The Perspectives Of Target Tracking Technology Based On Feature Space Research

Visual object tracking is one of the focuses of research in the field of computer vision. It has a wide range of applications in the field of video surveillance, intelligent traffic monitoring, human-computer interaction, video data retrieval and so on. Compared to single-camera system, multi-camera system can provide a lot of information redundancy to improve the tracking robustness. In recent years, multi-view object tracking is under great concern.Visual object tracking faces many challenges. The complexity of the tracking scene is the basis problem, including background interference, appearance variation, occlusion, illumination changing and so on. The following two aspects need to be done to achieve accurate object tracking in complex scenes. First, the appearance model of object must be built up effectively. It is the basis of model matching and updating in tracking algorithm, but the model updating can easily lead to model drift in complex scenes. Secondly, multi-view information must be fused effectively, including the integration of multi-view target position information and the integration of multi-view appearance models.In order to track object under complex scene, a distributed algorithm framework of Bayesian Inference is introduced and improved in this paper firstly. It can solve the problem of multi-view data fusion, especially the fusion of multi-view position information. The framework uses Bayesian graph model to effectively derive the posterior probability distribution of the target state (or location) from multi-view data, and use the particle filter method to approximate the posterior probability distribution. At the same time, the SSIM index and Kalman filter are adopted in the process of modal matching and updating. The experimental results show that the algorithm framework can effectively fuse multi-view position information, and track objects efficiently, robustly in real time under occlusion scenes. However, the algorithm framework doesn’t solve the problems of model drift and multi-view model fusing. Two algorithms were proposed to solve these two problems as below.Under the distributed Bayesian framework, a new adaptive multi-view subspace tracking algorithm is proposed for the model drift in model updating process. The algorithm updates the appearance model in subspace using the likelihood of the sample in order to eliminate the model shift. It processes and fuses the data in distributed way on different views under the Bayesian tracking framework, and use multi-part appearance model for matching and updating to achieve more accurate tracking result. Experiments show that the algorithm can solve the problem of model drift and track the object effectively, accurately especially under the scenes of occlusion and appearance variation.Then, a multi-view model fusing algorithm based on sparse space is proposed under the distributed Bayesian framework to solve the problem of multi-view model fusing. Sparse space is linear space for sparse representation. The algorithm updates and fuses templates in multi-view as the basis of the sparse space, and the object can be represented with sparse coefficient in this sparse space. The simulation results show that the algorithm can not only avoid the problem of model drift, but also effectively carry out multi-view model fusing.