Multi-object tracking by expanding long-tracklets

Multi-object tracking has made great progress recently by using global data association approaches which optimize all tracks simultaneously. But occlusions, false alarms and miss detections still cause many problems, e.g. track fragments, identities switches and track incompleteness. In this paper, we propose a three-stage hierarchical tracking framework based on global association, which can alleviate the above problems effectively. Firstly, short but reliable tracks (tracklets), are built using a globally-optimal association method under network formulation. Secondly, online boosting and a novel bidirectional trackelts similarity metric are introduced for tracklets association, which can naturally handle complete occlusion and miss detection. Finally, a tracking-by-long-tracklets approach using single object trackers is proposed to expand the linked tracklets (long-tracklets) generated in the second stage. Experiments on static and moving camera datasets verify our method.