Real-Time Tracking-by-Detection in Broadcast Sports Videos

This paper presents a novel algorithm for real-time tracking-by-detection of players in sports videos. The solution consists of a Convolutional Neural Network with optimized architecture and a new type of Inter Frame Connection logic for data association and tracklet handling. The proposed data association connects region proposals in the current frame with detected objects in the previous frames. The association is established before thresholding the region proposals and used to stabilize the detection performance. The proposed solution demonstrates superior performance over the existing pre-trained detection models and tracking concepts.