Extraction of Semantic Information from Events

Abstract. Behavior understanding from events can be considered as a typical classification problem using predefined classes. In our research presented in this thesis, the main idea is to deal with this classification problem using motion analysis for behavior recognition. We describe our research work from its first steps to the final accomplishments. Novel methods are introduced including a methodology whereby frame infor-mation is coded in graphs (called Optical Flow Proximity Graphs- OF-PGs), using only optical flow to form the feature vector. A symbolic method including two levels of graph representation is also proposed dealing with the same recognition problem. OFPGs are also proposed for video indexing. Furthermore, a bottom-up approach for anomaly de-tection using a multi camera system is proposed. In the framework of this system, we present the use of one class continuous Hidden Markov Models (cHMMs) for the task of human behavior recognition. An approximation algorithm, called Observation Log Probability Approximation (OLPA), is proposed to overcome numerical stability problems in the calculation of probability of emission for very long observations