value to use, the boards have an excellent feature called Device Tree that documents the current hardware configuration; enter the following command in a console window:

xxd -c 4 -g 4 /proc/device-tree/soc/ranges

The base I/O address is the second value returned, for example:

RPi zero:
00000000: 7e000000  ~...
00000004: 20000000   ...
00000008: 02000000  ....
RPi v3:
00000000: 7e000000 ~...
00000004: 3f000000 ?..
00000008: 02000000 ....
..and so on..

The clock scaling is less critical, since we’re generally aiming for around 1 MHz, which gives quite a bit of leeway in terms of being fast or slow. This is fortunate, because it is difficult to find a definitive explanation of the values that should be used for all hardware & clock settings. My understanding, from reading the source code, is that every I/O read or write instruction is followed by a loop containing NOP (CPU idle) cycles to space out the operations; this number is known as the ‘jtag_delay’, and is calculated by:

(speed_coeff / khz) - speed_offset;

..where speed_coeff & speed_offset are the two scaling parameters, and khz is the desired SWD/JTAG clock speed in kHz (all the values are integers). Obviously the delay is very CPU-dependant; the standard values in the files are:

Rpi zero and v1:
  bcm2835gpio_speed_coeffs 113714 28
RPi v2+:
  bcm2835gpio_speed_coeffs 146203 36

These do seem to give roughly the right answers, and there isn’t any great necessity for the delays to be accurate – when viewed on an oscilloscope, you can see some of the cycles being stretched by an incoming interrupt, so they never will be as accurate as a pure hardware solution.

Adaptor configuration files

Combining all the information above, here are the two adaptor configuration files: rpi1.cfg for RPi zero & v1, and rpi2.cfg for v2+

# rpi1.cfg: OpenOCD interface on RPi zero and v1

# Use RPi GPIO pins
interface bcm2835gpio

# Base address of I/O port
bcm2835gpio_peripheral_base 0x20000000

# Clock scaling
bcm2835gpio_speed_coeffs 113714 28

# SWD                swclk swdio
# Header pin numbers 22    18
bcm2835gpio_swd_nums 25    24

# JTAG                tck tms tdi tdo
# Header pin numbers  22  18  16  15 
bcm2835gpio_jtag_nums 25  24  23  22
# rpi2.cfg: OpenOCD interface on RPi v2+

# Use RPi GPIO pins
interface bcm2835gpio

# Base address of I/O port
bcm2835gpio_peripheral_base 0x3F000000

# Clock scaling
bcm2835gpio_speed_coeffs 146203 36

# SWD                swclk swdio
# Header pin numbers 22    18
bcm2835gpio_swd_nums 25    24

# JTAG                tck tms tdi tdo
# Header pin numbers  22  18  16  15 
bcm2835gpio_jtag_nums 25  24  23  22

Running OpenOCD

Finally we get to run OpenOCD, but in addition to the adaptor configuration, we need to give some details about the interface & target CPU.

The command line consists of configuration files prefixed by -f, and commands prefixed by -c. In reality, a configuration file is just a series of commands; for example you can select JTAG operation using the command-line option:

openocd -c "transport select jtag"

This is exactly the same as:

openocd -f select_jtag.cfg

where the file ‘select_jtag.cfg’ has the line:

transport select jtag

So we’ll use a mixture of commands and files on our command line. The following example is for an RPi v3 driving an SWD interface into a STM32F103 processor;  I’ve used backslash continuation characters at the end of each line to make the commands more readable:

sudo openocd -f rpi2.cfg \
             -c "transport select swd" \
             -c "adapter_khz 1000" \
             -f target/stm32f1x.cfg

Some hardware operations require superuser privileges, hence the use of ‘sudo’. The usual security warnings apply when doing this; you can try without, there will just be a ‘permission denied’ error if it fails.

For a list of supported CPUs, see the files in /usr/local/share/openocd/scripts/target

When OpenOCD runs, with a bit of luck, you’ll see something like:

BCM2835 GPIO nums: swclk = 25, swdio = 24
BCM2835 GPIO config: tck = 25, tms = 24, tdi = 23, tdo = 22
adapter speed: 1000 kHz
adapter speed: 1000 kHz
adapter_nsrst_delay: 100
none separate
cortex_m reset_config sysresetreq
Info : Listening on port 6666 for tcl connections
Info : Listening on port 4444 for telnet connections
Info : BCM2835 GPIO JTAG/SWD bitbang driver
Info : JTAG and SWD modes enabled
Info : clock speed 1001 kHz
Info : SWD DPIDR 0x1ba01477
Info : stm32f1x.cpu: hardware has 6 breakpoints, 4 watchpoints
Info : Listening on port 3333 for gdb connections

If there is a configuration or wiring error, OpenOCD usually (but not always!) returns to the command line, for example if the SWDIO line is disconnected:

BCM2835 GPIO nums: swclk = 25, swdio = 24
BCM2835 GPIO config: tck = 25, tms = 24, tdi = 23, tdo = 22
adapter speed: 1000 kHz
adapter speed: 1000 kHz
adapter_nsrst_delay: 100
none separate
cortex_m reset_config sysresetreq
Info : Listening on port 6666 for tcl connections
Info : Listening on port 4444 for telnet connections
Info : BCM2835 GPIO JTAG/SWD bitbang driver
Info : JTAG and SWD modes enabled
Info : clock speed 1001 kHz
Info : SWD DPIDR 0x02192468

..and then OpenOCD terminates back to the command line..

The clue is in the SWD Data Port ID Register (DPIDR) value. According to the datasheet for the STM32F103 CPU, this should be 1BA01477. With a data line fault, every time OpenOCD runs, a different value is returned, e.g. 0x00e65468, 0x02192468, 0x00433468 and so on; the software is just picking up noise on the data line.

A disconnected clock line is harder to diagnose, as OpenOCD just terminates after the ‘clock speed’ report, with no error indication. Try using the -d option to invoke a debug display, and you’ll see lines like

JUNK DP read reg 0 = ffffffff

which suggests that all is not well in the hardware interface.

Another thing to try in the event of a failure is adding or removing a reset line, and changing its configuration entries; if there is a reset problem you’ll probably see the DPIDR value reported correctly, but other functions may not work.

What now?

Having just written 2100 words and drawn 8 diagrams, I’m going to take a short break. However, first I ought to give some indication as to how you control this OpenOCD setup.

The sign-on text mentions a telnet interface on port 4444, so we can use that; the commands highlighted in bold:

sudo apt-get install telnet  # ..if not already installed

telnet localhost 4444
Trying ::1...
Connected to localhost.
Escape character is '^]'.
Open On-Chip Debugger

As standard, this interface only works when telnet is running locally on the Raspberry Pi. To open it up to a wider network, add the command ‘bindto’ to the configuration. However, this option comes with a major security warning; think very carefully before making the system accessible to everyone on the network.

Refer to the OpenOCD documentation for information on the large number of commands that can be used over telnet, for example displaying memory using ‘mdw’ or halting the processor using ‘halt’. When finished, close the telnet link with ‘exit’.

Open-source development toolchain

To learn more about the way OpenOCD can be used with GCC and GDB to program & debug ARM target systems, take a look at this post.

Copyright (c) Jeremy P Bentham 2019. Please credit if you use the information or software in here.