Pedestrian Detection using HOGVA Feature and Local Pedestrian Classifier

Hiroshi Kasai and Kazunori Onoguchi


Pedestrian detection, HOG, Stereo vision, Local pedestrian classifiers


This paper proposes the HOGVA (Histogram of Oriented Gradient-Vector-Addition) feature for pedestrian detection. The HOG feature is often used for pedestrian detection. However, it does not exploit which side of the contour is a pedestrian because it represents the contour of the pedestrian by oriented gradient between 0 and 180 degrees. And it is weak to the image noise because the histogram is created by voting alone the edge strength of each point. The proposed HOGVA feature describes the pedestrian contour more in detail because the oriented gradient of the pedestrian contour is represented in the range from 0 to 360 degrees. Moreover, it is strong to the image noise because it describes the combination of gradient vectors at nearby two points by the addition of these vectors and the histogram of oriented gradient is created by voting the strength of the vector addition. At first, the proposed method detects candidate regions whose height are different from a road plain by stereo vision. Then, pedestrians in candidate regions are detected by using the cascade connection of the coarse HOG feature named 2HOG classifier and the HOGVA classifier. These features are generated by only edges whose disparities are similar to those of candidate regions. We also propose the local pedestrian classifier learning only background images collected in each local section as non-pedestrian samples. Experimental results for real road scenes show the effectiveness of the proposed method.

Important Links:

Go Back