A Lightweight and Accurate Method for Detecting Traffic Flow in Real Time

Traffic flow detection provides significant information for intelligent transportation systems. However, as the mainstream research direction, vision-based traffic flow detection methods currently face the challenges of a trade-off between accuracy and speed. Furthermore, it is crucial that modularization be incorporated into the system design process to enhance the maintainability and flexibility of the system. To achieve this, we propose a modular design method that divides this task into three parts: vehicle detecting, vehicle tracking, and vehicle counting. As an important link of the system, vehicle detection greatly influences the accuracy and speed of the system. We therefore introduce a lightweight network called feature adaptive fusion-YOLOX, which is based on YOLOX. Specifically, in order to eliminate redundant information brought by bilinear interpolation, we propose a feature-level upsampling method called channel to spatial, which enables upsampling without additional calculations. Based on this module, we design a lightweight, multi-scale feature fusion module, feature adaptive fusion pyramid network (FAFPN). Compared with PA-FPN, FAFPN reduces FLOPs by 61% and parameters of the neck by 50% while maintaining comparable or even slightly improved performance. Through experimental tests, the traffic flow detection method proposed in this paper achieves high accuracy and adaptability in a series of traffic surveillance videos in different types of weather and perspectives and can realize traffic flow detection in real time.