Shape-based pedestrian detection and localization

Publisher Website
Google Scholar
Abstract
This work presents a vision-based system for detect- ing and localizing pedestrians in road environments by means of a statistical technique. Initially, attentive vision techniques relying on the search for specific characteristics of pedestrians such as vertical symmetry and strong presence of edges, allow to select interesting regions likely to contain pedestrians. These regions are then used to estimate the localization of pedestrians using a Kalman filter estimator. I. I NTRODUCTION The pedestrians detection is an essential functionality for intelligent vehicles, since avoiding crashes with pedestrians is a requisite for aiding the driver in urban environments. Vision-based pedestrian detection in outdoor scenes is a challenging task even in the case of a stationary camera. In fact, pedestrians usually wear different clothes with various colors that, sometimes, are barely distinguishable from the background (this is particularly true when processing grey- level images). Moreover, pedestrians can wear or carry items like hats, bags, umbrellas, and many others, which give a broad variability to their shape. When the vision system is installed on-board of a moving vehicle additional problems must be faced, since the observer's ego-motion entails additional motion in the background and changes in the illumination conditions. In addition, since Pedestrian Detection is more likely to be of use in a urban environment, also the presence of a complex background (including buildings, moving or parked cars, cycles, road signs, signals. . . ) must be taken into account. Widely used approaches for addressing vision-based Pedes- trian Detection are: the search of specific patterns or tex- tures (1), stereo vision (2)-(4), shape detection (5)-(7), motion detection (8)-(10), neural networks (11), (12). The great part of the research groups use a combination of two or more of these approaches (2), (13), (14). Anyway, only a few of these systems have already proved their efficacy in applications for intelligent vehicles. This work presents the first results of a new localization and association rule specifically designed to follow the detection process previously developed (15). In this work the strong vertical symmetry of the human shape is exploited to determine specific regions of interest which are likely to contain pedestrians. This method allows the identification of pedestrians in various poses, positions and clothing, and is not limited to moving people. In order to improve the reliability of the system and as preliminary work for pedestrain tracking, a pedestrian localization step has been added. Pedestrian localization iteratively computes the position of pedestrians in the 3D world. It has been conceived to be used for a tracking system. This paper is organized as follows. Section 2 introduces the structure of the algorithm. Section 3 describes the detection module, section 4 presents the localization procedure. Sec- tion 5 ends the paper with some final remarks.

This publication has 11 references indexed in Scilit: