A football player tracking model combining object detection algorithm and YOLOv3 algorithm

Abstract

With the increasing demand for analyzing accuracy and real-time performance in soccer games, traditional athlete tracking algorithms are challenged to perform in complex scenes. The main objective of the research is to cope with the challenges such as occlusion and lighting changes in complex scenes by improving the detection speed and accuracy. To this end, a novel soccer player tracking detection model is investigated. First, channel pruning and layer pruning techniques are used to optimize YOLOv3, which reduces the computational complexity and the number of parameters, while ensuring high detection accuracy. Second, the feature enhancement module and offset sampling mechanism are designed in combination with the improved DeepSORT multi-target tracking algorithm. By combining the Kalman filter and the Hungarian algorithm, the accuracy of target prediction and trajectory association is further improved. The experiment outcomes show that on the publicly available football network version 2 dataset, the F1 value of the model reaches 94.56%, and the average processing time is only 0.74 s. On the football player tracking and recognition dataset, the F1 value of the model is 95.27% and the processing time is 0.77 s. In addition, the loss rate is the lowest in occluded scenes, at 8.34%, and the highest trajectory consistency reaches 92.59% in lighting changing environments. From this, the model proposed by the research is significantly superior to existing methods in multiple key performance indicators, and has excellent detection capabilities and practical application potential. It can provide certain technical support for the analysis of penalty decisions and technical guidance in future football matches.