Lidar "gets on the bus" race, and variables emerge

At present, lidar has been recognized by the industry as a "leap" in the logic of high-level intelligent driving and automatic driving. Last week, Xiaopeng Motors' world's first mass-produced lidar smart car, the Xiaopeng P5, rolled off the production line and will open the first batch of car owners' delivery at the end of this month.

Unlike Audi a few years ago, which carries Valeo lidar,

Why is it possible?

To improve reliability, we will start from first-tier cities (except Beijing) to carry out calibration testing, and gradually extend to other cities, this process may last a long time;

(Some jurisdictions may take longer), the feature cannot cover all cities when it is pushed, and will be gradually opened according to the approval of each city.

According to the plans of other car companies, the next version of the Great Wall Mocha equipped with lidar will be listed before the end of the year. Next year, NIO ET7, Volvo XC90, Zhiji L7 and other models will also be listed.

Obviously, it is still an unknown for lidar to get on the car and actually apply it to mass production systems. However, this does not affect the lidar "hardware" on the car, as the software can be updated by OTA.

One

The cost of lidar as one of the main sensor options for autonomous driving has long been one of the biggest obstacles to large-scale boarding. Although L4 level autonomous driving almost all carry different types of lidar, even Waymo, the number of fleets is only a thousand.

For the passenger car market for private users, cost sensitivity is even higher. Below 1000 yuan is considered to be the benchmark line for large-scale mass production import.

At present, most lidar suppliers are still in the "struggle" of cost reduction from $1,000 to $500. For the $300 or even $100 target, some vendors have opted for a "performance-degrading" strategy.

What is the reality?

When you see an image of a large number of point clouds generated by lidar, you can see a lot of detail. For example, images on pedestrians, cars, stop signs, and even vehicles, but this is not the "real" effect of point clouds, because most of the data is discarded by point cloud processing, which is similar to traditional millimeter-wave radar.

At the same time, real point cloud objects are also "blurry", and even planes like buildings or signs will have small bumps (not a real plane), which means that the returned points will bounce off subtle changes in the surface texture, or because the algorithms used to remove the vehicle's self-motion will produce small errors.

Even for stationary objects, their position changes slightly. Because the lidar returns only a series of points on the surface of the object, algorithms such as Kalman filtering must be applied to the return value to filter out jitter.

But for algorithmic processing, this is a double-edged sword.

Because if there is too much filtering, the perceptual system may "miss" the actual moving object; but if the filtering is not enough, the system may calculate that an object has just moved a perceptible distance in milliseconds, which may misjudge the subsequent path planning.

When companies claim that their lidar "sees" 200 meters or 250 meters, what they really mean is that the sensor is sensitive enough to detect a pulse returning from a certain part of an object at a certain distance, but this means almost nothing to the system.

This means that point cloud density is critical.

Because lidar works means that the density of points returned as a function of sensor distance decreases linearly, either horizontally or vertically. Low point cloud density can seriously affect object segmentation and classification, making it difficult, if not nearly impossible, to determine object types.

For example, DJI Livox's non-repetitive scanning mode, the unique flower-like scanning mode, as the fusion time increases, the field coverage of the point cloud will continue to increase until the field coverage is close to 100%. Under the traditional mechanical scanning, when the line is not dense enough, there is a possibility of losing the object.

Previously, Livox's Horizon series had a detection range of 260 meters and a horizontal field of view of 81.7 degrees, covering a range of 10 meters in 4 lanes. However, in medium-high-speed scenarios, this method has the problem of delay in accurately identifying obstacles.

In the customized version of Livox HAP provided to Xiaopeng Motors, Livox added two scans of the ROI area to enhance the safety perception of pedestrians and cyclists.

For long-range detection (over 300 meters), Livox has also introduced a new solution called Avina, which can switch between different scanning modes, ranges and different scenes. Among them, the repetitive scanning mode is used to cope with the application requirements of high precision and dense point clouds in specific areas.

This improvement, the purpose is very clear, is to increase the density of point clouds in some areas.

From a technical perspective, lidar's ranging, points per second (PPS), and angular resolution within a given field of view (and the corresponding field of view angle) are the three main specifications, of which the third parameter is used to determine differences in the ability to detect and classify objects.

In the view of Liang Hongyi, vice president of autonomous driving sales marketing at Innovusion, why the emphasis on lidar to see far is not to see a vehicle at a distance of 250 meters, but to see a small object at a distance of 100-150 meters, so that the system can have a safe warning time and distance.

This requires point cloud density, especially for low-reflective objects.

For example, 100 photons go out and only 10 photons return. Only when a vehicle with a reflectivity of 10% can be seen at 250 meters, it is possible to see low-reflex tires on the side of the road at 150 meters, a cardboard box of 20 × 20 cm, and it is possible to see small black (lower reflectivity) objects about 100 meters closer.

Behind this, thanks to the long-distance detection and high-definition resolution of Innovusion lidar, pedestrians at a distance of 120 meters can obtain more than 20 points or more than 20 points on vehicles about 400 meters, which is crucial for subsequent perceptual recognition algorithms.

In addition, Innovusion lidar also has a dynamic focus function, through local pixel encryption, key targets and small objects in the area of interest to "gaze" to obtain more accurate three-dimensional information.

For example, the system encounters a pedestrian crossing the road in the path of the vehicle. Because lidar is able to achieve dynamic scene adaptive changes in temporal and spatial sampling density in the area of interest, more attention can be focused on this pedestrian and less on irrelevant information, such as vehicles parked on the side of the road.

Two

At present, one of the cost reduction strategies of lidar is to combine lidar with different performance specifications. The problem, however, is that the system may require more lidar to provide full field of view coverage, or to complement it with other sensors. This means more data processing, sensor fusion, high data transmission, and high computing power requirements.

Taking Ibeo's scheme as an example, the performance parameters of ibeoNEXT are to achieve 260 meters of target detection at an 11.2 degree horizontal field of view, and a 32 degree field of view is still under development. This means that two blinding lidars are needed to form a wide-coverage environmental perception of the forward road.

For example, continental group, on the basis of its own self-developed mass production of short-range FLASH lidar, last year through the participation of lidar company AEye, plans to start production of long-range lidar from 2024.

Yijing Technology has launched a full set of MEMS lidar solutions for long range + short and medium range + blind area, including large field of view MEMS lidar for short range applications, and forward long-range MEMS lidar based on 1550nm fiber lasers.

It is undeniable that large-scale applications of lidar also need to significantly reduce costs, improve product life, and need to break through technical bottlenecks such as higher detection distances (more than 250m, or even 300-400m) and ultra-high scanning wiring harnesses.

Using a 1550nm laser can not only increase the output power by several orders of magnitude within the safety range of the human eye, but also more effectively avoid the sunlight noise area, thereby reducing the background light noise.

It can also be seen from the 1550nm+ MEMS scheme lidar ML-Xs launched by Yijing Technology that all parameters have reached a new height. For example, the field of view reaches 120° ×25°, the angular resolution reaches 0.15°, the wiring harness reaches 200 lines, and the background light noise (in natural light conditions) is reduced by 70%.

In addition, the 1550nm transmitter is safer than the 905nm, which can increase the power of the laser, improve the signal-to-noise ratio, reduce the pulse width, and have higher safety to the human eye, and more importantly, improve the effective distance of the lidar.

However, the biggest obstacle to the 1550nm solution is the cost. The vertical integration of the core supply chain system is also a key part of reducing costs and ensuring upstream supply in the future.

Luminar is the acquisition of InGaAs chip company OptoGration Inc., chip design company Black Forest Engineering, the main layout of 1550nm InGaAs photodetector chips and dedicated data processing chips, ideally, large-scale mass production costs can be reduced to several dollars.

Yijing Technology also chooses independent research and development and innovative design such as underlying chips and components, and its self-developed LiDAR proprietary chip and core algorithm have been formed, thereby further reducing the cost of 1550nm lidar.

The 1550nm solution launched by AEye emphasizes that it can be placed behind the windshield, similar to the traditional forward camera. This is critical for future model designs, without compromising aesthetics and reducing the possible constraints on the drag coefficient imposed by external placements.

However, in terms of cost and supply chain maturity, the 905nm still has specific advantages for now, although this wavelength brings concerns about eye safety (such as increasing power) and limits the detection range.

In addition, from R&D to manufacturing (product yield, indirect impact costs), supply capacity and after-sales support, lidar suppliers need to prove to the market that continuous and efficient mass production is ready.

Before that, there were many variables in this market.

Three

Among the many variable factors, there is another very key, that is, standards.

According to the data forecast of The Institute of Intelligent Vehicles of Gaogong, as the proportion of new domestic vehicles equipped with L2 levels continues to grow rapidly from 2022 to 2023, and the lidar equipped with high-end intelligent driving has entered the first round of growth cycle, it is expected that by 2023, the scale of front-mounted lidar in domestic passenger cars will exceed 1.5 million.

This means that one thing is already very clear about how the next technical route will develop and what choices the market will make: the front-loading mass production of lidar has begun. The entire industry also urgently needs standardization and standardization to provide a reference benchmark for large-scale front-loading mass production.

It has initially established the standard system composition of vehicle-mounted lidar and the plan and division of labor for the development of standards.

Among them, in terms of national standards, Hesai is the lead unit and Baidu is the joint lead unit, which is jointly responsible for the formulation of the national standard GB/T "Performance Requirements and Test Methods for Vehicle-mounted Lidar". In addition, Hesai has also participated in the development of a number of lidar industry standards.

Just in September this year, Hesai's Pandar128 lidar received the world's first LIDAR ISO 26262 ASIL B functional safety product certification issued by SGS. As a globally recognized standard for automotive functional safety, ISO 26262 is one of the main entry thresholds for the front-loading of core components in the field of intelligent driving.

From the minutes of the standard drafting meeting of the Auto Standard Committee, a preliminary "opinion" has been formed for the improvement of the standard, and more practical functions such as self-test, fault alarm, start time, and wake-up function have been added to the basic framework, and the point cloud performance requirements of high-speed object motion and high-dynamic scenes have been further distinguished in combination with the actual scene.

Further specifications and suggestions have been made for the test conditions of lidar, such as the light source of the laboratory, the reflectivity of the test plate; a detailed description of the specific test layout; a reference to the IEC standard provided by Rheinland for human eye safety; and a comprehensive requirement for the environmental test of the vehicle regulations.

At the same time, in the global market, IEEE also launched the standard development of lidar performance test methods last year, focusing on performance test methods, including distance accuracy/accuracy/resolution, maximum/minimum distance, detection probability, angular accuracy/resolution, reflectivity, etc.

In addition, the "soft" power of lidar is receiving more and more attention, and having system-level supply capabilities may become a new requirement for lidar suppliers in the next stage of the market.

In March, Luminar reached an agreement with Zenseact, a software subsidiary of Volvo Cars, to integrate lidar hardware and perception software to provide a complete set of autonomous driving solutions.

Because, for the main engine factory, after screening out the lidar hardware that meets the requirements for the model, it is also necessary to solve the problems of how to develop algorithms for the hardware and how to test and verify the perception ability of the system to ensure that the mass production requirements are met. In a lidar mass production project, software and hardware are equally important.

For example, the perception algorithm of lidar, including target detection, target tracking, target classification, speed judgment, judgment of drivable area and even path planning, etc., can realize the development of specific functions on this basis. A typical case is that the raw point cloud data output by lidar also exists in The Corner Case.

This means that the lidar perception algorithm, like the algorithm of the automatic driving system, also needs a large amount of scene data feeding, and the algorithm is iterated through a wealth of various scenario verification tests to ensure the safety and reliability of the lidar system, detection rate and accuracy.

Previously, based on several years of lidar software "development + test verification" closed-loop mass production experience, Liangdao Intelligent launched a lidar system solution for vehicle mass production, supporting customers to complete the performance definition of lidar before mass production, product hardware selection, perception algorithm development and test verification, and system integration into the mass production target of the vehicle.