英文摘要 | In recent years, video cameras are widely used for monitoring in intelligent transportation systems (ITS) and intelligent video surveillance technology has become an important tool for the collection of transportation data, leading to the increased demand for automatic video analysis. How to track and recognize visual objects is an important and also a basic issue in video surveillance, which has been widely studied. However, there are still some problems to be solved in in urban traffic scenes. On one hand, due to the problem of illumination, weather, shadow, occlusion, posture and object-interference, it is changeable to track and recognize the target in complex traffic scenes. On the other hand, monitoring systems are expected to be more functional, in which more accurate and abundant information should be provided. Therefore, it is of important theoretical significance and practical value for us to do the research on visual tracking and recogntion. In order to be able to more effectively achieve our goal, our work has been focused on how to extract the robust and effective features, and how to obtain the model of tracking and recognition by fusing spatial-temporal context information and multi-feature, taking advantage of probabilistic graphical model (PGM). Therefore, we carry out our research on tracking and recognition of two representative objects, vehicles and pedestrians. The main purpose of this thesis is to implement region-level object tracking, solve the unidirectional transmission of error in traditional tracking model and the problem of occlusion, and improve the accuracy and robustness of vehicle recognition. Specifically, the main contributions of this thesis include the following aspects: 1. We have analyzed the model based on the fusion of spatial-temporal context information for region-level tracking. First, we construct a dynamic conditional random field, the nodes of which correspond to image pixels. Thus the region-level tracking problem can be transformed into a maximum a posteriori(MAP) problem of a dynamic conditional random field. Then we define potential functions based on the time consistency, motion coherence and local smoothness. Finally, we can use the probability distribution to describe the spatial-temporal context among pixels and we can use loopy belief propagation algorithm for the inference. Experimental results show that the algorithm can achieve the goal of region-level tracking and also improve the accuracy of ... |
修改评论