Multi-object Tracking under Occlusion Using Dual-Mode Graph Embedding

Jin Gao, Junliang Xing, Weiming Hu, Xiaoqin Zhang, Ruiguang Hu

Abstract: In this paper we address the problem of tracking multiple objects to know how objects are moving (e.g. occlusion relationships) while interacting with each other in a group, given models of their appearances that are learned online even when occlusion occurs. This aim is very different from the recently popular detection-based tracklets association approaches. In our approach, occlusion relationships between multiple objects are explicitly defined and deduction of the occlusion relationships is integrated into the whole tracking framework. Specifically, we deduce the joint state estimation problem in the multi-object tracking in a new decentralized strategy, that the single object tracking and the multi-object separating are viewed as one-versus-rest classification problems based on graph embedding framework. Two kinds of discriminative subspaces are learned: one for single object tracking which is robust to various appearance variations, the other for occlusion reasoning and decentralizing. Partial disappearance can also be addressed as an occlusion problem by this strategy. Experimental results demonstrate the effectiveness of our method.