Research On Object Long-term Tracking Algorithm With Sparse Appearance Model Learning

Visual tracking is a hot topic in computer vision. However, since the complex tracking environment and the motion uncertainty, there are still lots of the challenging problems to be solved. Recently, to keep the long-term tracking, many machine learning methods have been introduced to the tracking problem such as the subspace learning method, metric learning method, manifold learning method and deep learning et al, which make appearance model learn the variations of the tracked object well(such as the severe partial occlusion, drastic illumination changes, abrupt motion and motion blur et al.)so as to keep a better tracking performance. Meanwhile, the observation model is used to obtain the optimal tracking states from the effective object candidates.Such visual tracking method with appearance model learning manily includes three aspects: coustucting appearance model, designing observation model and providing a set of object candidates with the help of motion model and sampling mechanism, which is a new research trend. In this dissertation, under the tracking framework based on appearance model, the research is focued on two main aspects: one is how to design an appearance model to represent the object, the other is how to improve the efficiency of the object candidates. Our main contributions are described as follows:1. A novel visual tracking algorithm via Incremental Non-negative Matrix Factorization(INMF) and dual1l-norm constraints is proposed. The method constructs the appearacne model by combining the NMF subspace representation model and sparse representation model. The non-negative constaint is introduced to matrix factorization, and the atom in traditional sparse representation is replaced with the basic vectors. Then, an iterative algorithm is used to solve the optimization problem, which ensures the representation is more robust. Finally, we take partial occlusion into construct likelihood function, and combined with INMF learning to update appearance model for alleviating tracking drift. The proposed algorithm achieves favorable performance when the object undergoes large occlusion, motion blur and illumination changes.2. A novel visual tracking algorithm via robust multi-task sparse prototypes is proposed. The method improves the object representation by mining the inter-relation between basic vectors, and extends the visual tracking algorithm with sparse prototypes in multi-task learning framework. In addition,to avoid the problem that enforcing all subtasks to share the same structure may result in degraded tracking results, we impose group sparse constraints on the coefficients of PCA basis vectors and element-wise sparse constraints on the error coefficients, respectively. Finally, we show that the proposed optimization problem can be effectively solved using the accelerated proximal gradient method with fast convergence. The proposed algorithm achieves favorable performance when the object undergoes partial occlusion, motion blur, and illumination changes.3. A novel sparse visual tracking algorithm based on the motion prediction using the extended SIFT flow is proposed. The method extends the sparse tracking algorithm within the multi-tracker framework, and firstly introduce SIFT flow to tracking problem. Meanwhile, to cope with the large motion displacement between consecutive images frames, SIFT flow tracker is proposed. Moreover, to make our algorithm adapt target’s appearance variations, especially to partial occlusion, the extended 1l tracker designs a hybrid mechanism of sampling two predicted results, which improves the efficiency of object candidate sampling. The proposed algorithm achieves favorable performance when the object undergoes partial occlusion, motion blur, and abrupt motion.