Top Speaker announced for AutoSens USA this June
Director of Engineering, Automated Driving
David Doria holds a Ph.D. in Electrical Engineering from Rensselaer Polytechnic Institute, with a focus on LiDAR data analysis. He previously led computer vision and HD mapping teams at HERE Technologies and is now Director of Engineering at Magna Electronics, where he drives innovation in ADAS, sensors, and vehicle architecture. He has contributed to major open-source projects and has published extensively on LiDAR and computer vision.
At AutoSens USA, David Doria steps in with a thought-provoking presentation: Bridging the Gap Between ADAS and AV: Surprising Lessons Learned Developing Mid-Level Sensor Fusion. Whether you’re designing for today’s vehicles or preparing for tomorrow’s autonomy, this session offers a roadmap for smarter, more flexible sensor fusion development.
Ahead of this, be sure to check out our exclusive interview with David below…
1) What makes mid-level sensor fusion a compelling approach compared to early or late fusion, and how does it impact system performance?
To move from current ADAS configurations (e.g., 1V5R) that use late fusion to true low-level fusion, it requires significantly different compute hardware, as well as the addition of many new sensors. In contrast, we get a significant performance improvement by moving to mid-level fusion, but with only minor increases in compute requirements, and with no changes required to the existing sensor suites.
2) What were some of the most unexpected technical or operational challenges your team encountered while developing sensor fusion solutions?
With a 1V5R sensor suite, the camera and radars have a significantly different field of view, with the cameras being only forward-looking. When naively training a neural network for this fusion task, we found this non-fully overlapping field of view led to some unexpected performance issues. As is often the case with machine learning solutions, we had to develop some clever ways to coax the network into understanding this concept to bring the performance back into the expected range.
3) How are advancements in AI and machine learning shaping the evolution of sensor fusion strategies at Magna?
It’s widely accepted that a learning-based approach is the best solution for this type of problem. While it’s certainly not trivial to maintain state-of-the-art performance, the good news is that these approaches simplify the system architecture by forcing a lot of the design choices. This makes the evolution of the entire ADAS over time much more predictable. Further, it streamlines the engineering workflow with a standard “train ➡ test ➡ refine” approach instead of needing to develop heuristics throughout the process.
4) With the industry balancing cost, performance, and reliability, what’s your perspective on the role of sensor redundancy in ADAS and automated driving?
The types of sensors we use in automotive have wildly different characteristics and failure modes. There’s no question that taking advantage of the positive properties of each can yield much safer systems overall. The trick, then, is to do that in a cost-effective way, and this is exactly what a mid-level fusion approach does for us.
5) What key insights do you hope attendees take away from your session at AutoSens USA?
The key point to takeaway is that “better-than-late” sensor fusion is quite possible in modern L2 systems. Though it’s only one of many components in an ADAS software stack, I see it as a step that sets the groundwork for bringing in even more machine learning into the greater system. But it doesn’t come for free – significant offline infrastructure and process changes are necessary but justified as we work toward the safety goals that the industry and society have come to expect.
Mid-level fusion offers a practical path to safer, smarter ADAS. Magna is pushing the limits of current systems—without pushing costs sky-high. Don’t miss David’s session at AutoSens USA.
Interested in in-cabin monitoring technology?
With a pass to AutoSens USA, you’ll also get full access to our co-located sister event, InCabin. See the Agenda for InCabin USA here >>
