Automatic Learning of Semantic Region Models for Event Recognition

The semantic structure of scene is important information used for interpretation of object behavior or event detection in video surveillance system. In this paper, we propose an automatic method for learning models of semantic region by analyzing the trajectories of moving objects in the scene. First, the trajectory is encoded to represent both the position of the object and its instantaneous velocity. Then, the hierarchical clustering algorithm is applied to cluster the trajectories according to different spatial and velocity distributions. In each cluster, trajectories are spatially close, have similar velocities of motion and represent one type of activity pattern. Based on the trajectory clusters, the statistical models of semantic region in the scene are generated by estimating the density and velocity distributions of each type of activity pattern. Finally, using the proposed semantic region models, anomalous activities are detected in two scenes. Experimental results demonstrate the effectiveness of the proposed method.