Steering angle prediction via neural networks

Abstract

Methods to calculate the correct steering angle are an important aspect of developing self-driving vehicles. Recently, steering angle prediction methods based on deep neural networks have achieved accurate results that outperform other methods over a range of road types. This paper investigates steering angle prediction for an autonomous vehicle as a regression problem and solves it using deep neural networks. The proposed approach obtains data in the form of the coordinates of lane lines and examines if they belong to one or more lines. Based on that, the method locates the path line that the vehicle is moving on. It then determines the equation that represents that path. Finally, the coefficients of this equation are fed into a trained neural network, which predicts the steering angle for that video frame. Extensive tests were performed on the Comma.ai and Udacity benchmarks to evaluate the performance of the approach. State-of-the-art results were achieved, with mean absolute error 0.64 and root mean square error 0.87 for the Comma.ai dataset, and mean absolute error 1.04 and root mean square error 2.33 for the Udacity dataset.