Visual Tracking Based On Tensor Subspace Learning

Visual tracking is one of the most important topics in computer vision and artificial intelligence. It is not only the necessary precondition of the semantic-based video analysis, but also the basis of the content-based video processing. The key point in visual tracking is how to represent the object of interest efficiently, and exclude the influence, such as illumination and occlusion, in the video environment. Based on the traditional methods in visual tracking, we propose several new methods in the framework of Bayesian inference, by using tensor subspace, incremental/online learning and so on, in order to get the accurate description of the object. The main contributions of this paper are listed as follows.(1) An incremental tensor subspace learning based object tracking is presented. In the proposed algorithm, firstly the object appearance is modeled by initializing the tensor subspace, and extended to an incremental manner. Then the optimal estimation of the object state parameters is obtained by Bayesian inference. Finally, based on the optimal estimated object observation, the tensor subspace is updated incrementally. The proposed method is able to track targets effectively and robustly under pose variation, short-time occlusion and large lighting and so on in the experiments.(2) A novel object tracking method based on weighted Retinex tensor subspace is proposed. In this part, firstly the object is represented by the weighted Retinex tensor data, which could release the sensitivity of the traditional tensor to the illumination. Then an object appearance subspace is constructed based on the third order tensor for updating the appearance subspace online. Thus the state parameters of the object in the video can be predicted and estimated according to the observation model. Since both the original and Retinex images are kept in the appearance model, the proposed method could get the satisfactory results under the condition of drastic light.(3) A biased discriminant subspace selection is developed to track the object of interest on the basis that the tracking can be thought as the online classification problem as the time shifts. Firstly the traditional linear and biased analysis are extended to the tensor manner. Then an incremental biased discriminant tensor subspace analysis is presented. Finally, the object could be tracked accurately based on the biased discriminant analysis. The experimental results illustrate that the proposed method is able to track object under the complex background, pose and scale variation and so on.(4) An object tracking algorithm based on the pairwised constraint is desired. The relative appearance variation between the object and background is taken into account in this algorithm. Firstly, the object is modeled into an object-background pair. The traditional methods usually focus on the object, while the proposed algorithm puts forwards to the object-background, which could concern about both the object and background. Secondly, the consistency constraint of the appearance subspace is introduced when updating the subspace, so as to keep the model change stably, from the outlier due to the drastic variation in appearance. Thirdly, a pairwised constraint based object observation model is constructed to predict and estimate the object state. The proposed method could track object robustly for long time under the condition of the pose variation, complicated background.In this thesis, we apply and extend the machine learning method in the application of the object tracking. The proposed methods could get satisfactory results undergoing the condition such as occlusion, complex circumstance and drastic illumination, which can enrich the algorithms and applications of the effective and efficient object tracking.