Passenger Flow Statistics Algorithm of Scenic Spots Based on Multi-Target Tracking

Abstract

According to the real-time and accuracy requirements of obtaining passenger flow by surveillance videos in scenic spot, a model based on deep learning is proposed. Aided by Yolov4 and Deep Sort, passenger flow is counted by detection and tracking tourists. Aiming at the real-time requirement, the model compression method is used to replace the backbone network of Yolov4 with lightweight network mobileNetv3 to improve the detection speed. For the accuracy of the model, a detection scale is used to Yolov4 for extracting shallow features and the features are concatenated with deep features. Furthermore, Soft-NMS is used to process the detection results. The purpose of these improvements is to solve the dense tourists and small target problems in the surveillance videos. Then Deep Sort tracks the tourists target and obtains passenger flow information in the scenic spot. Through the experiment of the model in the surveillance videos, it is verified that this model meets the real-time requirements and has high accuracy in passenger flow statistics.