A Biological Inspired and Coarse-to-Fine Pedestrian Detection Method

In this paper, we present a biological inspired coarse-to-fine approach to detect pedestrians in a given image. In the coarse detection step, a probable pedestrian area is predicted with both global and local features. Global orientation features are extracted to infer the spatial configuration of an image scene. Specifically, the vertical location of the pedestrian area is retrieved based on the spatial relationship between the target and other objects, and the local skin color and edge feature points are computed to refine the horizontal pedestrian location. Employing the predicted locations as top-down spatial information, we can generate the skin color saliency map based on the center-surround computation so as to find out target related sub areas. In the fine detection step, we input the extracted HOG-LBP features of candidate areas into an SVM classifier to detect pedestrian. Compared with traditional target detection methods, our biological inspired method consumes significantly less time but does not lose much detection accuracy.