| 英文摘要 | As an important research topic of visual analysis of human motion, people tracking is to detect, localize and track moving people in video sequences captured by cameras. It has a wide spectrum of promising applications, such as visual surveillance, motion analysis, human computer interfaces, virtual analysis and so on. For a long time, it has received increasing attention from both academia and industry. However, in this field many theoretical and technical problems remain open. Firstly, fast motion segmentation, non-rigid human motion, self-occlusion of a person and occlusion between different moving objects are still vexing problems in people tracking. Secondly, as the need of visual surveillance in wide scenes, the use of multiple cameras in people tracking is challenged by a series of issues. The main application of people tracking is visual surveillance. In this thesis, we study people tracking in such an application, provide a novel method of principal axis based people tracking. This proposed method is applied to single view tracking, occlusion handling, and multi-view tracking. Experimental results show the effectiveness of the proposed method. The main work of the thesis is given below: ① Based on the constraint that human bodies are symmetric around principal axes, a novel principal axis based method is proposed for single view people tracking. For each frame image, the procedures of motion detection and object classification are to extract moving regions of people. Then, principal axes of people under different situations are automatically detected from moving regions. Kalman filtering is further applied to track people. ② Based on the idea that silhouettes of human bodies can be represented by projection histograms, a simple classifier is designed on the vertical projection histogram. In this thesis, we only consider two kinds of objects: human and vehicle, which are the most ordinary objects in lots of monitored scenes. During classification, vertical projection histograms are firstly created by projecting foreground pixels onto the horizontal axis of the image coordinate system. Based on vertical projection histograms, compactness of objects is then defined to classify the two categories. ③ A new method is provided to handling occlusion in single view people tracking. In this method, by introducing the Bayesian network during which the occlusion relation transition is represented by a hidden state process, the occlusion problem is modified as a posterior probability estimation during probability propagation. In observation measurement, 2D ellipse models and color histograms are used as prior knowledge. Then, observation likelihoods are defined on the Bhattacharyya distance between the models and observations. In tracking, the particle filtering algorithm is further applied |
修改评论