SOM based activity learning for visual surveillance system

This paper proposes a new object activity learning algorithm based on self-organizing map (SOM) to detect anomaly events and predict activities in intelligent visual surveillance system. Two SOM networks are used to construct the distribution patterns of sub-trajectories and trajectories respectively. Sub-trajectories are first sampled to reveal the local activities. Before constructing the distribution patterns, trajectories are represented based on the distribution patterns of sub-trajectories. Finally, the distribution patterns of trajectories are merged to form clusters using agglomerative hierarchical clustering algorithm. By using the patterns of sub-trajectories, the learning process is accelerated and the representation of trajectory is simplified. The patterns of sub-trajectories and trajectories learned are then used to detect local and global anomaly events. A fuzzy set theory based predicting method is also proposed to predict the activity of object. Experimental results on real scene demonstrate the effectiveness of the proposed algorithm.