Enhanced Faster R-CNN with Attention Mechanisms for Multidi- mensional Soccer Player Performance Assessment

Abstract

In this study, we propose a novel framework that optimizes the Faster R-CNN algorithm and constructs a multidimensional evaluation model for soccer player performance. Specifically, we redesign the ResNet- 50 backbone, integrate Feature Pyramid Networks (FPN), and embed SE and CBAM attention modules to enhance feature extraction in dynamic match environments. The enhanced model extracts key motion and spatial features from a dataset of 6,000 annotated match images, achieving a mean Average Precision (mAP@0.5) of 84.3%, precision of 89.1%, recall of 86.2%, and F1-score of 87.6%, outperforming base- line Faster R-CNN (mAP@0.5 = 75.8%) and YOLOv5 (mAP@0.5 = 79.3%). Our model also achieves mAP@0.75 of 72.4% and mAP@50:95 of 68.9%. Building on robust detection outputs, we develop an evaluation system across physical performance, technical skill, and tactical execution, each quantified through expert-defined indicators and weighted scoring. Validation on diverse match scenarios shows high correlation (r = 0.91, p < 0.001) with expert assessments and effective identification of fatigue and tactical behavior variations. This approach provides a data-driven tool for intelligent performance as- sessment and lays the groundwork for athlete monitoring and tactical planning in soccer.