Development of driving environment recognition system for autonomous vehicle using deep learning based image processing

Abstract

In this paper, we have dealt with the recognition technology of the surrounding environment for autonomous driving. By using the deep learning, the recognition accuracy from the image obtained by cameras was improved, and distance information can also be predicted. More specifically, the developed system recognizes drivable regions, road markings and lanes through the Fully Convolutional Neural Network. In addition, the Single Shot Multi-box Detector (SSD) is applied to detect pedestrian, vehicles, and suggests a way to improve recognition accuracy of distance information considering the driving environment conditions. In conclusion, it was shown that the recognition system developed through this study can be applied to real autonomous vehicles, and can estimate sufficient information for driving even by using only low-cost cameras, not using expensive RADAR and LiDAR