Robust object tracking based on FREAK and P3CA

We proposed a novel tracking algorithm based on the Fast Retina Keypoint (FREAK) and Principal Component-Canonical Correlation Analysis (P3CA). The proposed FREAK-based multi-mode dynamic model improves the prediction accuracy of the object location, reduces the searching space. P3CA-based appearance model is more robust in handling occlusion than holistic information based appearance model due to the adoption of the canonical correlation between sub-patches in an image, and the integration of principal component analysis (PCA), which is very excellent in data dimension reduction, successfully solves the small sample size problem and reduces the computation cost in the generation of canonical correlation analysis (CCA) subspace. Meanwhile, the tracker can deal with the appearance variations with time thanks to the novel online updating method for P3CA subspace. The comparison experimental results on several challenging video sequences demonstrate that our algorithm can cope with the appearance variations caused by illumination changes, occlusion, rotation and background clutters etc. and performs better than some state-of-the-art methods according to the tracking accuracy.