Californian startup AEye revisits LiDARs with AI
The iDAR combines what the company says is the world’s first agile MOEMS LiDAR, pre-fused with a low-light camera and embedded artificial intelligence. AEye’s founder and CEO Luis Dussan is keen on highlighting the fast and dynamic perception and path planning that the iDAR platform enables, for intelligent and agile data collection. Interviewed by eeNews Europe, Dussan wouldn’t reveal much about how the two sensors, camera and LiDAR are stacked together.
“The CMOS camera and the MOEMS LiDAR are mechanically aligned, they see the same thing. How we manage the light path is our own secret recipe and that’s why we have about 17 patents around the system”, the CEO said. “The camera overlays its 2D colour images onto our 3D LiDAR scans without any registration post-processing required to correlate the two data sets, it is done mechanically”.
The LiDAR part is based on a proprietary low-cost, solid-state beam-steering MOEMS operating at the 1550nm telecom waveband, which according to Dussan, is the only sensible choice for the mass adoption of LiDARs.
“This waveband is retina-safe, it is not only invisible but does not focus on the back of your retina, instead it is absorbed by the liquid between the cornea and the retina. It is the right choice if you are going to put millions of those devices on the streets, so that there is no chance to do any harm to unsuspecting pedestrians. In contrast, 900nm is the most dangerous wavelength to stare at”, the CEO noted.
Then embedded AI algorithms leverage computer vision on the 2D images to guide the LiDAR scanning patterns in real-time so it can keep in check particular areas of interest within the perceived environment. This prioritization on-the-fly and the capability to analyze co-located pixels (2D) and voxels (3D) within a same frame allows the system to target and identify objects within a scene 10 to 20 times faster than LiDAR-only products, claims AEye.
In essence, the 2D colour image overlay allows AEye to add computer vision intelligence to the 3D point clouds, providing the iDAR with sufficiently detailed information to interpret signage, emergency warning lights, brake versus reverse lights, and other scenarios that have historically been tricky for legacy LiDAR-based systems to navigate.
“What is important to understand is that we call it agile LiDAR because it is a random access LiDAR. By software we can drive any type of scanning pattern you want, the system then figures out what is the fastest way to do it. You can control 20 parameters completely by software so the scanning patterns can adapt to the environment as it changes. At high speed on the highway, that may be long range and situational awareness. In a city, that may be all around awareness. For us, it is just a software switch, this is a big departure from what other startups are doing” explained Dussan.
The CEO also envisages the iDAR platform to be used as the primary sensor, augmented by independent views from other sensor inputs such as Radar and ultrasonic sensors.
“Because we have introduced AI at the edge, the iDAR can intelligently decide what it needs to do. That makes huge savings in processing power, system cost and it operates with a reduced data bandwidth. In comparison, today’s LiDARs rely on a fixed path planning route, sending data to a central processing unit, that’s not fast enough to perform decision analytics.
We are solving the main problem of our customers, which is that LiDARs are either under-sampling or over-sampling. How do you get the right perception of a dynamic environment if you are having to scan the sky at the same resolution as a small brick or an obstacle on the road ahead? This is a very inefficient use of photons, lots of 1’s and 0’s that are useless. In the end, 80 to 97% of the data is thrown away as it cannot be processed more than a certain amount of time.
We are the first to fix that problem, the camera guides the LiDAR where to look, and the LiDAR can find objects that the camera can’t always see” continued the CEO.
“Today’s systems try to scan at the highest resolution at 10Hz. But we don’t want to miss anything in space and time between those frames. 10Hz is too slow. Imagine two converging cars driving at 70mph on the highway, at 10Hz, you’ll sample the next car once every 20 feet (every 100ms), this is too slow for a safe margin. We can selectively revisit any object twice within 30 microseconds, that’s 3000 times faster than current LiDARs.
Then we can perform velocity measurement through triangulation, by revisiting the same object three times, faster than what could be done using coherent detection to work out the Doppler shift on the returning light”.
What’s more, because it is so much faster than competing LiDAR solutions, the iDAR platform can be used to emulate incumbent LiDAR patterns, allowing designers to progressively improve on them through software. What’s more, the iDAR’s system architecture allows for remote updates of the firmware and software, which enables rapid prototyping without requiring hardware modifications.
Dussan founded the company back in 2013, using a simulation program he had written to raise seed money. Later, the company raised 20 million US dollars in series A funding from investors such as Intel, Airbus.
AEye first demonstrated iDAR last September, operating a 360 degree vehicle-mounted solid-state LiDAR system with a range up to 300m at high resolution. The startup then launched the iDAR Development Partner Program for OEM customers, Tier 1 partners, and universities interested in integrating iDAR into their vehicles and will be announcing its automotive product suite at CES 2018 in Las Vegas.
AEye – https://aeye.ai
Related articles:
Hi-resolution LiDAR gets compact and affordable
Velodyne makes LiDAR sensors affordable
3D LiDAR generates environmental model from time-of-flight measurement
Californian startup boosts LiDAR’s specs with proprietary InGaAs design
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
