Action Recognition of Basketball Players Based on Hybrid Attention Module and Spatial Feature Pyramid Module

Abstract

Watching basketball game videos is an important reference for coaches to analyze the team’s tactics. Detecting the athletes’ actions on the court in real time through an object detection algorithm can help coaches find team problems and formulate solutions. Aiming at the problems of fewer basketball players’ action detection datasets and the difficulty of action detection, this paper proposes a dataset of basketball players’ action detection, BPAD (basketball player action dataset), and an object detection algorithm, YOLOSS (YOLOv4 SimSE and SPPFCSPCG), in which the BPAD dataset consists of 2,151 pictures, which are obtained by extracting the videos of college basketball teams’ matches and annotated with the Labelimg tool, and include three categories, namely, walk, run, and defense. The YOLOSS algorithm is based on YOLOv4, with the more efficient Ghostnet as the backbone of the model, and add a new hybrid attention module SimSE and a new spatial feature pyramid pooling module SPPFCSPCG. YOLOSS can effectively improves the detection algorithm’s ability to recognize the typical basketball actions of athletes in the video stream-run, defense, and walk. The recognition ability of the YOLOSS algorithm on the BPAD dataset is as high as 82.4% mAP50, which can clearly express the action of each player on the court. By comparing the results of various experiments, it is proved that YOLOSS, the object detection algorithm proposed in this paper, can effectively detect the actions of basketball players.