A Method for Top View Pedestrian Flow Detection Based on Small Target Tracking

Abstract

In public spaces, monitoring pedestrian flow can effectively avoid the occurrence of crowding and stampede incidents, and can effectively improve public safety. To improve the accuracy and efficiency of small target tracking in pedestrian flow detection from a top-down perspective, this study integrates the Vision Transformer architecture and the Deep SORT tracking algorithm to improve the YOLOv5 model. This method aims to achieve efficient detection in complex pedestrian flow scenarios by enhancing the recognition ability and tracking accuracy of small targets. In the experiment, the improved YOLOv5-V-D model quickly converged after 61 iterations, achieving excellent operational efficiency with an average delay of only 7.2ms. Compared to CenterNet and RetinaNet, it has increased by 3.9ms and 6.4ms, respectively. In addition, the model performed outstandingly in terms of accuracy in predicting pedestrian flow, reaching 98.72%, which is superior to the comparison model of 20.59%-28.61%. In summary, the improved YOLOv5 model not only provided faster detection speed, but also significantly improved the accuracy of pedestrian flow detection. This progress provides an effective solution for high-density crowd monitoring, laying a solid foundation for the development of future real-time monitoring systems and effectively improving public safety.