Research On Object Tracking And Action Recognition Method Based On Vision

As hot research areas in the field of computer vision, visual tracking and action recognition which more and more researchers have been attracted to work on deeply, have been widely used in intelligent surveillance, human-computer interaction, the entertainment and movement analysis etc. Because of the complexity of these problems and the diversity of the scene, it still is a challenging task to get a robust result. Based on the research results in these fields, more methods, such as particle filter, sparse representation, graphical theory, incremental learning and so on, have been presented to solve the problems in the fields of visual tracking and action recognition in this paper.The main works can be summarized as follows:1. Because the dynamic occlusion often causes object tracking and counting errors in a monocular static scene, a two-step object detection tracking and counting method based on the Gaussian mixture model is presented to improve object tracking and counting accuracy under dynamic occlusion conditions. In this paper a two-step Gaussian mixture model detecting method is proposed to detect the object area firstly. Then, the states of merged object and split object are applied to deal with the problem of error counting under dynamic occlusion conditions. Finally, an improved tracker is proposed to complete object tracking and counting. The experimental results show that compared with the traditional methods, the proposed method performs better.2. An incremental learning visual tracking method based on block sparse representation is presented in this paper. Firstly, aimed at the tracking effect of the traditional PCA method which uses global image features is not ideal, a block weighted PCA sparse representation model is proposed to represent the object appearance. Because the method can extract the local feature of objects effectively, it is used to improve the effect of tracking. Then, aiming to the situations, such as the dramatic changes in illumination, object appearance and posture, occlusion and so on, frequently happened during tracking, the incremental learning algorithm which use the block weight to detect object in occlusion and update the observation model is proposed. Finally, to solve the problems of complex calculation and the slow speed in l1 tracker method, the l2 regularized least squares method is adopted to solve sparse representation coefficient. In order to evaluate the new tracking algorithm, we compared it with three popular algorithms in six test video. The qualitative and quantitative analysis results show that the proposed tracking algorithm is better than the three popular algorithms. It can overcome the effect of pose, illumination, occlusion, clutter and so on, and has good robustness.3. Aiming to the stable tracking problem when objects are temporally occluded partly in complicated scene, a particle filter tracking method using graphical model is proposed. This method first constructs particle filter framework by building object observation model through the fusion of color and edge feature. Then, by selecting object feature area, this method establishes the graphical models by dividing an object into several parts, and each part is a vertex of the graph. Finally, the graphical model is applied in particle filtering to complete the object tracking. The experiment results show that the proposed method can achieve stable object tracking by estimating the state of the occluded part on the conditions of part occlusion.4. To track the object stably, a multi-feature fusion tracking method based on mean shift and particle filter is proposed. In this method, by using mean shift method in particle filter method, the degeneracy problem of particle can be overcome. In addition, the color and sift information are adaptively fused to build the object observation model, which can solve the problems such as illumination change, scale changing and similar color area. Experimental results demonstrate that this method is significantly better than the traditional particle filter and mean shift method.5. To solve this problem that the similarity is easily lost after the similar features are coded in the traditional sparse representation, a new locality-constrained group sparse representation action recognition method based on local spatial-temporal features is proposed. In this method, the local spatio-temporal features are first extracted to build locality-constrained group sparse representation model according to the characteristics of local similarity for the human behavior feature in spatio-temporal domain. Meanwhile, the dictionary is constructed and the group sparse representation model is solved. On this basis of these works, we realize the action recognition by using the classifier based on locality-constrained group sparse representation model. Finally, the method is evaluated in two action recognition database. The results show that the new method can obtain good recognition effect.