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

초록

This paper proposes an integrated system of vehicle detection and distance estimation for real-time autonomous emergency braking system(AEBs) using stereo vision. The main two modules, object detection and distance estimation, share the disparity extraction algorithm to satisfy the real-time processing. The object detection module consists of the object candidate region generator and the classifier. The object candidate region generator uses the stixel from the disparity. The surface normal vector is computed for validation of the candidate regions, consequently reducing the false alarms in object detection. In order to classify the candidates into foreground and background, we use the classifier based on convolutional neural network(CNN) classifier is employed. Then, the distance to the object is estimated based on the relations of the disparity value and the camera parameters. After the distance estimation, the height constraint is applied with respect to the distance using geometric information. The detection accuracy and the distance error rate of the proposed method are evaluated over the KITTI datasets and the results demonstrate promising performance.