A Deep Tracking and Segmentation Approach for Soccer Videos Visual Effects

Abstract

The applications of deep learning algorithm in sports contain enormous potential. Specifically, in soccer, tracking algorithm could record the tracks of players, which could play as an assistant to assess team performance and evaluate strategies. Moreover, through segmentation model, we could extract semantic attributes of players. This auxiliary information may contribute to the special visual effects processing in broadcasting or entertainment area. Unlike general tracking tasks, soccer videos contain much more cases of deformation, blur, and occlusion. In this paper, we propose a novel model which could combine tracking and segmentation together. A novel deformable cross-similarity correlation (DF_CORR) is adopted to estimate the deformation of players. A new soccer tracking dataset is established to evaluate the performance of top-ranked trackers in soccer videos. In soccer tracking dataset, our model outperforms the state-of-the-art trackers whose accuracy is decreased significantly compared with the general tracking tasks. Moreover, our extensive experiments show comparable segmentation performance against SiamMask, while running in a real-time speed of 36.2FPS.