Adaptive Object Tracking with Incremental Tensor Subspace Learning
Xinbo Gao · Dianzi xuebao · 2009
The conventional subspaces based tracking methods usually have low precision of object matching and tracking,because they lose the inherent partial structure and neighborhood information.In this paper,an incremental tensor subspace learning algorithm is proposed to model and update the object appearance in tensor subspace.Simultaneously,by combining the proposed learning algorithm with Bayesian inference,an adaptive object tracking method is presented.Firstly,we represented the appearance of the object in tensor subspace;secondly,obtained the optimal estimation of the state parameters by Bayesian inference;finally updated the tensor subspace by using the optimal observation.Due to the construction information is maintained,the proposed method is able to track targets effectively and robustly under pose variation,short-time occlusion and large lighting and so on in the experiments.