Pose Feature Fusion Based Short Track Speed Skaters Multi Object Tracking

The multi-object tracking is more challenging due to the large-scale changes, frequent occlusions,motion blurring, and similar appearance of the skaters during the skating. A multi-object tracking dataset SSSMOT is constructed in the short track speed skating scene, and a multi-object tracking method is proposed based on pose information. The SSSMOT dataset is made up of 37 videos of short-track speed skaters competing. Firstly, the multi-object tracking method optimizes the YOLOv5 detection model in terms of Anchor, loss function, and non-maximum suppression to improve the detection accuracy. Secondly, the feature extraction network PRNet is de-signed to extract features based on the pose information to improve the feature robustness. Finally, the data association matching method is improved by using the pose key point to alleviate the matching error caused by the similar appearance of athletes to a certain extent. The accuracy of multi-target tracking is 96.43% when compared to other approaches in the SSSMOT and SKMOT datasets,and additional assessment indicators and ablation experiments show the effectiveness and superiority of the proposed method.