Real-time event recognition from video via a “ bag-of-activities ”

In this paper we present a new method for highlevel event recognition, demonstrated in realtime on video. Human behaviours have underlying activities that can be used as salient features. We do not assume that the exact temporal ordering of such features is necessary, so can represent behaviours using an unordered “bagof-activities”. A weak temporal ordering is imposed during inference, so fewer training exemplars are necessary compared to other methods. Our three-tier architecture comprises low-level tracking, event analysis and high-level recognition. High-level inference is performed using a new extension of the Rao-Blackwellised Particle Filter. We validate our approach using the PETS 2006 video surveillance dataset and our own sequences. Further, we simulate temporal disruption and increased levels of sensor noise.