Vision Processor

A replacement for the aging (ssl-vision)[https://github.com/RoboCup-SSL/ssl-vision]. The shape based blob detector and decentralized software architecture is intended to minimize setup time and improve detection rates in uneven illumination conditions. It currently supports Teledyne FLIR (Spinnaker), Matrix Vision (Bluefox3/mvIMPACT) and OpenCV camera backends.

The vision_processor is the image processing component that processes a camera feed to multicast the detected robot and ball positions and a debug video livestream. The geometry publisher geom_publisher.py publishes the field geometry for all vision_processors, teams and the game controller. cam_viewer.py opens the mpv video player with the camera streams from the vision_processor instances.

Dependency installation and compilation

geom_publisher.py cam_viewer.py

Required only for the geometry publisher and camera viewer. mpv is optional and only required for the camera viewer.

Debian/Ubuntu based distributions: apt install protobuf-compiler python3-protobuf python3-yaml mpv

Arch based distributions: pacman -S make python-protobuf python-yaml mpv

Installation with PIP: pip install protobuf pyyaml

vision_processor automatic

1. Install the camera SDK required for your camera type (Arch user repository mvIMPACT: mvimpact-acquire Spinnaker: spinnaker-sdk)
2. Debian/Ubuntu/Arch Linux/Manjaro: Run ./setup.sh. If the script wants to install an OpenCL driver for the wrong GPU (e.g. integrated graphics card) or you want, need or have a different OpenCL driver skip the driver installation with SKIP_DRIVERS=1 ./setup.sh.

vision_processor manual

1. Install the camera SDK required for your camera type (Arch user repository mvIMPACT: mvimpact-acquire Spinnaker: spinnaker-sdk)

2. Install the required dependencies:

   • cmake
   • Eigen3
   • ffmpeg
   • gcc
   • OpenCL (GPU based runtime recommended)
   • OpenCV
   • protobuf
   • yaml-cpp

3. Compile vision_processor:

   cmake -B build . make -C build vision_processor

Setup

1. Complete the dependency installation and compilation section.
2. Configure one config-minimal.yml or config.yml for each camera, skip the geometry section for now. The camera ids are assigned like in ssl-vision:
3. Start build/vision_processor config[X].yml for each camera.
4. Tune the orientation and position of each camera. You can view the camera feeds with python/cam_viewer.py --cameras <X>.
5. Restart the vision_processors for the generation of a new sample image img/sample.[X].png and complete the geometry section of each camera config with it. Visual explanation how to determine line_corners:
6. Restart all vision_processors to reload the config.
7. Modify geometry[X].yml to match your field geometry. (for simple use cases configuring the field size, penalty area and goal will suffice)
8. Start python/geom_publisher.py geometry[X].yml.

Troubleshooting

The video livestream cycles through 4 different views:

1. Raw camera data If the data is very bright, dark or miscolored consider adjusting the camera gain, exposure and white_balance in your config[X].yml.
2. Reprojected color delta If the visible reprojected image extent does not match the field boundary the geometry calibration has issues. If color blobs are desaturated in the center your image might be overexposed/too bright (gain, exposure).
3. Gradient dot product All color blobs should be visible here as black and white checkered rings.
4. Blob circularity score If the blob score of some undetected blobs is too faint consider reducing the circularity threshold.

If blobs are attributed the wrong color (or balls are seemingly undetected despite high blob scores) adjust the reference colors under color.

If nothing helps: Activate stream: raw_feed: true in your config[X].yml and record the video livestream with ffmpeg -protocol_whitelist file,rtp,udp -i python/cam[X].sdp cam[X].mp4. Publish the resulting video including your config[X].yml and geometry[X].yml for further remote analysis.