LiDAR in Self-Driving Cars
What is LiDAR?
LiDAR stands for Light Detection And Ranging. It pulses laser lights and receives the reflection by objects. These reflections create a point cloud that represents these objects in 3D (as can be seen above) or in some configurations 2D. LiDAR is one of the more crucial and argumentative sensors that is used on self-driving cars. It’s used for surround view, detection of objects like pedestrians and other vehicles, and also to create detailed maps that self-driving cars need to get around (They have been used more in space and terrain mapping applications). The best LiDAR sensors can see details of a few centimeters at distances of more than 100 meters.
Physics of LiDAR:
LiDAR captures the surrounding environment by firing laser beams at the objects that surround it, in the same way that SoNAR (Sound Navigation And Ranging) uses sound or RaDAR (Radio Detection And Ranging) uses radio waves. They mostly use the wave in the 900nm IR length but some use longer wavelengths that work better in fog and rain. Time of Flight (ToF) allows the sensor to measure how far away objects are based on how long it takes the laser beams to bounce back. The distance is calculated by multiplying the speed of light by the bounce back time. LiDAR gives data points including x, y, and z coordinates, time and intensity which is based on the reflectivity of the object (the amount of reflected light or radiation produced by an object)
Types of LiDAR:
There are different ways to scan laser beams. One way is to scan using mechanical rotating mirror or using a vibrating micro-mirror. Instead of mechanically moving the laser beam, another way is using a phased array in which each laser beam is divided into multiple waveguides, the phase relationship between the waveguides can be altered and therefore the direction of the laser beam can be shifted. This type has the advantage of smaller form factor and no moving parts. Another possibility is to use the laser as a huge flash like with a camera and then measuring the arrival times for all the objects with one big imaging photodiode array similar to a 3D camera. In high-level LiDAR comes in four configurations:
1. Mechanical/ Spinning LiDAR:
They have IR light emitted from a laser, which is then parallelized and circularized into a round beam with optics. Each beam is matched with a receiver, typically an avalanche photodiode (APD). Multiple emitter-detector pairs are mounted on a column that is spun by a motor between 5 and 20 Hz. The duty cycles are low to ensure eye safety (LiDARs produce Class 1 or Category 1 lasers. There are four categories of lasers from the high-intensity class 4 to the safest and lowest intensity class 1. This class is harmless for the eyes). The vertical FoV (VFoV) is determined by the number of emitter-detector pairs stacked vertically, while the horizontal FoV (HFoV) is determined by the duty cycle and the motor rotation speed. They typically have the cleanest signal and noise ration to date and generally provide 360 HFoV. Their downsides are size and cost. Also there is a need for self-calibration as the motor bearings wear. A Mechanical LiDAR unit may have 8, 16, 32 or 64 lasers positioned at specific angles to produce an accurate 360 degrees, spherical view of the environment. Below is a typical internal architecture of a mechanical LiDAR.
2. Optical Phased Array/ Solid State (OPA/ SS):
They use an OPA instead of a MEMS mirror to scan FoV. OPA is a new technology compared to MEMS. Quanergy is one of the companies working on this technology. They are also known as Solid State. They’re much cheaper and smaller than typical LiDARs since they use a design that doesn’t use spinning mirrors to direct the laser beams out into the world, they steer the lasers electronically, they’re also more robust because they don’t have moving parts. However, Velodyne has a different view on SS LiDARs:
Solid state, fixed sensors are driven by the idea that you want an embeddable sensor with the smallest size at the lowest possible cost. Naturally, that also means that you have a smaller field of view. Velodyne supports both fixed and surround view sensors. The fixed sensors are miniaturized to be embedded. From a cost standpoint, both contain lenses, lasers and detectors. The lowest cost system is actually via surround view sensors because rotation reuses the lens, lasers and detectors across the field of view, versus using additional sensors each containing individual lenses, lasers and detectors. This reuse is both the most economical, as well as the most powerful, as it reduces the error associated with merging different points of view in real-time — something that really counts when the vehicle is moving at speed.
3. Microelectromechanical System (MEMS) mirror:
A MEMS mirror is used to scan the FoV with the laser beam, after it has been parallelized and circularized to a round shape. The detector side is typically an APD array. MEMS is proven technology in commercial use. It promises small form factor and relatively lower cost. Below is a typical internal architecture of a MEMS LiDAR.
4. Flash LiDAR:
A flash sensor sends out a burst of IR light through a laser to a fixed field of view. The reflected light from objects in the path is received usually in an array of p-i-n photodiodes. Using this signal the distance and angular location of the object is measured.
What are their strengths?
- Higher resolution, higher quality than RaDAR because of the more focused laser beam and larger number of laser beams in vertical scan and high density of laser points per layer. Therefore it can clearly distinguish between objects.
- They have 360 degrees of visibility and accurate depth information.
- They have a strong reliability like RaDAR but with much higher resolution and accuracy.
- They have been a successful source of ground truth, reliable sensor for vehicles that cost doesn’t matter.
- Longer range and wider field of view compared to other sensors
- Simultaneous tracking of objects.
- They track the retro-reflectivity of the objects it sees which can be used to ascertain the nature of targets i.e. road signs are highly reflective and fauna (trees and bushes) is not. Most license plates, street signs, even street line paint have retro-reflective surfaces, which provide a larger laser return signal.
- Perform well in conditions with low light or glare unlike cameras.
What are their weaknesses?
- Bulky and mechanically complex. They’re much bulkier than other sensors and therefore more difficult to integrate with the vehicle.
- Lack the performance at higher speeds or in challenging weather. They are affected by weather conditions and dirt on the sensor which requires keeping them clean.
- The cost. They’re expensive. Costing thousands or tens of thousands of dollars apiece.
- They can’t measure the velocity of objects directly and have to rely on different precisions between two or more scans.
- They struggle in bad weather. Note that LiDAR is able to send laser light through the gaps between raindrops and snowflakes, but it would collect less data since they can’t penetrate severe weather effectively.
- Currently demand is a problem and they’re backordered although only a small number of test vehicles are on the road. LiDAR manufacturers are struggling to keep up and force companies to wait six or more months for a new sensor.
Who are the players?
There are many companies involved in developing LiDAR sensor technology such as:
- Velodyne, both Ford and Baidu have invested heavily in this company
- Quanergy — focused on SSD LiDARs
- Waymo developed three different sensors
- IBEO
- Strobe bought by Cruise Automation
- Luminar
- LeddarTech
- Continental
- Valeo
- and more …
Below you can see the major players around the world.
It’s worth mentioning that Tesla doesn’t consider LiDAR sensor to be essential for self-driving cars. Their cars rely on Camera, RaDAR and SoNAR for self-driving tasks.
Future of LiDARs:
Future objective of research and development for this sensor can be summarized into:
- Decreased cost.
- Increased range.
References:
AEye, LeddarTech, OSRAM, Quanergy, Udacity SDCND, Velodyne, Voyage, Waymo
