Deep Learning for Autonomous and Driver Assistant Systems

Abstract

The main goal of autonomous driving is the complete removal of human supervision from the work-flow of autonomous vehicles. This objective represents an opportunity for enhancing quality of life by reducing traffic, removing parking spaces in cities, increasing collective fuel efficiency, and reducing accidents. As autonomous driving is progressively getting integrated into our daily lives, viable solutions are required for its challenges. Artificial intelligence is the main technology that provides intelligent agents with the capability to perceive visual information in a way similar or even superior to human agents. In recent years the deep learning methods showed their outstanding power in dealing with various data processing tasks. Most of the open problems in autonomous driving are focused on the surrounding environment, and some are within the cabin. This dissertation presents solutions to selected problems in both domains using deep learning methods with various sensor modalities. We introduce a model that is able to extract the geometric relationship between two camera images. These results then allow us to proceed with the development of a model to solve geometric transformation in a sequence of point-cloud observations to address the odometry problem. Our proposed method is directly consuming the point-clouds in real-time. Further, we develop the first publicly available comprehensive Radar dataset and propose an open space segmentation model for this task. Lastly, we present a method that uses thermal imaging within the vehicle to count the number of passengers. The thermal images are hiding most of the visual features of passengers and better respect their privacy.