A Lightweight DeepSORT-Based Model for Pedestrian Tracking

With the rapid development of intelligent surveillance, pedestrian tracking technology has become a significant research direction in the field of computer vision. However, traditional pedestrian tracking algorithms face critical issues of tracking accuracy and stability. Existing mainstream models often adopt complex architectures with high computational complexity to meet high-precision requirements, leading to high computational costs. This paper proposes a lightweight DeepSORT pedestrian tracking model. First, a detection network based on the YOLOv5s model is designed. By integrating the CBAM attention mechanism, it focuses on more important features of pedestrians from both spatial and channel dimensions, reducing interference from redundant information in pedestrian detection. Second, a novel person re-identification network is proposed to replace the CNN feature extraction network in the DeepSORT object tracking algorithm. This network uses the Shufflenetv2 basic architecture and optimizes it by introducing the ECA attention mechanism to compensate for accuracy loss caused by lightweighting. Experimental results show that the detection network designed in this paper improves model accuracy with almost no increase in computational complexity. The proposed person re-identification network achieves a mean average precision of 73.2%, which is 8.5% higher than the pre-improvement CNN network, while reducing the number of parameters by 2.9 million and floating-point operations by 642.02 million FLOPs. The person re-identification network presented in this paper can achieve a better balance between lightweighting and model performance.