Stereo Guided Pedestrian Detection

[Introduction | Demo | Publication ]

Introduction

Object detection and pedestrian recognition is essential to avoid dangerous traffic situations. In a driver assistance system for warning the driver of potential collision with nearby objects---especially pedestrians, we have to accomplish the following two tasks in real time. The first is to separate foreground objects from the background; the second is to distinguish pedestrians from other objects in order to protect pedestrians in danger. The first task is a segmentation procedure, the second one a recognition procedure.

In this project, I developed a real-time pedestrian detection system that uses a pair of moving cameras to detect both stationary and moving pedestrians in crowded environments. This is achieved through stereo-based segmentation and neural network-based recognition. Stereo-based segmentation allows us to extract objects from a changing background; neural network-based recognition allows us to identify pedestrians in various poses, shapes, sizes, clothing, occlusion status. The experiments on a large number of urban street scenes demonstrate the feasibility of the approach in terms of pedestrian detection rate and frame processing rate.

Demo

Here is a movie demonstrating people detection from a pair of moving cameras:

Publication

L. Zhao, C. Thorpe, "Stereo- and Neural Network-Based Pedestrian Detection", Proc. 1999 Int'l Conf. on Intelligent Transportation Systems, Tokyo, Japan, pp. 298-303, Oct. 5-7, 1999.