In order to overcome the disadvantage that traditional subspace methods usually lose the two-dimensional information of the objects in image,a novel adaptive object tracking method is proposed.The appearance of the object in tensor subspace is modeled and the object model is updated with online learning method.The object is tracked by using particle filter and the prior of affine motion,and the optimal observation is feeded back to the tensor subspace updating.Moreover,DPF is introduced into the subspace updating to reject outliers so as to keep the object subspace precise and compact.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.