Object Tracking via Tensor Kernel Space Projection

Although there has been significant progress in the past decade, object tracking under complex environment is still a very challenging task, due to the irregular changes in object appearance. To alleviate these problems, this research developed an object tracking algorithm via tensor kernel space projection. In the initial stage of tracking, a template matching algorithm was used to obtain a priori images of the appearance of the object. The steps taken were as follows: define the tensor kernel function based on a multi-linear singular value decomposition, view the object appearance color image as tensor data, calculate the kernel matrix for the priori appearance image samples, use KPCA to obtain the projection matrix of the image samples in kernel space, and finally, obtain an optimal estimate of the object state through Bayesian sequence interference. Meanwhile, the projection matrix in kernel space was updated on-line. Experiments on two real video surveillance sequences were conducted to evaluate the proposed algorithm against two classical tracking algorithms both qualitatively and quantitatively. Experimental results demonstrate that the proposed algorithm is robust in handing occlusion and object scale changes.