A Real-time Framework of Multi-Camera Vehicle Tracking System*

Hanyang Zhuang,Xingyuan Chen,Jianglin Lan,Ying Wang,Zhaoxuan Lu,Chunxiang Wang,Ming Yang

DOI: https://doi.org/10.1109/rcar61438.2024.10671316

2024-01-01

Abstract:Multi-camera vehicle tracking (MCVT) system is a key technology to build an intelligent city and intelligent transportation system. The MCVT system utilizes roadside monitoring devices and computing platforms to achieve city-level road traffic perception at a finer scale by tracking each vehicle on the road to obtain the trajectory. The existing MCVT system mainly focuses on dealing with the algorithm modules separately, without studying the real-time performance and system architecture of the entire pipeline. However, in real-world applications, many of the tricks in improving evaluation metrics are not applicable since there is endless traffic flow running all the time. Therefore, this paper proposes a real-time framework of the MCVT system to streamline all the algorithm modules with proper data flow. Firstly, a multi-threaded multi-vehicle tracking (MVT) module is developed to detect vehicles in each camera. Re-identification is implemented during MVT to provide the feature for tracking in the same camera and across different cameras. Then, after the vehicle leaves the camera field of view, the feature of vehicle trajectory is calculated by averaging the feature at each frame. Meanwhile, a background thread runs continuously to match the incoming feature of vehicle trajectory with past features by conducting a topology searching method. In the end, the vehicle trajectories with well-matched features are considered to be the same vehicle. The entire pipeline has been validated on two different datasets, the precision is 0.64 in the CityFlow dataset which uses 6 cameras simultaneously. In the customized dataset, the same method can still reach real-time when using 16 cameras at the same time with smaller network modules.