Parameter anglysis for steering angle prediction using neural networks

Abstract

The paper presents analysis of steering angle prediction using neural networks on a curved country road. Data from real car driving on a country road is used both for training and testing of a neural network. Among parameters current steering angle, current speed of the car, and curve in front of the car are analyzed. The study aims to add to development of driver assistance systems for country road driving. Steering rules for flat and sharp curves are compared. Results show that predictions of sharp curve steering on a test set is much more accurate if learning was performed on a sharp curve, and predictions for the flat curve are not so sensitive to the learning set, but is to some extent better if learning is made using flat curve data. Further, steering prediction parameters for flat and sharp curves are compared. For flat curves “looking ahead” of 1s works best, while for sharp curves “looking ahead” of 0.5 provides the optimal prediction. Yet, the differences in error in the flat curve case were not large, and the optimum was shallow, while for the sharp curve it was much sharper.