Hot-spot detection by group interaction extraction from trajectories
Fan Chen · 2013
We present a method for detecting hot-spots from surveillance videos via the extraction of group interactions (defined as stable and continuous spatial proximity of multiple objects). With a method that we propose for multi-object tracking in the multi-view scenario, we collect the trajectories of objects, from which we detect the group interactions. We assume that the movement of each object is driven by its interest of interaction, and model a group interaction by the mutual interests between objects. We solve detection of group interactions as a tracking problem, which first extracts unit-interactions by grouping objects at each individual frame, and then temporally associates them into continuous group interactions. We perform experiments on a publicly available dataset, and show that our tracking method achieves an accuracy around 95% and our detected group interactions could recall 80% of manually annotated hot-spots.