A Machine Learning Based Forward Collision Avoidance System of the Kayoola EVS

Abstract

Autonomous vehicle control presents a signi cant challenge for arti cial intelligence and control theory. The act of driving is best modeled as a series of sequential decisions made with occasional feedback from the environment. Reinforcement learning is one method whereby the agent successively improves control policies through experience and feedback from the system. Reinforcement learning techniques have shown some promise in solving complex control problems. Forward collision avoidance system is an advanced driver assistance system that ma- neuvers for safe motion in case of the occurrence of an imminent forward collision. In this research an e cient reinforcement learning algorithm that actuates the car to move forward, keep lane and brake was designed for autonomous vehicles. Currently forward collision avoidance systems are based on input commands from the sensors like Lidar and Camera to the system and the output is based on the commands initialized. With this model the vehicle gathers data using an RGB Camera and collision sensor while mov- ing on the road in a simulated environment. Scenarios are developed which include car moving around corners, straight road and in a more urban layout with other obstacles like cars within the environment. Reward ags are given for no collision and penalty for collision with obstacles within the environment. Model testing was done in CARLA simulator and analysis of the model was done on a tensorboard and recorded simulation as the vehicles moves within the environment.