A Vision-Based Method for Improving the Safety of Self-Driving

Abstract

As the accuracy in sensors and powerful in controller keep improving, there is more room for developing the perception of the road environment and the operation in complex traffic conditions of Connected Automated Vehicles. In this paper, we propose a control strategy with environment identification to minimize the cost but achieve the effect of expensive Multiline Lidar. We use computer vision and deep learning to train existing data sets in this paper. More specifically, we use efficient neural network trained the data in German Traffic Sign Recognition Benchmark and KITTI respectively to realize classification of traffic signs and detection of vehicles, and functions of OpenCV are used to identify and locate traffic identification lines. To plan and make decisions on the driving route, the vehicle driving simulator based on the Model Predictive Control also is used to collect, control and train the data. Finally, our method can be proved practically from the case study and data in Udacity's Self-Driving Car Nanodegree project and the road scene in real life.