Occlusion Sequence Mining for Activity Discovery from Surveillance Videos

AbstrAct Complex multiobject interactions result in occlusion sequences, which are a visual signature for the event. In this work, multiobject interactions are tracked using a set of qualitative occlusion primitives derived on the basis of the persistence hypothesis: objects continue to exist even when hidden from view. Variable length temporal sequences of occlusion primitives are shown to be well correlated with many classes of semantically significant events. In surveillance applications, determining occlusion primi-tives is based on foreground blob tracking and requires no prior knowledge of the domain or camera calibration. New foreground blobs are identified as putative objects that may undergo occlusions, split into multiple objects, merge back again, and so forth. Significant activities are identified through temporal sequence mining; these bear high correlation with semantic categories (e.g., disembarking from a vehicle involves a series of splits). Thus, semantically significant event categories can be recognized without assuming camera calibration or any environment/object/action model priors.