Abstract
Abstract This study develops a genetic algorithm (GA)-optimized U-Net framework for sports motion-region segmentation, trajectory extraction, motion anomaly detection, and potential injury-risk indication. The private dataset consisted of 320 sports video clips recorded at 1920 × 1080 resolution and 30 frames/s, covering track-and-field, basketball, and badminton scenarios. To examine external transferability, an additional validation experiment was conducted on the public JHMDB human motion dataset, which provides person segmentation masks and joint annotations. The GA searched the segmentation threshold, convolutional kernel size, stride, filter configuration, loss-weight combination, temporal-stability coefficient, and feature-fusion coefficients within predefined ranges. Based on the optimized segmentation masks, centroid position, principal-axis direction, boundary curvature, joint-angle variation, and displacement increment were extracted from knee, ankle, elbow, and trunk regions. On the private test set, the model achieved a weighted average mIoU of 85.6%, PA of 91.9%, and BF of 87.2%. Compared with the default U-Net, the average centroid error decreased from 3.42 pixels to 2.18 pixels, while anomaly-warning sensitivity increased from 82.5% to 91.2%. The warning labels used in this study were derived from visible motion abnormalities and trajectory changes; therefore, the model should be interpreted as a visual motion-anomaly detector and potential injury-risk indicator, not as a clinical injury prediction tool.