Object Detection Based on Binocular Vision with Convolutional Neural Network

Abstract

Autonomous vehicles are widely accepted as one of the most potential technologies in alleviating traffic problems. In most existing autonomous vehicles for object detection and distance measurement, compared with radar or LIDAR which obviously increases the cost, camera combined with Convolutional Neural Network (CNN) has advantage in accuracy and low cost. However, most object detection methods applied on camera cannot perform distance measurement. In this paper, we simultaneously carry out real-time object detection and distance measurement (DDM) in one system by utilizing CNN on a binocular camera. Firstly, a binocular camera is used to acquire disparity maps. Secondly, a set of high-quality region proposals is generated by those disparity maps and the number of region proposals is reduced. Thirdly, CNN is utilized to classify those region proposals and get the bounding box of detected objects. Consequently, those reduced region proposals generated by disparity maps lead to improved computational efficiency. Finally, the object distance is measured by the disparity map and the bounding box. The experiment results show that the proposed method can achieve an accuracy of 87.2% on KITTI dataset and an accuracy of 68% in the real environment for object detection. The average relative error of the distance measurement is 0.85% within 10 meters in real environment. The operation time of the whole DDM system is less than 80 ms.