Multi-Target Multi-Camera Vehicle Tracking for City-Scale Traffic Management

Abstract

Multi-target multi-camera (MTMC) tracking is one of the important fields in computer vision, where multiple objects are tracked across multiple cameras. MTMC tracking can be applied to various tasks such as crowd analysis, city-scale traffic management, and transportation systems analysis for intelligent city planning. However, it is challenging due to the large variety of conditions of each camera, such as perspective and illumination. Furthermore, MTMC tracking for vehicles is more problematic because of the relatively large inter-class similarity and intra-class variability. In this paper, we tackle the MTMC tracking problem for vehicles by dividing it into three main steps: (i) vehicle detection and feature extraction, (ii) multi-target single-camera tracking using the appearance feature of each vehicle, and (iii) multi-camera association of local trajectories from each camera. Our method shows comparable results with other highly-ranked methods in AI City Challenge 2021 and outperforms a recent MTMC tracking method that ranked first place in AI City Challenge 2020.