Automotive Lidar: Lidar for self-driving cars comes in many competing forms
For autonomous vehicles, a choice remains between spinning scanning, solid-state scanning, time of flight, flash, and coherent lidar.
“The jury is still out on architecture” for lidars used in self-driving cars, says Scott Davis, manager of electro-optical systems development for automotive applications for Analog Devices (Norwood, MA), after a session on laser applications in mobility at the OSA Laser Congress (November 5–8, 2018; Boston, MA). The speakers agreed that lidar can collect data vital for automotive autonomy, but the technology remains in flux.
The near-term focus is on limited tests of robo-taxis. Waymo (Mountain View, CA) has most of its fleet in the Phoenix, AZ area, where hundreds of people have joined an “early rider” program, says Waymo lidar systems manager Simon Verghese. Tests have begun with paying customers, yet Waymo’s approach remains slow and steady. Its cars travel on city streets in areas that have been mapped in great detail (see figure), which allows the car’s computer control systems to draw on an extensive dataset based on the fixed environment to supplement real-time observations of other cars, pedestrians, traffic lights, and other changing features by its suite of onboard sensors.
Lidar’s role is to record a point cloud of precise 3D locations for vehicles, people, wildlife, and other objects. A typical goal is collecting a million points per second, which an artificial-intelligence (AI) system then analyzes together with data from onboard cameras and microwave radar to identify what’s in the local environment and how it might move in the next few seconds. Engineers are seeking the best way to do that.
Waymo’s fully self-driving Chrysler Pacifica hybrid is seen on a public road in Chandler, CA. (Courtesy of Waymo)
905 vs. 1550 nm
One big question is what wavelength to use. So far, the leading lidar sources have been 905 nm diode lasers, which are inexpensive, readily available, and easily detected by silicon photodiodes. However, power levels used in open areas are limited by eye-safety standards because 905 nm light penetrates to the retina, which in practice restricts lidar range to tens of meters. Top lidar pulse powers can be 10X higher for 1550 nm, which does not reach the retina, says Umar Piracha, a senior researcher for IMEC (Kissimmee, FL). The longer wavelength also produces more photons per unit power, and less sunlight reaches the ground at 1550 nm than at 905 nm. Piracha says that strong cases exist for both wavelengths, but Davis expects 1550 nm to win ultimately.
The key attraction of the longer wavelength is its longer range. Luminar (Orlando, FL and Palo Alto, CA) is targeting a range of 200 m for objects reflecting only 10% of incident light, says Matthew Weed, its director of technology strategy. That range is needed to identify risks far enough away that a car traveling at highway speed can stop itself before hitting anything or anyone. “A human needs five to seven seconds to adequately predict what [another] car is going to do,” Weed says. Luminar has focused on 1550 nm so that it can achieve high performance first, then reduce the high cost that has been an issue for the longer wavelength. To overcome the limited brightness of diode-laser beams, Luminar uses linear avalanche photodiodes in its receivers.
The most common automotive lidar design puts a spinning scanner on the car roof to cover a full 360° horizontal view around the vehicle, but only has a limited vertical range. Scanners that include up to 128 separate lasers and spin mechanical mirrors at rates to 30 Hz can sell for tens of thousands of dollars, an important practical concern. Placement on top of the car gives the lidar a full 360° view at a distance, but leaves blind spots near the vehicle, so many cars include supplementary lidars mounted on their sides and corners.
Solid-state scanners can avoid the complexities of spinning mechanical mirrors, but at the cost of more limited range. Analog Devices has developed liquid crystal waveguide scanners that scan beams across 25° horizontally and 5° vertically, and Davis says the range could be expanded to 40° horizontal and 20° vertical. Several such scanners could be mounted around a car to cover a full 360° range.
Most scanning lidars are time-of-flight devices, which fire short pulses and measure the round-trip time from the laser to the object and back to a receiver that spins along with the laser. This approach works best if only one pulse is in flight at a time, which limits the pulse repetition rate to around 1 MHz for a range around 150 m. Data collection can be increased by stacking up to 128 laser emitters and spinning them to cover many parallel stripes simultaneously. Multibeam systems can also perform separate tasks, with some beams looking ahead to identify objects and other aimed downward to record road markings.
Flash lidar is a variation on a time-of-flight system that illuminates the field of view with a single burst of light and focuses the return signals onto a 2D detector array. Each pixel measures the time of flight, or how long the pulse reflected from an area on the target takes to return to the detector.
Time-of-flight lidars can detect movement by comparing successive scans, but coherent lidars can directly measure velocity with high precision. In a frequency-modulated continuous-wave (FMCW) lidar developed for automotive use, a single CW beam is split into two parts and modulated with a sawtooth-like frequency pattern. One part is focused onto an object, with the other part serving as a reference signal. For velocity measurements, Piracha says, chirping one pulse up in frequency and the other down in frequency can measure an object’s velocity. Some developers believe FMCW lidar could offer important performance advantages.
It will take time to sort out those details. Automotive lidars also face other technical challenges in delivering high-quality data that complements the capabilities of other sensors and the processing system to allow cars to take over the driving.