Tuning#
Now that everything is installed the system needs to be tuned for your use-case. This can be done by running the tuner:
# Go into the dynaikontrap code directory
cd dynaikontrap
# Activates the virtual environment
source ./venv/bin/activate
# Start the tuner
python tuner.py
You will be asked some questions to help determine the best parameters for the system to use. These settings are saved in a settings.json file, which is loaded when you start the actual camera trap program. Below is relevant information for each question in the tuning script:
Pipeline
This dictates the processing pipeline DynAikonTrap will use, information about each pipeline may be found in How it Works, for a low-powered HD resolution video camera trap, choose the LOW_POWER option.
Framerate
Number of frames that are captured by the camera every second. Testing indicates this should not exceed 20 FPS for HD resolutions. If you require a higher frame-rate than this, it is recomended to use the legacy pipeline and drop the resolution down considerably, ie to 640x480 (VGA).
Resolution
Dimensions of the captured images and video. This is specified using width and height in the tuning script. Take a look at the relevant PiCamera documentation for information on valid width and height combinations for your camera model. Also check that the chosen resolution supports video mode at the desired framerate. Note that certain dimensions limit the field of view of the camera.
Visible animal area to trigger
The expected visible area (on a plane parallel to the camera lens) of a subject animal, expressed in square-metres. This is used to determine what amount of motion will trigger the camera.
Expected distance of animal from sensor
Roughly the distance you would expect your subject animal to be from the camera sensor. A sensible value here may be around a metre, although you are encouraged to try out different values.
Animal trigger speed
The minimum speed, in metres per second, you would require your animal to travel at for motion to be triggered. Ensure this is not too low, so the animal detection stage is not overwhelmed with motion sequences. If you would like to set a low threshold make sure the background is not likely to experience movement, for example by pointing the camera at a wall rather than a bush.
Camera focal length
Focal length, in metres, as stated by the manufacturer. You may find this summary table helpful.
Pixel size
Pixel size, in metres, (single dimension) as stated by the manufacturer. You may find this summary table helpful.
Number of pixels
Sensor resolution width as stated by the manufacturer. You may find this summary table helpful.
SoTV small movement threshold
The initial threshold applied to all movement vectors independently. This should be a small value and is given in pixel dimensions.
SoTV smoothing IIR order
Order for the IIR filter on the output of the SoTV motion filtering stage. Testing has shown that an order of 3 is appropriate to minimise delays whilst still achieving the desired smoothing effect.
SoTV smoothing IIR stop-band attenuation
Amount by which the frequencies in the stop-band are to be attenuated by, in dB. These are the higher frequencies that are to be removed, leading to a smoother output. -35dB has been found to work well here.
Animal confidence threshold
Confidence value to be exceeded for the animal detector to declare a frame as containing an animal.
FASTCAT-Cloud animal detect
This option configures the camera trap to query our FASTCAT-Cloud API for deep neural network animal detection rather than running model inference on device. Our models available online are much larger and more accurate at detecting species and can be used so long as the device has a valid internet connection during deployment.
Filter humans
As well as filtering for animal detections, we also have a model available on-device which can distinguish humans from animals. If this option is selected, DynAIkonTrap will attempt to throw away any video/image detections which it deems as containing a human to protect individual privacy in deployed locations.
Human confidence threshold
Confidence value to be exceeded for the human detector to declare a frame as containing an human.
Maximum motion sequence period
Maximum length for a single motion sequence, in seconds. A new motion sequence is started if the current one exceed this limit.
Motion context buffer length
This is a low-powered pipeline parameter. The number of seconds selects the amount of video for head and tail context to detections. For example, a produced video may have a number of recorded seconds before animal enters frame and some seconds of video after it has left. We call this context time.
Fraction of event to process with neural network.
This is a low-powered pipeline parameter. This is the fraction of raw frames which are processed with a neural network in the worst case in our spiral inference scheme. Higher fractions will result in more required computation as a trade off for higher recall of animal events. It is reccomended to set this value to 0.0 for low-compute capable devices, such as Raspberry Pi Zero W and to 1.0 for more capable devices, such as Raspberry Pi 4B
Sensor board port
Port to be used to communicate with the USB sensor board. This will usually be /dev/ttyUSB0.
Sensor board baud rate
Baud rate to be used to communicate with the USB sensor board.
Sensor reading interval
Interval, in seconds, at which the sensor board is read.
Output mode
Choose between saving to disk (d) or sending data to a server (s) via HTTP requests. If picking the latter you will need to configure a server to use the simple API.
FASTCAT-Cloud upload
This option configures DynAIkonTrap to upload its observations to your FASTCAT-Cloud account. If no internet connection can be established, detections will be written to disk instead.
Output path
A location for all recordings to be saved to. Leaving this empty saves them in the DynAIkonTrap project directory, by default, it’s set to a folder called output.
Server address
URI of the server to which captures are to be transmitted using the implemented API.
Output format
Whether or not output is to be saved in video format. The alternative is to output still images.
Device ID
An identifier to use for the camera trap. This is not used other than in output meta-data. This could be used to uniquely identify camera traps if multiple of these are in use.
Delete metadata
In the low-powered pipeline, DynAIkonTrap buffers video to disk which is analysed with a background process. It may be desirable to keep these metadata for further processing/debugging. This option allows the user to disable deleting metadata.
Logging level
Choose the minimum threshold for logging. Messages with a level below this will not be output. The recommended level is INFO as this provides informative, but not excessive, output.
Logger output file
This dictates the file DynAikonTrap will output log messages to. By default, this is set to /dev/stdout, which will cause log messages to appear at the terminal. If you wish to save a system log, add a file name of your choice.