Pedestrian detection in dynamic scenes based on intersection kernel SVM

This paper presents a new algorithm to detect pedestrians in dynamic scenes. The algorithm includes three steps. First, the road surface is detected in an illumination invariant feature space. Second, regions of interest(ROIs) are extracted through the pinhole imaging principle and the scale of ROIs is normalized. Finally, the Histogram of Oriented Gradients (HOG) of the ROIs is used as an input to a trained histogram intersection kernel support vector machine(IKSVM) classifier for the pedestrian recognition. Experimental comparisons have been carried out to demonstrate the feasibility of our approach.