End to End Learning based Self-Driving using JacintoNet

Abstract

Automated driving functions, like highway driving and parking assist, are getting increasing deployed in high-end cars with the trend moving towards the self-driving car. With the advent of deep learning, many traditional computer vision techniques have been replaced by deep convolutional neural networks (CNN). End to end learning is one of the paradigm for self-driving, in which user provides a input images from the front facing camera to the given neural network and the network outputs the car control signals such as throttle, steering and braking. The paper proposes an embedded friendly convolutional neural network, `Jacintonet', to demonstrate self-driving using end to end learning paradigm in a virtual simulation environment. Paper discusses key learning during the training methodology and presents the results on embedded platform. Texas Instruments (TI) TDA2x System on Chip (SoC) is used as embedded platform for running `Jacintonet', real-time to demonstrate self-driving car in the virtual simulator.