A study on neural networks for vision-based road following of autonomous vehicle

This paper describes a neural network control for an autonomous vehicle to follow roadways with a vision camera. Without a complex geometric reasoning from visual image to vehicle-centered representation in conventional studies, the control system transfers the inputs of visual information into the output of steering angle directly. The neural networks replaces the human driving skill of nonlinear relation between vanishing lines of road boundary on the camera image and the steering angle of vehicle on the real ground. In straight and curved road, the driving performances by the proposed control scheme are measured in the configuration of disturbance and white noise correspond to vehicle joggling and sensing error.