AI for Intelligent Vehicles

Dr. Bram Bakker (Cygnify BV)
Title:
The disruption of driving: the various paths to high levels of vehicle automation, and the various roles of AI and the human driver

Abstract:
Artificial Intelligence (AI) is transforming many industries these days; but few industries are as affected as the automotive industry. In the race toward self-driving cars traditional automotive giants (both automakers themselves and their suppliers) struggle to stay relevant in the face of upstarts like Waymo and Tesla and smaller start-ups, with their Silicon Valley roots and strong AI background. AI, and its strengths and weaknesses, play a crucial role in this shake-up. On the one hand, modern AI like deep learning-based perception and prediction methods are instrumental in current vehicle automation efforts. On the other hand, some of those methods are known to be brittle (e.g. can easily be fooled by rare situations or easily designed adversarial examples). In part because of self-driving car accidents, some of which were likely caused by that brittleness, there is now increasing caution in most of the industry. There is also an increasing realization that human drivers will likely still be needed for a long time, at least for certain situations or environments--leading to new questions on how to divide the work between the human and AI. In this talk, these issues will be addressed in some detail, connecting them to various approaches to achieve higher SAE levels of vehicle automation as taken by different players in the industry.

Prof. dr. Dariu Gavrila (TU Delft)
Title:
Self-Driving Vehicles in the City: The Vulnerable Road User Challenge

Abstract:
Self-driving vehicles promise large benefits to society, as far as traffic safety, convenience and efficiency are concerned. The technological progress in this area has been remarkable over the past few years, fueled by better and cheaper sensors and processors, advances in machine learning and Big Data. Despite what it may appear from some company announcements or media reports, however, the problem is far from “solved”. This especially holds for complex city traffic and up-close contact with vulnerable road users (VRUs: pedestrians, cyclists, mopeds). VRU appearance varies widely, making reliable detection difficult, especially in adverse visibility conditions (occlusions, lighting). Moreover, VRUs are highly maneuverable and hard to predict. This complicates the development of a driving style, which is safe and comfortable yet also time-efficient.
This talk will discuss a processing pipeline from 3D environment reconstruction, VRU detection, VRU motion modeling and prediction, up to motion planning and control. I will present the results of a recent experimental study on VRU detection, which uses the latest deep learning methods and the large and diverse EuroCity Persons dataset (12 countries, 31 cities). I will subsequently cover methods for VRU path prediction, and discuss the importance of incorporating scene context. The talk spans the last decade of VRU-related research that I performed at Daimler R&D and TU Delft. I conclude with some thoughts on the road ahead.