Deep Learning for Self-Driving Cars

MOMI 2017
Day: Tuesday, 28th of February

Emilien TLAPALE
PhD

Title:
Deep Learning for Self-Driving Cars

Contact:

Hitachi
emilien.tlapale@airbus.com

Abstract:

Human factors appear in the vast majority of traffic accidents involving human injuries or deaths. Self-driving cars have been proposed as a possible solution to reduce traffic accidents, as well as improving commuting comfort, relieving traffic congestion and increasing fuel efficiency. In addition to the ethical and cybersecurity issues, numerous technical challenges arise in the design of autonomous vehicles. Among them, getting an efficient understanding of the scene and responding to unexpected events are vital to correctly planning trajectories.

Traditional approaches focused on explicit modeling of the environment and hand-crafted rules, but the last few years have seen the development of numerous extremely effective methods based on deep learning. Indeed, with the large number of available training datasets and the efficient processing power available on embedded systems, it has now become possible to train a deep neural network solely based on the recordings of the various sensors mounted on a car and use it on a real autonomous vehicle.

In the second part of the presentation we discuss current work on the desired properties on a deep neural network. Making sure that the autonomous systems are generic enough to account for the variety of driving environment and resilient to sensor dysfunctions can be handled directly in the training stage. Keeping the neural networks implementation efficient is a permanent struggle with the discrepancy between standard and embedded computing power: We describe how smaller neural networks can be generated from fully trained ones while keeping their desired properties. Although several autonomous systems focus on standard feedforward convolutional neural networks, recurrent neural networks have shown superior performance in steering angle and speed predictions, in part due to their ability to process temporally dependent events.

Building an autonomous navigation system is and will remain a challenge for the years to come, in particular if we want to understand it. We believe it will not only give us safer roads and vehicles, but also help us uncover the fundamentals of human cognition.

Comments are closed.