Hierarchical Data Association Framework with Occlusion Handling for Multiple Targets Tracking

Abstract

The problem of tracking multiple targets in video is addressed, and a novel hierarchical data association framework with occlusion handling is presented. The association hierarchy first divides the detections, targets, and candidates into several disjoint branches, then progressively associates the detections to the targets and candidates within each branch, and finally initializes the targets through candidate upgrade. Furthermore, the depth disorder inference and targets motion pairing prediction are introduced to explicitly tackle target-target and target-environment occlusions, respectively. Experiments verify that our method improves the runtime performance significantly while keeping competitive tracking accuracy and precision compared with several state-of-the-art methods.