Pedestrian detection and counting tries to recognize, locate, track and count pedestrians in signal taken by vision sensor ( i.e. image sequence taken from camera ), without or seldom with human’s participation and interference. The usage of this automatic approach in detecting and counting pedestrians emancipates people form boring and onerous surveillance tasks with human eyes. Besides, this technique has various applications, such as driving assistance, human-computer interaction and public security. As a result, research on pedestrian detection and counting based on computer vision is realistically significant. This thesis mainly concentrates on pedestrian detection and counting using computer vision approaches. We analyze the primary challenges in this domain, discuss the related works in recent years as well as their defects, introduce some proposed methods, and present a pedestrian counting system. Based on pervious achievements, this thesis attempts to deal with different problems from distinct views, including solving the problem of counting in heavy occlusion, constructing pedestrian counting system, and proposing a pedestrian detection algorithm facing to overlap situation in video surveillance. Referring to each difficulty, an improved algorithm is proposed. The main contents and contribution of this thesis include: 1)An Algorithm dealing with pedestrian counting in extreme occlusion situation is presented. This approach makes use of motion difference between different people’s trajectories to achieve counting goal. In the algorithm, feature tracking method is utilized to generate motion trajectories, and my experiments show that these trajectories could sufficiently express the motion difference between distinct people. Besides, I attempts to improve the performance of the algorithm in the situation that people tend to move likely. We inspired by the though of bag-of-words and fuze motion information and texture information together under a unique framework. 2)We realize a pedestrian counting system with zenith camera of vertical perspective, which includes motion detection, template matching, object recognition, global optimization under Bayes framework, and tracking techniques. The system gets foreground of each frame from motion detection part and searches for human head in the area. Besides, we also compare different detection methods according to the special application need. A global optimization step is used to enhance detectio...
修改评论