The Occlusion Handling And Research Of Object Tracking In Dynamic Scenes

Object tracking in video sequences is one of the important research directions in the field of computer vision. However, target occlusion is a common but difficult problem in video tracking. Whether target occlusion could be solved efficiently or not is playing an important role in improving the accuracy and robustness of target tracking algorithm. Aiming at the occlusion problem in target tracking, two tracking methods are proposed by incorporating sparse representation and deep learning in this paper. The quantitative analysis of the proposed algorithms are carried out through the evaluation system of the tracking algorithm. Meanwhile, the real-time target tracking and evaluation are completed in the visual cloud platform. The main research work is as follows:On the basis of deep learning, a target tracking algorithm based on the Stacked Sparse Autoencoders(SSAE) is presented. First, this algorithm is pre-training the SSAE and Logistic Regression Classifier by using the greedy layer-wise training. And then it is using the Back Propagation to fine-tune the whole network. In the particle filtering framework, we extract the characteristics of particles by unsupervised training, and use the Logistic Regression Classifier to select the particle with the highest Confidence Value as the tracking result, which has succeeded in achieving the robustness of tracking.Based on sparse representation and particle filter theory, the Block-Sparse Representation and HSV Color Model are proposed. First, we obtain a dictionary from the first frame through using the online dictionary learning method. Then we fuse the features of the block sparse and the HSV color histogram, making the extracted information have both the local and global features. We select the particle with the maximum likelihood as the tracking result. What’s more, we deal with the occlusion problem efficiently in the process of tracking and apply a practical template update method, which has improved the robustness of tracking.The quantitative evaluation of the proposed algorithms and other popular algorithms are carried out by the evaluation system of the tracking algorithm. The results show that the two tracking algorithms proposed in this paper have obvious advantages compared with the current popular algorithms. Both of them could achieve robust tracking under various types of occlusions. Among them, the Block-Sparse algorithm and HSV color histogram have lower Center offset error as well as faster tracking speed. And both of the two algorithms presented in this paper turned out to be excellent in aspect of tracking success rate.At the end of this paper, the visual tracking system is introduced. Meanwhile, the real-time tracking test of the proposed Block-Sparse algorithm and HSV color histogram as well as other popular algorithms are carried out by using the visual tracking system. The test results show that the tracking algorithms proposed in this paper have better performance than other algorithms in aspect of real-time tracking. They could achieve stable robust tracking and are more suitable for non-rapid moving objects’ tracking.