关键词: 行人定位；YOLOv5；融合注意力机制；mAP@0.5:0.95

Abstract: To address the issues of low accuracy and difficulty in achieving tracking localization in current vehicular binocular vision systems, a pedestrian localization and tracking method based on improved YOLOv5s is proposed in this study. Target localization and tracking functionalities are designed in the binocular vision-based YOLOv5 object detection network. Additionally, a Transformer-based adaptive fusion attention mechanism is incorporated into the main network to enhance target feature representation. The network structure of the feature fusion module is modified to improve fusion accuracy. The improved YOLOv5s model is named YOLOv5sBCT, which exhibits approximately 4.20% enhancement in precision and about 2.91% improvement in mAP@0.5:0.95 compared to the baseline YOLOv5s model. Furthermore, the measured distance error rate is reduced by approximately 6.19%. Consequently, this study provides an effective method for enhancing pedestrian tracking localization accuracy in vehicular intelligent systems, thereby enhancing driver safety and driving performance and promoting advancements in the field of intelligent transportation.

Key words: pedestrian localization; YOLOv5; fusion attention mechanism; mAP@0.5:0.95