GPU-Based Real-Time Pedestrian Detection and Tracking Using Equi-Height Mosaicking Image

Abstract

In this paper, we present a GPU-based real-time pedestrian detection and tracking system using a novel image representation called the equi-height mosaicking image [1]. This representation improves the processing time of the existing acceleration approach to pedestrian detection without decreasing accuracy. In equi-height mosaicking image generation, we first detect the horizon and crop a set of image strips from the road at uniform distance intervals. The height of each image strip is computed by projecting the predefined average height of a pedestrian at that distance onto the image plane. Then, all cropped images are resized to a uniform height and concatenated into a panorama image. Next, we detect the pedestrians on an equi-height mosaicking image using 1D based SVM classification. The SVM classifier is trained by an image dataset generated from various heights of pedestrians. After finishing this detection, we track the detected pedestrian in the previous frame. We performed the matching process in the neighbor block area of the equi-height mosaicking image to restrict the computation region. The detected or tracked results mapped onto the original image and grouped into multiple, overlapping regions.