Motion constraint Markov network model for multi-target tracking
Mingjun Wu, Xianrong Peng · 2008
The typical Markov network for modeling interaction among targets can handle error merge problem, but it suffers from labeling problem due to the blind competition among collaborative trackers. In this paper, we propose a motion constraint Markov network model for multiple target tracking. By augmenting the typical Markov network with an ad hoc Markov chain which carries motion constraint prior, this proposed model can overcome the blind competition for image resources and direct the label to the corresponding target even in the case of severe occlusion. In addition, the motion constraint prior is formulated as a local potential function and can be easily incorporated in the joint distribution representation of the novel model. Finally, this model is inferred within the framework of variational mean field method. Experimental results demonstrate that our model is superior to other methods in solving the error merge and labeling problems simultaneously and efficiently.