A pedestrian detection system combining motion detection, spatial grouping and classification

Motion is a strong cue for the pedestrian detection task. Several motion detection approaches exist which segment moving foreground regions quite reliably, nevertheless, correct estimation of a class label for the segmented objects still represents a challenge. Certain object classes such as pedestrian groups and vehicles are difficult to discriminate from each other based on the geometric properties of foreground segments only. While appearance-based detection approaches enable class-specific detectors, pose and view-point variations, small-sized objects adversely affect the detection performance. In this paper we combine a (i) motion-based detector - having the generic ability of detecting and outlining arbitrary moving objects-, and (ii) an appearance-based detector exhibiting class specificity. Scale-adaptive mean shift clustering is used to delineate regions of moving foreground. Within the delineated clusters the appearance-based pedestrian detector is used to estimate the label of the object class. For the pedestrian class, such as a group of pedestrians, a simple model-based verification is used to estimate the location of humans. The proposed use of spatial context shows to improve the performance of the overall pedestrian detection system in terms of significantly lower false alarm rates while maintaining about the same detection rate, as evaluated in the paper quantitatively. Real-time performance is achieved.