Self-driving cars need ground penetrating radar, says WaveSense
The pitch: current systems exclusively rely on above ground sensors like GPS, lidar and cameras to identify the exact position of the vehicle, but those can be confused and handicapped by poor visibility issues and adverse weather conditions such as snow, heavy rain, fog, sand and dust. But more reliable and stable data can be tapped beneath the road’s surface where unique subsurface textures can be mapped and leveraged to accurately position a car on a track, regardless of the weather conditions and visibility above ground.
The technology, initially developed by researchers at the MIT Lincoln Laboratory for military applications (accurately detecting and mapping mines and other sub-surface explosives) is described by WaveSense’s CTO and co-founder Byron Stanley in a 2015 paper “Localizing Ground Penetrating RADAR: A Step Toward Robust Autonomous Ground Vehicle Localization” published in the Journal of Field robotics.
The localizing ground-penetrating radar (GPR) consists of a 152cm×61cm plate array of 12 antennas (just under 8cm thick) sequentially pulsing RF patterns across the 100MHz to 400 MHz range. Just about the width of the car and positioned under the vehicle, behind the front wheels, the GPR operates at a sweep rate (across all its channels) of 126Hz, detecting underground features at a 2 to 3m depth to generate baseline maps on a first pass, tied with GPS tags.
Subsurface features as they are detected by the GPR systems (measuring the pulse reflections from many scattering points below the surface) are tied to the natural inhomogeneity in subterranean geology, including every discrete object and soil features, which makes them unique and identifiable.
On a second pass, the on-board electronics uses particle swarm optimization techniques to best match newly scanned data with the existing subsurface map, searching across five degrees of freedom (including latitude, longitude, height, heading, and roll) for the vehicle pose that best fits the new sweep to prior data.
This optimization technique, Stanley reports in the paper, offers centimetre-level positioning accuracy even when the vehicle is only partially overlapping a previously scanned path (the paper reports an accuracy within 6cm for an overlap of only 60cm (across a 152cm wide scan).
“Since equipped cars would only needs partial overlap, the usable map area in the last public design is much larger than the width of the vehicle” is keen to clarify WaveSense’s CEO Tarik Bolat.
Of course, it will take some time and money to map out every possible road before such a system can be reliably used, and heavy data handling may be one of the fundamental limitations of this approach. But unlike LIDAR mapping, one pass may be enough to cover most of each lane width (since partial overlap is enough to position a vehicle). Because surrounding vehicles do not block the sensors’ view (the GPR always gets a direct view at the ground) and because underground geology is relatively stable, each pass equates a valid map, without blind spots. Stability also means that a subsurface map would need to be scanned and updated much less frequently than a surface map.
When asked if the company would target other use cases such as utility or construction companies for which such a system may be more readily applicable to smaller fleets before aiming at the autonomous vehicle market, Bolat emphasized WaveSense’s primary focus on creating maps for autonomous vehicles. Yet the CEO admitted in an email exchange, “WaveSense maps also reveal vital information for utilities, municipalities, departments of transportation, and similar types of organizations who could leverage a comprehensive underground map to improve decision making around maintenance and budgeting”.
The company says it has several pilots underway with leading global players in the automotive and technology sectors and is currently closing a $3M seed round led by Rhapsody Venture Partners.
WaveSense – www.wavesense.io