Design of a Multi-Object Wearable Recognition and Tracking Algorithm Based on PHSM-YOLO

To address issues such as missed detections and false positives in safety equipment wear detection tasks caused by complex backgrounds, severe occlusions, and small target sizes, as well as frequent ID switching and low tracking accuracy during personnel tracking due to occlusions and drastic changes in target scale, this paper proposes a multi-object wear detection and tracking algorithm based on an improved YOLOv11 and Byte Track. During the detection phase, a novel C3k2_PPA architecture is constructed by introducing a parallel patch-aware module. This architecture leverages a multi-branch parallel structure to enhance feature representation capabilities for multi-scale objects and improve the fusion of local and global information. It effectively suppresses background interference while enhancing small object features. The lightweight downsampling module HWD is introduced to reduce redundant model parameters while preserving more useful information, further boosting the model’s feature expression capabilities. At the backbone network’s terminal layer, the MPCA attention mechanism captures cross-scale contextual information through multi-scale attention, improving small object spatial localization. Finally, a new detection head incorporating a separation and enhancement attention module is introduced to boost feature extraction capabilities. During the tracking phase, the aspect ratio in the state variables is replaced with width w, and the state update and observation models are redesigned to better align with target motion characteristics. The covariance matrix is adjusted, and noise estimation is unified using height h as the scale. Finally, DIOU replaces the traditional IOU for data association. Experimental results demonstrate that the improved YOLOv11 model achieves a 2.62% increase in mAP50, a 10.5% reduction in parameters, and a 7.9% decrease in computational load, exhibiting enhanced robustness and detection accuracy. The enhanced Byte Track algorithm achieves an average MOTA of 71.69%, average MOTP of 76.89%, and an average of 5 ID jumps per test data point, validating the superiority of this algorithm and its practical applicability.