Region of Interest Generation for Pedestrian Detection using Stereo Vision

Abstract

Pedestrian detection is an active research area in the field of computer vision. The sliding window paradigm is usually followed to extract all possible detector windows, however, it is very time consuming. Subsequently, stereo vision using a pair of camera is preferred to reduce the search space that includes the depth information. Disparity map generation using feature correspondence is an integral part and a prior task to depth estimation. In our work, we apply the ORB features to fasten the feature correspondence process. Once the ROI generation phase is over, the extracted detector window is represented by low level histogram of oriented gradient (HOG) features. Subsequently, Linear Support Vector Machine (SVM) is applied to classify them as either pedestrian or non-pedestrian. The experimental results reveal that ORB driven depth estimation is at least seven times faster than the SURF descriptor and ten times faster than the SIFT descriptor.