City-Scale Multi-Camera Vehicle Tracking Based on Space-Time-Appearance Features

Abstract

Multi-Camera Multi-Vehicle Tracking (MCMVT) is an essential task in the field of city-scale traffic management, which usually consists of three sub-tasks: object detection and re-identification (ReID), single-camera tracking, cross-camera trajectory association. Compared with existing methods, two challenges are considered and addressed in this paper: (1) low-confidence objects could be missed without extra data annotation, (2) precise association of trajectories from different cameras is affected by multiple factors. For the first challenge, a cascaded tracking method based on detection, appearance features and trajectory interpolation is proposed, exploiting potential real targets in low-confidence objects to improve detection and identification recall. For the second challenge, space, time and appearance features are proposed to be the most crucial factors for trajectory association, so a zone-gate and time-decay based matching mechanism is proposed to adjust original appearance matrix to link tracklets more precisely from different cameras. Extensive experimental results validate the effectiveness of the proposed innovative technologies.