Pedestrian detection via contour fragments

Abstract

Pedestrian detection has been the key point of several hot research fields, including intelligent surveillance, driving assistance, intelligent robot, etc. Due to pedestrian appearance, multi-pose, angle of view, environment and other factors, the existing pedestrian detection algorithms still have some unsolved technical difficulties. In this paper, we focus on contour features. Based on traditional feature descriptors and feature selection methods, we proposed a novel oriented Chamfer distance (OCD) feature to describe pedestrian contours. It improves anti-interference ability of image noise using Bag of Words (BOW) model and image multi-scale structure. Besides, we conducted experimental comparisons using scale-invariant feature transform (SIFT), histogram of oriented gradient (HOG), speeded up robust feature (SURF) and OCD feature in feature extraction and parameter optimization, in Boosting and support vector machine (SVM) classification tasks. Experimental results demonstrate that the proposed OCD feature has great discriminant ability of pedestrian contours.