Pedestrian detection and counting tries to recognize, locate and count pedestrians in static images or image sequences taken by cameras of fixed perspectives, without or seldom with human's participation and interference. The usage of digital camera and computer in detecting and counting pedestrians emancipates people from boring and onerous surveillance tasks with human eyes and have 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. We analyze the primary challenges in this domain and sum up the related works in recent years as well as their defects. Since the rapidly development of motion detection, near-perfect foreground images is available. Toward image sequences from realistic scenes, approaches of pedestrian detection and counting based on foreground image segmentation improve the state-of-art. Along with these approaches, we proposes several novel methods so as to deal with the phenomena which are frequently happening in realistic scenes, for instance,inter-person occlusion. The main contents and contributions of this thesis include: 1. Adaptive human model is proposed for pedestrian detection and counting. We model integral pedestrian and body parts based on contour information, and use two grid masks to infer visibility of torso sides, and then construct a branch-structure pedestrian classifier. Using the foreground image segmentation method to formulate a Maximum a Posteriori estimation problem, we verify and optimize the pre-detection results provided by adaptive human model. Because of the part detectors and adaptiveness of pedestrian model,our approach is capable of tackling inter-person occlusion. Besides, good pre-detection results accelerate the speed of convergence of Markov Chain Monte Carlo algorithm during the optimization, thereby satisfying the real-time application. 2. We proposed a group context based pedestrian counting method. With foreground image, we construct correspondence matrixes between consecutive images in order to detect and track groups as well as their relatives. Group context is modeled by foreground masks of a given group and its relatives to integrate spatial and temporal information together. Further, we assemble a series of context masks and formulate a joint Maximum a Posteriori ...
修改评论