Research Of Low-Rank Sparse Representation Based Object Tracking Algorithms

The visual tracking algorithms have got great progress in the past decades under the follow-up research of lots of researchers. However, object’appearance in the complex environment usually would be changed by various factors, such as, occlusions, varying illumination, shape deformation, background noise. That to develop a visual tracking algorithm with high accuracy, stability and robustness is also a very challenging work. This thesis proposes two kinds of low-rank sparse representation based object tracking algorithms on the basis of the existing algorithms.We firstly introduces the low-rank sparse representation model into the generative tracking framework, and then proposes a principle component analysis(PCA) basis vectors template and square template based low-rank sparse representation model. The model has fully used the advantages of PCA basis vectors template and square template using to build the dictionary matrix in the object appearance model. Moreover, low-rank constraint can well reflect the structure information between the candidate particles. In the fact, different object tracking result has different influence to subspace update. So a weight based on target reconstructed error is introduced into every tracking result. Then this new result can be used to update the target representation model with incremental PCA algorithm. Finally, a low-rank sparse representation based weighted incremental object tracking algorithm is proposed.A joint coefficient matrix and low-rank sparse representation based object tracking algorithm is also proposed in this thesis. In this algorithm, it proposes a novel low-rank sparse representation model based on joint coefficient matrix. In the process of multi-task learning, the coefficient matrix is decomposed into the sum of two components. One component uses to capture the shared features among tasks and two constraints, low-rank constraint and sparse constraint, is imposed on this component. The second component uses to discover the outlier tasks, which only sparse constraint is imposed on. By embedding this novel model into the particle filter framework, a new object tracking framework is proposed. Considered the problem of particle degradation in the particle filter, all candidate particles in every frame have to be resampled. Compared with the existing tracking algorithms, it has improved the particles’succession.Both qualitative and quantitative evaluations on most challenging and international open videos demonstrate that the proposed algorithms have better robustness of object tracking in the complex environment.