As I look at the passenger cars today, I see two compelling needs. First is the need for increased safety. We tragically lose 20pedestrians per day and in total almost 43,000 Americans a year on our roadways.To help put that in perspective, the total losses are just about the same as the number of people we lose to breast cancer each year. Despite the adoption of automatic emergency braking, rear cameras, and other regulations driving safety technology, we still don’t see a substantial drop in these numbers. Second, the number of hours we spend having to drive a car – especially in traffic – is still too high. We would rather be thinking about important things, or less important things, or the weekend coming up, or last weekend. But we can’t because we have to drive our car and pay attention. “Hands-off / Eyes-on” still means pay attention, only it is arguably more difficult to do so since you aren’t driving (Google “startle effect in aviation”for more insight into how it can be worse).
At the top level both needs are addressed by having better sensors and better decisions about what actions to take. There is a lot of good work going on for both of these, and for today I’m just going to discuss the first solution – better sensors – and specifically radar sensors.
At Waymo we used 6 radars around our vehicle. Each of these was a really impressive long-range imaging sensor that my team designed and built. And although that approach works for a robotaxi, it doesn’t work for a passenger car because of the cost of the sensors. For society to make meaningful progress on making driving safer and less attention-consuming we need to make advanced sensing technology widely available – not just in the high-end luxury vehicles; I love the phrase we often use here, calling this need the “democratization of safety.” So here's the question: instead of using 6 very expensive radars around the vehicle, is there another architecture that can produce the right capabilities at a more attractive right price point? It turns out that there may be.
For all the good that radars bring (instantaneous velocity measurement, better performance in inclement weather, ability to “see” under cars via waveguide effects, etc.) it has the worst spatial resolution – or “pixel size” – of the camera / laser / radar automotive sensing trifecta. To address this, most imaging radars utilize physically large (think expensive) antenna arrays with high channel count(again, expensive). This high cost buys them several benefits, including long range capability and high-quality images. However, we don’t need really long range all around the vehicle for a passenger car, only in the front. So a better architecture for passenger cars would be a long-range imaging radar in the front and several much less expensive radars elsewhere around the vehicle, provided these radars still produce adequate imagery.
This exact problem was proposed to Spartan a little over a year ago by a major automotive Tier 1. They had customer (the OEM) and regulatory safety requirements needing more from their radar system while at the same time their procurement team was requiring them to keep their hardware costs affordable, which is difficult at any time but especially challenging in an inflationary economy. They asked Spartan to create a cost-effective, higher resolution radar solution that can see pedestrians, children, and bicyclists next to or near another vehicle, sign, or curb. They needed more accurate information for their ADAS systems to reduce accidents without an expensive “brute-force”hardware redesign.
It turns out that there are three methods one can use to increase the spatial resolution of a radar without changing its size, carrier frequency, or channel count. First, one could replace angle estimation with some sort of machine learning approach that is trained on the spatial channel data with various numbers of targets present, and produces as its output the angles, powers, number of targets present, etc. As with many AI applications, this can be very promising and powerful. And as withAI applications we will have the challenges of explainability and extensibility(do we need to retrain for every radar?). A second approach would be to use classic super-resolution techniques. IAA, MUSIC, ESPRIT – most good super-resolution approaches are an acronym! – are all common and well-known algorithms that perform well in MATLAB, but sometimes come with high computational cost and degraded performance in the complex and dynamic automotive environment. At SpartanRadar, we created a third approach and designed Clarify™, our own resolution enhancement software layer that focuses on robust performance and super-fast execution. But in theory any of these three methods can enable a much less expensive radar to be used around the vehicle to provide the required imaging capability.
We still have several challenges in achieving the safety and autonomy we all want in our passenger cars. It is easy to view radar as being an “old” technology with nothing new to offer, but it is actually (in my opinion) still the most underutilized sensor on the vehicle today. But by using advanced radar software and processing we can keep the hardware costs affordable while enabling the safety we need and autonomy we want.
We'll be at CES 2024!
Contact firstname.lastname@example.org to set up a meeting.