| Soccer is considered the world largest sport.It has important practical and commercial value in detecting and tracking objects in soccer videos.In recent years,deep learning technology has been widely used in computer vision tasks and has achieved significant success.Therefore,based on deep learning technology,targeted model improvements were made in this thesis to enhance the detection and tracking performance of the model by analyzing the challenges in current object detection and tracking tasks.The main work of this thesis includes:(1)To address the issues of slow detection rate,occlusion,and small object size affecting detection accuracy in soccer video object detection tasks,the YOLOv5 detection algorithm was improved.Firstly,the backbone structure was simplified,reducing the model’s parameters by 50%.Secondly,multi-scale features were fused,reducing the missed detection rate for small objects.Finally,attention mechanisms were embedded in the network structure to enhance the feature extraction ability of the network and improve the overall detection accuracy of the model.The improved YOLOv5 detection algorithm was validated on a dataset,showing a 17% increase in detection speed and 3% increase in accuracy.(2)To address the issues of fast-moving and irregularly moving player in soccer videos,based on DeepSORT tracking algorithm,the Unscented Kalman Filter algorithm and DIoU matching were introduced to enhance the model’s tracking ability for non-linear moving objects.The improved tracking algorithm was validated on a dataset and showed an improved performance compared to the original algorithm.(3)Building on the foundation of object detection,player teams assignment experiments were conducted in downstream tasks of video analysis.The color features of detected players were extracted,and both similarity measure and clustering algorithms were used for player teams assignment.The experiments showed that clustering algorithms achieved an assignment accuracy of 97%. |
PDF Full Text Request