Feature-based ROI generation for stereo-based pedestrian detection

Abstract

Region of interest (ROI) generation is an important step in stereo-based pedestrian detection systems. In this paper, we propose an ROI generation method by fusing the color and depth information obtained from a stereo camera mounted on a vehicle. In our proposed method, a feature-based method which uses contour properties of the image is used to find the ROIs. In our feature-based ROI extraction method, we extract four features which are contour density, maximum area, maximum perimeter and matching score. Then we create a feature vector from these features and classify them using SVM. ROIs are then classified into the pedestrian and non-pedestrian classes using Histogram of Oriented Gradients (HOG)/Linear SVM. We have tested our proposed method on the Daimler dataset and experimental results show that our proposed method has a 96.5% accuracy for 1 false positive per frame and outperforms existing monocular and stereo-based methods.