Integration of Vehicle Detection and Distance Estimation using Stereo Vision for Real-Time AEB System

Abstract

We propose an integrated system for vehicle detection and distance estimation for real-time autonomous emergency braking (AEB) systems using stereo vision. The two main modules, object detection and distance estimation, share a disparity extraction algorithm in order to satisfy real-time processing requirements. The object detection module consists of an object candidate region generator and a classifier. The object candidate region generator uses stixels extracted from image disparity. A surface normal vector is computed for validation of the candidate regions, which reduces false alarms in the object detection results. In order to classify the proposed stixel regions into foreground and background regions, we use a convolutional neural network (CNN)-based classifier. The distance to an object is estimated from the relationship between the image disparity and camera parameters. After distance estimation, a height constraint is applied with respect to the distance using geometric information. The detection accuracy and distance error rate of the proposed method are evaluated using the KITTI datasets, and the results demonstrate promising performance.