An Effective Volleyball Trajectory Estimation and Analysis Method With Embedded Graph Convolution

Volleyball trajectory prediction and analysis based on deep learning has become a hot topic in sports video research. However, due to a large amount of calculation in video processing and the fast speed of volleyball movement with the target scale changing rapidly, these challenges lead to low performance. To this end, this paper proposes an effectively variant YOLOv4 framework to predict and analyze the volleyball trajectory based on video sequences. In the proposed framework, the authors adopt the pre-trained YOLOv4 to select some proposal regions with a high confidence score. Then, the authors embed graph convolution to effectively aggregate deep features. Moreover, to improve the detection and localization capacity of small targets, they introduce a new loss function by modeling the target area with Gaussian distribution. The experimental results show that the proposed framework can effectively prompt the performance of volleyball detection.