Road topology classification for autonomous driving

Zusammenfassung

Road topology classification is a crucial point if we want to develop complete and safe autonomous driving systems. It is logical to think that a thorough understanding of the environment surrounding the ego-vehicle, as it happens when a human being is a decision-maker at the wheel, is an indispensable condition if we want to advance in the achievement of level 4 or 5 autonomous vehicles. If the driver, either an autonomous system or a human being, does not have access to the information of the environment, the decrease in safety is critical, and the accident is almost instantaneous, i.e., when a driver falls asleep at the wheel. Throughout this doctoral thesis, we present two deep learning systems that will help an autonomous driving system understand the environment in which it is at that instant. The first one, 3D-Deep and its optimization 3D-Deepest, is a new network architecture for semantic road segmentation in which data sources of different types are integrated. Road segmentation is vital in an autonomous vehicle since it is the medium on which it should drive in 99.9% of the cases. The second is an urban intersection classification system using different approaches comprised of metric-learning, temporal integration, and synthetic image generation. Safety is a crucial point in any autonomous system, and if it is a driving system, even more so. Intersections are one of the places within cities where safety is critical. Cars follow secant trajectories and therefore can collide; most of them are used by pedestrians to cross the road regardless of whether there are crosswalks or not, which alarmingly increases the risks of being hit and collision. The implementation of the combination of both systems substantially improves the understanding of the environment and can be considered to increase safety, paving the way in the research towards a fully autonomous vehicle.