Vision based steering angle estimation for autonomous vehicles

Abstract

Estimating the steering angle is a fundamental but challenging problem for autonomous vehicles. The main challenges come from different road conditions and capturing sensors. In this paper, we first present an autonomous vehicle prototype for different tasks in small and narrow indoor environment. We then perform a comparative study on two approaches for vision based steering angle estimation to drive the designed vehicle. One approach bases on conventional image processing techniques such as edge detection and Hough transform while the other one bases on advanced deep learning (convolutional neural network). We evaluate both methods on a common dataset. Experimental results show that in the scenario with simple and static background, image processing techniques give lightly faster and more precise steering angle. However, deep learning based approach is well generalized to background changes. We applied and integrated our proposed methods on the designed vehicle and show their capacity to drive the vehicle accurately.