Visual attention and explainability in end-to-end autonomous driving

Autonomous driving for, the past few years, has achieved remarkable re- sults in terms of performance. Driven by researchers from around the world, it is to date a destination not as far away as it seemed not so many years ago. How- ever, the problem of autonomous driving can be approached and attacked from different angles. The goal of this thesis work is to shed light on a principle that to date, to the best of our knowledge, is still underestimated, and contributing to fueling feelings of distrust about these technologies, that is, Explainability. In this thesis, we propose an approach to autonomous driving based on imitation learning using Visual Attention mechanisms. Visual Attention allows, just as in a human approach to driving, regions of the image to be selected and weighted differently. The purpose is to prioritize useful regions rather than regions, pos- sibly noisy or otherwise irrelevant in a driving scenario. Interpretability and Explainability, to date, are for us fundamental properties that an autonomous system must necessarily possess. Throughout this thesis, we will analyze and propose resolution strategies for this problem to make these systems even safer and more reliable for themselves and all entities interacting in an urban driving environment. We will present increasingly detailed architectures and method- ologies that can make these systems transparent and fully understandable. We will present methods for analyzing failures that can help us, in some cases, to predict and anticipate them. In addition, for each proposed insight we will eval- uate its effectiveness by comparing its results with the state of the art. We will demonstrate with both quantitative and qualitative results that using visual at- tention allows for excellent performance as well as greater interpretability and explainability.