Road Detection as a part of the Reward System for Reinforcement Learning-based Autonomous Cars

Abstract

Human drivers are subject to numerous flaws. It is common that the driversget tired and lose control of the vehicle, and some even get drunk or high, whichresults in dangerous situations for themselves and others.Autonomous driving is a field of study which has gained notoriety lately, as itattempts to get more reliable than humans at driving; Though there is still muchresearch to be done in the field. We are far away from replacing human driverswith safe AI-equivalents.In this thesis, we aim to validate a road detection algorithm as a part of thereward system of the autonomous vehicle1. We introduce a road detection asa simple supervised model, which predicts if the car is off or on the road, andis specifically designed to support Reinforcement Learning Algorithms such asProximal Policy Optimization, Deep Q-Networks, and Deep Deterministic PolicyGradient.