Video plays a key role in security surveillance. Especially in ITS (Intelligent Transportation Systems), traffic surveillance is very important to traffic emergency and traffic signal control. With the expansion of cities and increasing use of vehicles, traffic accidents occur more frequently, taking great lose to the society, and if we can get the information of traffic accidents in time, it will be much helpful to handling traffic accident and dispersion. Therefore it is necessary to understand traffic activities in traffic surveillance video. This dissertation studies models to understand the activities of video based on statistical learning theory and probabilistic graphical models individually. To do video understanding and activities analysis we need some low-level features of video. So in this dissertation, we first introduce the work to collect features of activities from video. Integrating multi-frame tracking information and the three-dimensional features of vehicles, we complete the vehicle classification and counting with occlusion. In traffic surveillance, the low-level features are mainly the trajectories of each vehicle, and first of all, we use statistical learning algorithms to model trajectories to do activities understanding. In our latter work, we use the hierarchical Dirichlet processes (HDP), one popular Bayesian model in probabilistic graphical models, to model the trajectories. The main work in the dissertation is as follows, 1. We develop a method by integrating multi-frame tracking information and tree-dimensional size of the vehicle to classify and count the vehicles into tree types, and complete occlusion detection. In this way, we can avoid the influence of noise in just one frame to classify and count the vehicles. 2. We model the trajectories using kernel density estimation function based on the nearest points. After using spectral clustering algorithm to cluster trajectories hierarchically, in the testing stage, using the kernel density function, we complete the work to understand the activities of vehicles and detect abnormal successfully. Based on the recognition and recording of activity pattern, we can do video retrieval easily. 3. We survey the hierarchical Dirichlet processes and their applications. HDP are non-parametric Bayesian models, and can cluster data with various numbers of clusters. We review the development of HDP, other models using HDP as a prior, and their applications. 4. We cluster topics of the tr...
修改评论