High Speed and Tiny Objects Tracking System in Racquet Sports Videos Using Deep Learning with Trajectory Rectification Feature

Racquet sports like tennis, badminton and table tennis are popular worldwide. Players spend countless hours practising to perfect their skills and become the best in their domain. Players and coaches analyse multiple professional games to improve their own. To help with this, multiple deep-learning models have emerged that track the ball or shuttlecock in these games. However, they still face many challenges. We present the TrackNetV3 model in three different racquet sports - tennis, badminton, and table tennis. TrackNetV3 is a SOTA (State Of The Art) model used previously for shuttlecock trajectories. We aim to determine the tracking system’s robustness by testing it on multiple sports, including scenarios with rapid movements and occlusions, which are common in racquet sports. The TrackNetV3 has been trained and tested on the 2017 Summer Universiade tennis tournament’s men’s singles championship, 18 broadcast videos which include professional and amateur gameplays for badminton and 12 broadcast videos from the OpenTT dataset for table tennis. We prove the model’s robustness and performance through extensive experiments compared to TrackNetV2 and YOLOv9.