Abstract
Semantic event detection is one of the most important parts in video surveillance system, because people usually interest in some semantic events and desire to detect them. This thesis proposes a method to detect a specific event. After foreground object extraction, we get foreground mask for each video frame. With the foreground masks, we calculate the block-based temporal histograms for background color, foreground color, motion magnitude, and motion direction. Each temporal histogram is modeled by a parametric model, GMM, and the parameters of GMM are taken as features. With these GMM parameters, an HMM is trained to detect the specific event. Because of the properties of histogram representation, our proposed block-based temporal histogram is insensitive to noise and errors produced in foreground extraction. Furthermore, this block-based temporal histogram retains spatial information, but general histogram representation has the drawback of losing spatial information. Besides spatial information, the block-based temporal histogram incorporates the temporal information of color and motion as well. The effectiveness of temporal information will be proved in experimental results.