Bus Passenger Flow Detection Model Based on Image Cross-Scale Feature Fusion and Data Augmentation

Objectives In order to resolve the issues of large size gap, serious occlusion and overlapping images in the objection detection research of bus passenger flow, an improved object detection model YOLOv5s_P is proposed.

Methods The PANet structure in YOLOv5 model is replaced with the BiFPN structure to strengthen features information of different scales maps in order to extract complex target features. At the same time, Mixup data augmentation method is used to increase the training samples of occlusion and overlapping images to improve model generalization capabilities and to reduce the detection errors caused by the fragmentation of passenger flow images. In order to verify the performance of the YOLOv5s_P model, it is compared with four other models: Faster R-CNN, SSD300, RetinaNet and YOLOv5s to detect bus passenger flow in real bus scenario, in which the image sets are labeled to detect the upper body of the human instead of the head.

Results Experimental results show that the average accu⁃racy of YOLOv5s_P model reached 96.9% without affecting the detection speed, and the average missed detection rate was reduced by 3.43% compared with YOLOv5s model, which improved the detection accuracy of bus passenger flow.

Conclusions In the future, research will be integrated with the attention mechanism to further improve the detection accuracy, and combined with the tracking algorithm to solve the problem of the bus passengers number fluctuation.