A Deep Learning Framework for Video-Based Vehicle Counting

Traffic surveillance can be used to monitor and collect the traffic condition data of road networks, which plays an important role in a wide range of applications in intelligent transportation systems (ITSs). Accurately and rapidly detecting and counting vehicles in traffic videos is one of the main problems of traffic surveillance. Traditional video-based vehicle detection methods, such as background subtraction, frame difference, and optical flow have some limitations in accuracy or efficiency. In this paper, deep learning is applied for vehicle counting in traffic videos. First, to solve the problem of the lack of annotated data, a method for vehicle detection based on transfer learning is proposed. Then, based on vehicle detection, a vehicle counting method based on fusing the virtual detection area and vehicle tracking is proposed. Finally, due to possible situations of missing detection and false detection, a missing alarm suppression module and a false alarm suppression module are designed to improve the accuracy of vehicle counting. The results show that the proposed deep learning vehicle counting framework can achieve lane-level vehicle counting without enough annotated data, and the accuracy of vehicle counting can reach up to 99%. In terms of computational efficiency, this method has high real-time performance and can be used for real-time vehicle counting.

Automated summary

This paper applies deep learning for vehicle counting in traffic videos. A method based on transfer learning is proposed to solve the problem of lacking annotated data. A vehicle counting method based on fusing the virtual detection area and vehicle tracking is proposed. Suppression modules are designed to improve the accuracy of vehicle counting.