Design And Implementation of a Computer Vision-based Autopilot In a Simulation Environment
AbstractAutonomous cars are complex automotive systems with multiple subsystems working together in harmony to produce the desired result. The autopilot then works as the brain that governs the behavior of the car when the self-driving mode is switched on. In this work 1 , the design of a computer vision-based self-driving car autopilot was conducted in a simulation environment using an End-to-End approach, where a single neural network model maps the images from the environment directly to some sort of action. The NVIDIA End-to-End model was chosen, because it takes the input picture, does end-to-end processing, and gives an output of a steering angle. This is done without any human intervention. Two models influenced by NVIDIA's model were developed. The first one is based on VGG architecture while the second is based on the ResNet architecture. Unlike their baseline model, each of our models is supposed to not only determine the steering angle but also control the acceleration and the break. The two models were compared using different matrices.