PEDESTRIAN DETECTION USING HOGVA FEATURE AND LOCAL PEDESTRIAN CLASSIFIER

摘要

This paper proposes the HOGVA (Histogram of OrientedrnGradient-Vector-Addition) feature for pedestrian detection.rnThe HOG feature is often used for pedestrian detection.rnHowever, it does not exploit which side of the contour isrna pedestrian because it represents the contour of the pedestrianrnby oriented gradient between 0 and 180 degrees. Andrnit is weak to the image noise because the histogram is createdrnby voting alone the edge strength of each point. Thernproposed HOGVA feature describes the pedestrian contourrnmore in detail because the oriented gradient of the pedestrianrncontour is represented in the range from 0 to 360 degrees.rnMoreover, it is strong to the image noise becausernit describes the combination of gradient vectors at nearbyrntwo points by the addition of these vectors and the histogramrnof oriented gradient is created by voting the strengthrnof the vector addition. At first, the proposed method detectsrncandidate regions whose height are different from arnroad plain by stereo vision. Then, pedestrians in candidaternregions are detected by using the cascade connectionrnof the coarse HOG feature named 2HOG classifier and thernHOGVA classifier. These features are generated by onlyrnedges whose disparities are similar to those of candidate regions.rnWe also propose the local pedestrian classifier learningrnonly background images collected in each local sectionrnas non-pedestrian samples. Experimental results for realrnroad scenes show the effectiveness of the proposed method.