Multi-camera Vehicle Tracking from End-to-end based on Spatial-Temporal Information and Visual Features

In large-scale traffic video analysis, continuous tracking of vehicles across cameras overcomes the time and space limitations of a single camera, and is conducive to transportation design and traffic flow optimization. In this work, we propose an end-to-end framework for multi-camera vehicle detection, tracking and re-identification in complex traffic environments with urban multi-junctions, which integrates visual features and temporal-spatial information of the trajectories for optimization. Based on detection and tracking of multi-vehicles in a single camera, our method distinguishes and marks the vehicle trajectories from different intersections where they enter and exit. Then, the visual features of the same vehicle keyframes are extracted to match between the cameras of the specific matching link, while taking into account the constraint of the trajectory time. In the end, our algorithm shortens vehicle trajectories' average matching time in two cameras to 2 seconds, and the accuracy is 81.59% in the test scenarios, which greatly improves the efficiency and accuracy of vehicle re-identification.