Online Multi-Object Tracking With Visual and Radar Features

초록

Multi-object tracking (MOT) constructs multiple object trajectories by associating detections between consecutive frames while maintaining object identities. In many autonomous systems equipped with a camera and a radar, an amplitude and visual features can be measured. Therefore, our goal is to solve a MOT problem by associating detections with both features. To achieve it, we propose a unified MOT framework based on object model learning and confidence-based association. For improving discriminability between different objects, we present a method to learn several visual and amplitude object models during online tracking. By applying the learned object models for the affinity evaluation, we improve the confidence-based association further. In addition, we present a practical track management method to initialize and terminate tracks, and eliminate duplicated false tracks. We implement several MOT systems with different object model learning and association methods, and compare our system with them on challenging visual MOT datasets. We further compare our method with the recent deep appearance learning methods. These comparisons verify that our method can achieve the competitive tracking accuracy while maintaining a low MOT complexity.

키워드