Self-Driving Cars Using CNN and Q-Learning

Abstract

DrivingMatter is an experiment carried out to understand the deeper side of an autonomous car. In 1900s, idea was to drive car on Moon from Earth. This was initial motivation which grew from there and now expanding to complex system of roads in the real world. A book-sized Raspberry Pi based autonomous car is built to carry out the experiment on hardware. Software side was accomplished by developing a Python based library for controlling and communicating with car over a network or locally within the car. For environment learning two methodologies are practiced; Supervised learning: Drove the car on an environment/road and collected 3, 000+ data-points. Based on this a CNN model was trained which achieved 73 % test 89 % train accuracy. Reinforcement learning: Car is trained for three different road signs; Stop, No left, and Traffic light using DQN with existing CNN model. These road signs are detected in the environment using OpenCV cascade classifiers.