Blob Motion Statistics for Pedestrian Detection

Abstract

Video analysis aiming at efficient pedestrian detection is an important research area in computer vision and robotics. Although this is a well studied topic, successful detection still remains a challenge in outdoor, low resolution images. We present efficient detection metrics which consider the fact that human movement presents some characteristic patterns. Unlike many methods which perform an intra-blob analysis based on motion masks, we approach the problem without necessarily considering the pixel distribution inside the blob. Therefore, we apply periodicity analysis not to the pixel luminances inside the blob, but by analyzing the motion statistics of the tracked blob as a whole. We propose the use of three cues: (i) a cyclic behavior in the blob trajectory, (ii) an in-phase relationship between the change in blob size and position, and (iii) a correlation between blob size and vertical position, assuming that the camera is set up sufficiently high. These features are combined according to the Bayes classifier for improved performance. Experiments present numerical error rates and comparisons with other methods, illustrating the applicability of the proposed method.