Abnormal Human Behavior Detection in Video Surveillance of Conference Halls

Abstract

We present a novel approach for the analysis of human behaviors in order to provide characterization of apparent events in the video surveillance scenes of conference halls. A topic model is used for activity analysis and for normal model learning. In particular, spatiotemporal features that carry trajectory, direction, and shape information of co-occurring individuals in the scene are treated in the Probabilistic Latent Semantic Analysis (pLSA) framework, in order to learn normal activity topics. Using the learned model, this approach is extended to detect abnormal behaviors. Two different strategies for anomaly detection are adapted and examined within the entire modeling system on a video dataset built as part of this work. The experimental study validates the ability of our system to model all normal activities and distinguish the abnormal ones.