Transfer Learning for Friction Estimation : Using Deep Reduced Features

Abstract: Autonomous cars are now becoming a reality, but there are still technical hurdles needed to be overcome for the technology to be safe and reliable. One of these issues is the cars’ ability to estimate braking distances. This function relies heavily on one parameter, friction. Friction is difficult to estimate for a car since the friction coefficient is dependent on both surfaces in contact - the tires and the road. This thesis presents anovel approach to the problem using a neural network classifier trained on features extracted from images of the road. One major advantage the presented method gives over the few but existing conventional methods is the ability to estimate friction on road segments ahead of the vehicle. This gives the vehicle time to slow down while the friction is still sufficient. The estimation pipeline performs significantly better than the baseline methods explored in the thesis and provides satisfying results which demonstrates its potential.