Figure 4

The features used in supervised case are described in [19]; a single feature vector (right) is computed for each sequence, by concatenating data coming from each frame of the sequence (left). In each frame, Harris interest points are used to recover a tight bounding box, which is vertically partitioned in three regions. The topmost 1/5 of the bounding box approximately contains the head and the neck. The middle 2/5 contains the torso and hands, whereas the bottom 2/5 of the bounding box contains the legs. Such segmentation is obviously approximated, and the resulting features would still be usable in cases where the assumptions are not met. Flow data in each region is summarized in separate histograms for the horizontal and vertical directions.