Probabilistic Matrix Factorization for Visual Tracking
Xinyi Wei, Zhen-Rong Lin, Tao Liu, Le Zhang · Pattern Recognition and Image Analysis · 2022
Abstract In visual tracking, designing an effective and robust target appearances remains a challenging work due to variations of complicated target appearances and other challenging factors, such as drastic illumination variations, object rotations, partial occlusions, nonrigid deformation, and fast motion. Existing tracking algorithms use previous tracking results as target templates to represent a target candidate. Such target representations are not robust to drastic appearance variations. In this paper, we represent a target candidate as a linear combination of a set of matrix basis. These bases are obtained from a probabilistic matrix factorization, which are learned from previous tracking result examples. The learned matrix basis can capture target structure information and local information. The matrix factorization based on previous tracking result examples over frames can learn target appearance variations in the tracking process. Such target representation is robust to shape variations and outliers. With this kind of target appearances, we propose a novel tracking algorithm in a particle filter framework. Extensive experiments conducted on some challenging video sequences demonstrate the proposed tracking algorithm is effective and robust to outliers and other complicated appearance variations.