009 : Computer vision system for pool (2024)

In this work a low-cost computer vision based tracking solution is designed for pool, which enables ball localisation, number detection and cue stick tracking in real-time. DNN training for ball labeling is perfomed by using synthetic images. Synthetic images are rendered using path tracing from ball models which have been reconstructed from the real set. Also real images can be included into training after annotation using a custom-made annotation tool. DNN is used to extract the ball numbers in real-time from the ball detections. Cue tracker executes cue stick location and direction extraction when visible in the view. Cue tracking is performend on CPU side in parallel to the main process. Ball and cue action is channeled via WebSocket outside the system, for example to Chrome browser in real-time. Browser based GUI enables many features such as UVC camera parameters, camera calibration, snapshot fetching, video recording and streaming. Camera housings are modeled and 3D printed using Formlabs 3T. Platform: Jetson Orin/TX1, Intel NUC.