Object tracking algorithm via incremental orthogonal projective non-negative matrix factorization
1. College of Civil Aviation, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu 210016, China; 2. Key Lab of Aviation Information Technology in Universities of Shandong, Binzhou University, Binzhou, Shandong 256603, China
Abstract：To avoid the tracking performance degradation of traditional trackers by occlusion, scale change and illumination change in videos, an object tracking algorithm was proposed based on incremental orthogonal projective non-negative matrix factorization(IOPNMF) and L1 norm. The L1 norm was introduced into IOPNMF subspace for reconstruction to ensure the learned part-based appearance model to tolerate different noises. Another L1 norm was enforced on orthogonal projection coefficients by iterative operators to obtain IOPNMF basis vectors, which enabled the proposed algorithm to tackle the dynamic video stream with robust tracking. The tracking object was represented by linear combination of IOPNMF basis vectors, and the partial occlusion factor was introduced into the observation model to update the selective IOPNMF subspace. The proposed algorithm was realized by MATLAB, and the comparison experiments with six other algorithms were conducted on challenging videos with occlusion, illumination change, scale change, motion blur and background clutter. The experimental results show that the proposed algorithm can achieve good robust and stable object tracking with the lowest center location error of 4.3 pixels and the highest overlap rate of 0.84.
2016, Vol. 37 
