Abstract
A new method is proposed for analyzing pedestrian behavior via image information through robust reference estimation and reliable subtraction. The method is based on subtraction between a reference image and each frame including objects, where an estimation based on the least median of squares (LMedS) is utilized for making the reference image. For extraction of pedestrian, the peripheral increment sign correlation (PISC) is used in order to descriminate stationary background pixels and pedestrians. The effectiveness of the proposed method is confirmed with experiments with real image sequences. It is found that the density and the occupation rate obtained from the real data are consistent with the ones calculated manually. Based on the data, we can classify regions in the field of view into some meaningful categories.
