Design and Implementation of Visual Detection System for Driving Environment

Abstract

With the growth of car ownership, traffic safety problems become more and more serious. Traditional passive safety measures can only reduce losses, but fail to effectively avoid occurrence of traffic accidents in driving. Aiming at deficiencies in traditional target detection methods, this paper uses a deep learning based target detection method on an established driving environment dataset, so as to realize recognition and positioning of road moving targets under complicated road conditions. In addition, this paper analyzes and verifies the feasibility of the method for detection and classification of different targets in the driving environment, and compares the detection speed and accuracy of different target detection algorithms. Based on the analysis of experimental results, a detection method based on OHEM (Online Hard Example Mining) Algorithm and combined with Faster R-CNN is proposed. It has been verified in the experiment that, the improved algorithm of OHEM + Faster R-CNN proposed in this paper prevails over YOLOv3 in detection efficiency of small targets. Its recognition accuracy for larger targets remains over 90%, and mAP reaches up to 0.906.