Pedestrian Detection Based on Depth Information

Abstract

Pedestrian detection is a technique that uses computer vision techniques to determine if there are pedestrians in an image or video sequence and introduces precise positioning. In this paper, a pedestrian detection algorithm is designed for depth information images collected by a low-resolution Time-of-Flight (ToF) camera. First, GoDec algorithm is applied to remove noise and extract the foreground. Then Maximally Stable Extremal Regions (MSER) is employed to roughly segment the head region. Due to adhesion between head and shoulder, some shoulder regions may be falsely segmented by MSER. To overcome this problem, an improved water filling algorithm is proposed to get a fine detection result. Two datasets are constructed in an indoor environment to validate the validity of the proposed method. In the situation of crowding and people with diverse postures, the proposed method gets better detection performance compared with existing methods.