Some thoughts on automotive-grade solid-state lidar

Currently, many OEMs have launched L3 autonomous driving cars based on LiDAR, such as the Alpha S, NIO ET7, and Zeekr 001. To meet automotive grade requirements, most manufacturers choose to use solid-state LiDAR as a perception sensor and cooperate with cameras for perception. This article will introduce the relevant properties of solid-state LiDAR, point cloud quality issues, and future development trends.

01. Solid-state LiDAR technology

Solid-state LiDAR is a type of LiDAR that has no moving parts. Optical Phased Array and Flash are its typical technology routes and are also considered to be pure solid-state LiDAR solutions. Currently, MEMS (micro-electromechanical system) LiDAR is the main one used.

Figure 1 Layout of solid-state laser radar

02. General selection parameter requirements for solid-state laser radar

Solid-state LiDAR can generally detect objects up to about 200 meters away. The basic parameter requirements for LiDAR can be divided into long-range radar and short-range radar requirements.

Requirements for long-range radar:

(1) Vertical FOV is 20-40°, horizontal FOV is around 120°

(2) The maximum distance for measuring distance is about 200m, and the best effect is achieved at 120m-150m

(3) Absolute ranging accuracy: The lower the better. The general radar on the market is within ±5cm.

(4) Relative ranging accuracy: 5cm-10cm@1σ

(5) Horizontal and vertical angle resolution: The higher the resolution, the better. Currently, the maximum resolution can generally reach 0.1°

(6) Can be powered directly from the vehicle power supply to meet automotive-grade temperature requirements

(7) Meet automotive-grade DV and PV test requirements

Requirements for short-range radar:

(1) The vertical FOV is preferably in the range of 70°-120°, and the horizontal FOV is around 140°

(2) The maximum ranging distance is about 50m, and the best effect is achieved at 10m-30m. (3) Absolute ranging accuracy: the lower the better, and the general radar on the market is within ±5cm. (4) Relative ranging accuracy: 2cm-5cm@1σ is the best. (5) Horizontal and vertical angle resolution: the higher the resolution, the better, and the current maximum resolution can generally reach 0.1°. (6) Can be directly powered by the vehicle power supply to meet the automotive grade temperature requirements.

(7) Meet automotive-grade DV and PV test requirements

03. Point cloud quality issues of solid-state lidar

In actual driving scenarios, highly reflective objects are very common, such as traffic signs. LiDAR is very sensitive to the high-intensity echoes reflected by them, which can easily form "ghosts" and "expansions" in the point cloud. When a real high-reflectivity object enters any area of ​​the LiDAR field of view, the output point cloud may have an image of a "ghost" in other directions in addition to the image at the real high-reflectivity position. In different scenarios, the whereabouts of different types of LiDAR "ghosts" may be different.

The "expansion" phenomenon usually manifests itself as the point cloud outline of a normal high-reflection sign spreading outward, imaging a point cloud shape that is larger than the real object, and the reflection intensity of the extra point cloud part is lower. (Source: RoboSense)

Figure 2: The “ghosting” phenomenon of solid-state laser radar

Figure 3: The “expansion” phenomenon of solid-state LiDAR

The ability to detect near-field obstacles is also very useful in intelligent driving. For example, in the traffic jam following (TJP) function, the smaller the laser radar's closest detection distance value, the shorter the following distance can be, and the less likely it is to be squeezed in.

However, medium- and long-range LiDAR detection may experience "point absorption" (inaccurate distance measurement) and "hole" (no detection) phenomena when detecting close objects. Small areas where the above two problems occur are usually set as "blind areas", and the output point cloud data is not recommended for use.

"Voids" describe the phenomenon that the LiDAR sometimes loses the detection of low near-field obstacles when moving from far to near. The "occasional absence" of the original point cloud of obstacles makes it difficult for the perception algorithm to track continuously, which can easily lead to sudden braking or frequent "deceleration and acceleration" of intelligent driving.

Figure 4: Solid-state laser radar "sucking point" phenomenon

Figure 5: “Void” phenomenon of solid-state laser radar

04. Development trend of solid-state lidar

Most of the lidars currently in use are of one-dimensional motor scanning architecture. This type of one-dimensional motor scanning architecture has been used for more than ten years. The laser transceiver unit is completely fixed before leaving the factory, resulting in the scanning beam distribution and maximum frame rate being fixed at the factory.

Subsequent solid-state LiDARs will have two-dimensional MEMS scanning functions, which can arbitrarily change the horizontal and vertical scanning speeds to change the scanning form, and the switch can be completed in the next frame after receiving the command. The line number distribution can be changed arbitrarily, and the LiDAR can freely adjust the angle range and resolution size of the ROI area based on different driving scenarios.

Figure 6 2D MEMS spatial scanning array

Figure 7 2D MEMS scanning array

In order for the intelligent driving system to realize the HWP (Highway Pilot) function on the highway, it is necessary to obtain a longer effective detection distance for the above obstacles, which requires the LiDAR to have a high ranging capability and a high effective resolution (i.e., a high resolution in the ROI area where the obstacle is located). The vertical resolution of the ROI area is intelligently improved, the imaging density of the obstacle point cloud is doubled, and the high accuracy of small objects in front is measured to make decisions in advance. On straight roads, improving the vertical resolution means that the perception algorithm can recognize the vehicle at a longer distance (up to 200m), and the detection distance for static small obstacles can also be farther (up to 160m). For intelligent driving vehicles traveling at a speed of 120km/h, this is a qualitative change from a safe braking distance to a comfortable braking/lane change distance, which effectively improves the safety and driving comfort of the vehicle.

Figure 8 Schematic diagram of high-resolution difference in ROI area

05. Conclusion

With the development of autonomous driving technology and the update of sensor hardware technology, solid-state lidar will be installed in large quantities in mass-produced vehicles in the future to meet the requirements of automotive-grade autonomous driving.