A novel method for multimodal context identification from video-surveillance footage of multipurpose halls is presented is this paper. After presenting a dedicated definition of ‘Context’ in computer vision systems, the goal is resumed by detecting the active context type among predefined ones. To this end, a spatial modeling of context is performed by extracting five discriminative semantic features according to depth zones. These zones are detected by depth-based scene segmentation method. These features are processed with the Transferable Belief Model (TBM) to propose a classification. Results show the validity of the method for context recognition.