Research On Video Object Tracking Based On Low Rank Sparse

With the development of artificial intelligence, video target tracking technologies have become a research topic in computer vision. And the technologies are widely applied to intelligent monitoring, traffic detection, human-computer interaction, behavior recognition, military, medical and so on. Although the visual tracking algorithms have acquired great progress, due to the existence of various complex scenes,such as the changes of scale and light, background disorder, shielding, rotating, and rapid moving, to name a few, designing a robust tracking algorithm is still a very challenging task.Recently, the tracking algorithms of the low rank sparse representation based on particle filter framework are received extensive attention. The algorithms first use the template dictionary to build the appearance model of target. Then the candidate targets are expressed as a linear combination of the template dictionary. Finally, the algorithms use the sparse representation theory or low rank representation theory to acquire the linear coefficient of candidate particles. In current frame, the candidate particle which has the minimum reconstruction error as the tracking result. The algorithms have been demonstrated that they can effectively handle shielding and the changes of appearance, for a variety of complex scenes have robustness. This paper proposes two kinds of robust target tracking algorithms, which based on low rank sparse representation theory under particle filter framework.(1) For the overall appearance model tracking algorithms easily drift even tracking failure problems in shielding or non-rigid deformation. So this paper proposes a video target tracking algorithm based on local multi-task joint sparse representation. The overall appearance model needs to obtain the complete information of the target, the local pixels appearance model can use the visible part information or small deformation local pixels to determine the target location. The algorithm first partition the each candidate targets, the observation vector of the local pixels of the same location is expressed as the linear combination of the local pixels of the target template. Then we design a local multi-task joint sparse representation model which based on multi-task learning and sparse representation theory. In order to capture the stray particles, the coefficient matrix of the model is decomposed into two different sparse matrixes. And we apply Accelerate Proximal Gradient algorithm (APG) to solve coefficient matrix. Finally, combining with the stray particles detection mechanism and the target template update mechanism achieve the robust tracking. Note that the local pixels contain part background information, so the algorithm gives a lower weight to the edge reconstruction error of the local pixels.(2) Due to single feature description target capacity is weak, different features description target capacity is different. Thus this paper proposes a video target tracking algorithm which based on multi-task and multimedia joint low rank sparse representation. This algorithm first extracts four complementary visual features which include brightness, color, edge and texture. The observation matrix consists of all the same visual characteristics of candidate targets, and the observation matrix is expressed as linear combination of the corresponding characteristics dictionary. Then, we construct a multi-task and multimedia joint low rank sparse representation model which based on multi-task learning and sparse representation theory. In the model, the coefficient matrix is decompressed into the sum of two matrixes, one matrix has low rank and row sparse characteristic, the other matrix has column sparse characteristic. And we apply Imprecise Augmented Lagrangian Method (IALM) to solve coefficient matrix. Finally, combining with the stray particles detection mechanism and the target template update mechanism achieve the robust tracking.The above proposed tracking algorithms were tested in vast amount of challenging videos, and comparing with seven popular tracking algorithms. These results show that the proposed algorithms have better robustness in many complex situations.