Applications of Deep Learning for Autonomous Driving
Senthil Yogamani
Technical Lead / Computer Vision Architect / Valeo Expert
Computer Vision Department (Automated Parking Segment)
VALEO Vision Systems
In this talk, I will provide an overview of applying deep learning for various autonomous driving problems and discuss some of the research we have done in this area mainly on computer vision. Compared to other successful applications of deep learning, autonomous driving has its own set of challenges to achieve high levels of accuracy and reliability needed for a safety system. The first success of Deep Learning was mainly in visual perception via CNNs which has enabled applications like semantic segmentation which wasn't deemed possible before and expanding into classical geometric vision problems like Optical Flow and Structure from Motion. The other application areas like motion planning, sensor fusion, etc are in early stages of research. There is also the ambitious side of solving autonomous driving by a single deep learning model (end-to-end learning) and its variant of modular end-to-end with auxiliary losses for semantics. From a deployment perspective, processing power is still a bottleneck and there is steady increase of computational power where next generation platforms are targeting 10-100 TOPS.
Seminar in Hamilton Seminar Room, 3rd Floor, Eolas Building, North Campus, Maynooth University - Tuesday 27th June 2pm-3pm.