Visual object tracking based on incremental kernel PCA

This paper provides a novel method for visual object tracking based on incremental kernel principal component analysis. The proposed method is particularly robust in the case that the tracking object performs pose variation or there exists occasional occlusions. The whole method is constructed in the framework of particle filter and the state of object is defined by the position and shape of a parallelogram, with which the tracking result is located in every frame. For every particle, we compute its reconstructed preimage based on KPCA which can regressively estimate the de-noising pattern in the input space constructed by located objects in previous frames. The difference between the preimage and its original patch is finally adopted to measure the particle weight. Compared to the other state-of-the-art methods, the proposed method can cope with occasional occlusion and pose variation without significantly increasing the computation.