Automatic Steering Angle and Direction Prediction for Autonomous Driving Using Deep Learning

In this paper, we first examine the literature of steering angle and direction analysis and prediction using deep learning under diverse hazardous conditions or adverse weather conditions. We present our insights learned and propose a new deep neural network to automate steering angle and direction prediction based on real-world environmental perceptions and dynamic driving representation. We systematically explore the proposed deep neural network and its neurons using DeepTest comparing Rambo and Chauffeur models. There are two sets of data that we have used to test our network and trained driving models. One is the dataset from the Udacity self-driving challenge and the other is the dataset collected when we are driving in and around Commerce, Texas with the goal of ensuring the robustness of the proposed deep neural network against hazardous and adverse driving conditions. We then experimentally evaluate our network and models compared to the state-of-the-art on two datasets. The evaluation provides clear evidence and meaningful scientific insights to address grand challenges for safe autonomous driving.