With the projects of " Security China", the intelligent surveillance system with its unique advantages, which can analyse the object' s behavior automatically, is becoming more and more necessary. Pedestrian is one of the most important objects in surveillance system. Detecting people in images is considered as one of the hardest tasks in object detection. The articulated structure and variable appearance of the human body, combined with illumination and pose variations, contribute to the complexity of the problem. Nevertheless pedestrian detection has been studied for decades, it has not been handled properly yet. This thesis targets the detection and segmentation of pedestrian in video surveillance, especially studies extracting suitable features for pedestrian detection with good ability of discriminative and low computational complexity, with employing the pedestrian detection and segmentation approach, to improve the performance of these approaches in intelligent surveillance system. The main contributions of this work are as follows: 1） The thesis proposes a cascade hierarchical rejection learning framework combining the SVM and the Adaboost discriminative model. By using this coarse to fine rejection structure, a fast and robust method for pedestrian detection has been evaluated by the experimental results. Combined with the motion information, the method can not only reduces the computational cost of sliding windows, but also reduces the false alarms in detection. 2）The thesis employs the Markov Random Field (MRF) to segment the moving foreground. In contrast to the segmentation using background subtraction, the method can improve the final result, which considers neighboring smooth information that will refine the final segmentation. Combined with the pedestrian detection approach, the silhouette of the pedestrian can be extracted properly. 3）The thesis proposes a method to handle the moving cast shadow in the background subtraction, based on a local texture descriptor called Scale Invariant Local Ternary Pattern (SILTP). The likelihood of cast shadows is derived using information in both color and texture. Finally, the posterior probability of cast shadow region is formulated by further incorporating prior contextual constrains using a Markov Random Field (MRF) model. The optimal solution is found via using graph cuts. An online learning scheme is introduced to shadow learning process in both texture and color space, which ...
修改评论