Video Traffic Volume Extraction Based on Onelevel Feature

In this paper, a single-level feature detector network based on YOLOF is built to detect objects and extract traffic volume information for videos. YOLOF does not have a multi-scale feature fusion structure (such as FPN, PAN). It uses the dilated encoder and uniform matching method to replace the fusion feature module, which simplifies the network structure and improves the response time while maintaining good accuracy. Finally, the target tracking algorithm is used to obtain the traffic flow statistics in the video sequence. Generally speaking, YOLOF has higher detection performance and faster speed than RetinaNet, DETR and other networks of similar structure.