| Video surveillance is of great significance for intelligent urban management and cracking down crimes to build a safe and smart city. With the rapid development of digital cameras and storage technologies, the video surveillance systems spread rapidly and have become the major and core component of the monitoring system for social and public security. Nowadays, high-speed increasing cameras have covered every corner of cities and brought us vast amounts of video data. Under such circumstances, intelligent video surveillance is the most promising direction in contrast to the traditional technologies which are less efficient and labor-consuming.Discovering and tracking potential targets automatically from unstructured video data are basic problems in intelligent video surveillance, and have attracted much attention from both academia and industry. Because of the complexity of application environments, these problems are still far from solved. In this dissertation, we study the problems of discovering and tracking moving objects for the intelligent video surveillance, and propose several models and algorithms. Main achievements and contributions are listed as follows:1. Proposed a dynamic background modeling method based on SCBP feature. This method takes advantages from two aspects. On one hand, the proposed SCBP feature, a fusion of color and texture information, is discriminative, thus can distinguish foreground from background easily. On the other hand, we proposed an efficient method to refine the contour of object, which overcomes the disadvantage that region based methods usually can’t detect the real contour of object accurately. Experiment shows that proposed method is robust to dynamic background and achieves high detection rate along with low false alarm rate.2. Proposed an algorithm for foreground detection based on spatio-temporal continuity constraint. There always exist high correlations both in time and space in video data, which means that neighboring pixels tend to be of the same category, either foreground or background. Traditional background modeling methods take incomprehensive consideration on this problem, if not neglect. We proposed a generic spatio-temporal continuity constraint framework. Using Markov random field model, we take into account the correlations between adjacent pixels in time and space, as well as the similarity between features and background model. The final foreground is achieved by global optimization. Experiments show that proposed algorithm significantly improves the accuracy of foreground detection.3. Proposed a generic approach to real-time part-based visual tracking. Priori knowledge, such as a set of part detectors and a structure model, is essential and critical to traditional part-based methods. By using a novel matrix model, this problem is successfully solved. The matrix model divides target into several non-overlapping patches while ignoring its physical structures. Each patch is corresponds to a part of the target, and is associated with two attributes:a detector and a weight. To capture the variations of objects’appearance, these attributes are updated online. By way of the matrix model and weight strategy, our approach alleviate drifting problem significantly. Experiments show that the proposed approach can work generically in real-time and is robust to illumination changes, pose, occlusion and motion. It outperforms other methods.4. Based on our research achievements, an object oriented video surveillance prototype system is built. By taking real moving objects in videos as searching target, it provides convenient service for surveillance video retrieval and playback. Given a query image by the user, our system returns videos containing similar objects.At the end, we made a summary of this dissertation, and prospected the further studies in the future. |
PDF Full Text Request