Dense Pedestrian Detection and Tracking Algorithm Based on Improved YOLOv8 and DeepSORT

To address the issues of weak small-target feature response, low bounding box regression accuracy, and poor feature discrimination in pedestrian re-identification within traditional dense pedestrian detection and tracking systems, this paper presents an improved algorithm based on YOLOv8 and DeepSORT. In the detection branch, the Global Attention Mechanism (GAM) is integrated into YOLOv8’s backbone to strengthen feature extraction for small and occluded targets, while the Inner-WIoUv3 loss function is adopted to boost regression precision and convergence. For tracking, the Convolutional Block Attention Module (CBAM) is embedded into DeepSORT’s ResNet50 re-identification network to enhance discrim- inative features under occlusion. Experiments on CrowdHuman and MOT17 demonstrate significant im- provements: on CrowdHuman, Recall reaches 92.6%, mAP50 86.5%, mAP50-95 59.2%, exceeding the base- line by 4.4%, 3.3%, and 4.5% respectively. On MOT17, MOTA, MOTP, and IDF1 achieve 75.2%, 82.7%, and 78.6%, outperforming the baseline by 6.7%, 3.5%, and 8.5%, with IDSW reduced from 956 to 521. By jointly optimizing detection and tracking, the method effectively improves scale adaptability and spatio- temporal robustness in dense pedestrian scenarios.