Studied The Visual Behavior Of The Hidden Markov Model-based Analysis And Anomaly Detection

Visual activity analysis and unusual event detection has recently attracted wide attention in computer vision. In this thesis, we address the problem of visual activity analysis and unusual event detection. Unusual events are characterized by a number of features such as rarity, unexpectedness and relevance, which limit the application of traditional model-based approaches. A Continuous Hidden Markov model (CHMM) framework was formed to solve the problem. To deal with spatial and temporal data, CHMM outperform other models. The stochastic time-sequence recognition framework of the Continuous Hidden Markov Model forms the basis of activity recognition and anomaly detection. The system detects suspicious human activity in a scene. The system is designed to detect and track people in the scene; and then recognize the "normal" activity in the scene; finally, it detects anomalous activity by finding sufficiently large deviations from the normal activity patterns. In the framework usual events are first learned from a large amount of training data, we apply K-means clustering in the initialization of CHMM. Anomalies can be detected by the use of likelihood scores for CHMM representing activity patterns. Experimental result shows that our method accurately finds typical motion patterns and unusual events.