Chess Detection (and tracking)

The goal is, given a top-down video of a chess game, detect and track the board itself as well as the pieces (without Deep-Learning/Neural-Netowork methods, only classical Computer Vision techniques)

Methodology

After some experimentation, the method we settled on was:

• for chessboard detection: Harris Corner detection + Hough Lines detection with some fancy postprocessing
• for chess pieces detection: Hough Circles with rudimentary hyperparameter autodetermination
• for chess piece classification attempt: template matching at detected centers to a database of images (to be extended for better results)

You can observe the process in more details with intermediate results in the tracking_experiments.ipynb notebook

Result Example

The detected elements are the chessboard itself, as well as the piece centers with attempted classification by piece color/type. Also, there is preliminary event detection with the detected events written in the top-left (but its still in beta, as you can see)

Usage

python -m chess_detection.video_processing 
    ./my_chess_video.mp4 
    ./my_chess_video_result.mp4 
    --start-time 10 
    --end-time 50 
    --redetect_seconds 0.5