Vehicle’s steering signal predictions using neural networks

Back-propagation trained neural networks, as well as extreme learning machine (ELM) were used to predict car driver’s steering behavior, based on road curvature, velocity and acceleration of a car. Predictions were performed using real-road data, obtained on a test car in a country-road scenario. We made a simplification using gyroscopically measured curvature of the road instead of visually extracted curvature measures. It was found that an optimum exists how far one has to look onto a curvature signal, according to neural network prediction accuracy. Velocity and acceleration did not improve steering signal prediction accuracy in our framework. Traditional neural networks and ELM performed similarly in terms of prediction errors.