Depth Based Context Modeling and Classification in Video-surveillance

With a dedicated definition of ‘Context’ in image understanding systems, we present in this paper a novel context modelling and classification system. The main goal behind multimodal context modelling is to identify the context type from video-surveillance footage of multipurpose halls. First, the distribution of the different zones in a multipurpose hall is automatically captured using a dedicated depth based segmentation method. The discriminative description is illustrated by extracting five semantic features according to depth zones. These features are processed with the Transferable Belief Model to propose a classification. Results show the validity of the method for context recognition.