Learned probabilistic image motion models for event detection in videos

We present new probabilistic motion models of interest for the detection of relevant dynamic contents (or events) in videos. We separately handle the dominant image motion assumed to be due to the camera motion and the residual image motion related to scene motion. These two motion components are then represented by different probabilistic models which are further recombined for the event detection task. The motion models associated to pre-identified classes of meaningful events are learned from a training set of video samples. The event detection scheme proceeds in two steps which exploit different kinds of information and allow us to progressively select the video segments of interest using maximum likelihood (ML) criteria. The efficiency of the proposed approach is demonstrated on sports videos.