Video content representation as salient regions of activity

In this paper, we present a generic and robust representation of video shots content expressed in terms of salient regions of activity. The proposed approach is based on salient points of the image space, thus minimizing the computational effort. Salient points are extracted from each frame. Their trajectories are computed between successive frames and the global motion model is estimated. Moving salient points are selected from which salient regions are estimated using an adaptive Mean-Shift process, based on the statistical properties of the point neighborhoods. The salient regions are then matched along the stream, using the salient points trajectories. The information carried by the proposed salient regions of activity is evaluated and we show that such a representation of the content forms suitable input for video content interpretation algorithms.