Is the scanning module the core technological barrier of LiDAR?

Since the main architecture of LiDAR is transmission and scanning, let's trace back to the source and understand how the performance, reliability, and cost of LiDAR are affected by the transmission and scanning modules, in order to fundamentally clarify the primary and secondary aspects.


We further break down performance, reliability, and cost into small indicators as the vertical axis, and the influence of the transceiver module and scanning module on this indicator as the horizontal axis: where the horizontal axis is 1, indicating that this indicator is mainly determined by the transceiver module; If the horizontal axis is 5, it indicates that the indicator is mainly determined by the scanning module, as shown in the following figure.

First, let's look at performance indicators that users are concerned about, including ranging ability, accuracy, field of view, resolution, refresh frame rate, power consumption, volume, point cloud regularity, etc.

Distance measurement capability

The detection distance is the range that LiDAR can detect, reflecting the ability of laser transmission and signal processing. The higher the efficiency of laser transmission and reception, the stronger the signal processing ability, and the stronger the ranging ability.

In the actual use of LiDAR, the ranging ability is also related to the reflectivity of the object being measured. The higher the reflectivity, the more reflected light is received, and the farther the ranging is. So the detection distance generally appears in groups with reflectivity, for example, 150 meters @ 10%, which means the detection distance is 150 meters when the target reflectivity is 10%.

Of course, some manufacturers, in order to exaggerate product performance, advertise detection distances based on reflectivity of 50% or 80%, and even avoid discussing reflectivity. In response to these disturbances, some insiders roast when communicating with the author that those who talk about the detection distance without specifying the reflectivity are hooligans!

In Figure 8, the coordinate value of the ranging capability is 2, indicating that the ranging capability is mainly determined by the transceiver module, including the emission power of the laser, the emission band, and the detection sensitivity of the detector.


Why isn't it all determined by the transceiver module? Because sometimes the scanning module can also affect the ranging ability. For example, it is widely believed in the industry that if the aperture of the scanning mirror is too small, the ranging ability will be limited. The MEMS scheme has a smaller aperture of the galvanometer aperture, while the rotating mirror scheme has a relatively larger aperture.

accuracy

Accuracy refers to the accuracy of the detection distance, and the higher the detection accuracy, the more accurate the characterization of the target object.

The accuracy reflects the waveform processing capability and mainly depends on the transceiver module, which has a weak relationship with the scanning module. Therefore, the coordinate value of the ranging accuracy in Figure 8 is 1.

Field of view angle, resolution, and refresh frame rate

The field of view angle, resolution, and refresh frame rate that users are particularly concerned about are often interdependent and interdependent, so we will discuss them together.

Field of View, also known as FOV (Field of View), includes both horizontal and vertical directions, just like using a projector to illuminate a wall, the range covered by the projector's light energy is the same, the larger the range, the better.


Resolution refers to the angle interval between two adjacent detection points in a LiDAR, which is divided into horizontal angle resolution and vertical angle resolution. The smaller the angle interval between adjacent detection points, the stronger the ability to distinguish details of the target object.

The refresh frame rate refers to the scanning frequency of the LiDAR on the target object, usually expressed in frequency (Hz), indicating how many times it is scanned in 1 second.

In fact, these three parameters are interdependent and mutually influential, and the other two parameters can be adjusted by changing one of them. It would be unfair or even misleading to emphasize one of the other two parameters without specifying them.

At the 2019 CES exhibition, a manufacturer set the refresh frame rate to 1Hz to showcase the high resolution of their LiDAR equivalent to 640 lines. However, in reality, at a refresh frame rate of 1Hz, the LiDAR's tracking ability for dynamic targets lags behind and cannot be used.

Some manufacturers also emphasize the "highest resolution" that can be achieved within a small range, which is actually a clever move. What is the angular resolution in the sparsest part of the beam? They don't say.
Is there an industry recognized indicator that can comprehensively and comprehensively reflect the perception ability of LiDAR?
Yes, this is called "dot frequency".


Point frequency refers to the total number of detection points completed and obtained by a LiDAR per second, also known as points or points per second, similar to the concept of the total pixel count of a camera. Using point frequency to describe cannot be achieved by reducing the refresh frame rate, but can also avoid the one-sided description of "highest resolution" caused by uneven field of view angle size.

According to the definition, we can derive the formula for calculating point frequency:

The average number of points in the horizontal and vertical directions is equal to the field of view angle in the horizontal and vertical directions, divided by the corresponding angular resolution. Therefore, for two-dimensional scanning, the above formula can be expressed in a different way:

For one-dimensional scanning, the wire harness in the vertical direction is fixed, and the calculation formula is:

Taking Hesai's hybrid solid-state LiDAR AT128 as an example, its horizontal FOV is 120 °, horizontal angular resolution is 0.1 °, and it has 128 lines. Scanning once generates 120/0.1x128=153600 points, with a refresh frame rate of 10Hz. Scanning 10 times per second results in a point frequency of 153600x10=1536000 pts/s, which is consistent with the official announcement of over 1.53 million point frequency data.

It is obvious that, just like the pixels of a camera, the more point frequencies there are, the better the perception ability of the LiDAR towards the target object.

It can be said that just as the "hundred kilometer acceleration" of a car is the most core performance indicator to measure its power performance, point frequency is also the core performance indicator of LiDAR.

Some people may have doubts: can the refresh frame rate be increased by increasing the scanning frequency of the scanning components?

Increasing the scanning frequency can increase the refresh frame rate, but it does not change the point frequency.

Because if only the scanning frequency is changed, the refresh frame rate will increase, but the resolution will decrease, and the point frequency will remain unchanged.


For the sake of understanding this question, let's still use the example of a machine gun: dot frequency is the number of bullets fired by the machine gun per second, and the scanning frequency is the frequency of the machine gun swinging left and right. Can we change the frequency of the machine gun swinging left and right to change the number of bullets fired by the machine gun per second? Obviously not.

In other words, although dot frequency may seem like the calculation result of three numbers, it is a determining term (the independent variable in the function, determined by external input factors) rather than a result term (the dependent variable in the function can be determined by changing internal parameters). The field of view (FOV), resolution, and refresh frame rate, these three parameters, cannot determine or change the point frequency. On the contrary, the point frequency determines the constraint relationship between the three. If one parameter changes, the other two must also change accordingly.

Let's make another analogy: if the refresh frame rate is changed from 10Hz to 1Hz, the machine gun will swing slower left and right, resulting in denser natural scanning points and higher horizontal angle resolution; On the premise of keeping the refresh frame rate constant, if the FOV increases, the bullets will naturally become sparse, and the horizontal angle resolution will decrease. If the FOV decreases, the area of bullets scanned will decrease, the points scanned will become denser, and the horizontal angle resolution will also increase.

What determines the frequency? In Figure 8, the point frequency score is 1, which means that the point frequency is mainly determined by the transceiver module, as follows.

The upper limit of the point frequency for each laser is a comprehensive design choice that needs to consider performance requirements, total power consumption, lifespan, signal processing capabilities, etc.

The total number of lasers is determined by the architecture of the LiDAR transceiver module. So, can we achieve higher point frequency of LiDAR by increasing the total number of lasers, which means using multiple machine guns to simultaneously fire more bullets per second? The answer is yes, but every machine gun is not cheap and requires a large number of lasers. Without chipization, it is difficult to achieve controllable total costs, which is also one of the industry barriers.

The following figure shows the point frequency of each laser converted from the total number of lasers and the point frequency of several hybrid solid-state LiDARs obtained from official product manuals and industry discussions. The lower the point frequency of each laser, the lower the proportion of time each laser is excited (duty cycle), and the less negative impact it has on the lifespan of the laser.

When communicating with industry insiders, the author revealed that in order to achieve high lifespan, it is necessary to control the upper limit of each laser point frequency. However, for solutions with a small number of lasers, the problem of insufficient total point frequency needs to be solved. There are many manufacturers in the market that reduce transmitters, and the point frequency of each laser is already very close to or even exceeds a reasonable upper limit. There is no problem with doing demos, but there are challenges in the reliability of long-term work in the future.

power dissipation

The power consumption of the electronic module in LiDAR is much higher than that of the scanning mechanical module, because the electronic module needs to emit and receive light at a rate of millions of times per second, and each transmission and reception requires complex analog and digital circuit processing to convert it into a 3D point cloud signal. In contrast, the motion of the mechanical part does not output power to the outside world, but is generally a moving component with a uniform speed and ultra-low resistance, just like a stable gyroscope that does not require high power to maintain. Therefore, the coordinate value in Figure 8 is 2. Generally speaking, the power consumption of one-dimensional scanning is lower than that of two-dimensional scanning.

volume

The scanning module of a LiDAR is generally a small proportion of its overall volume, whether it is a mechanical radar's overall rotating disk or a one-dimensional or two-dimensional rotating mirror. In the transceiver module, the optical and electronic modules responsible for transmitting and receiving occupy the main space, and their volume is often difficult to further compress due to limitations in optical focal length, vertical field of view angle, and ranging capability requirements, or constraints caused by low integration due to the large number of channels.

The internal structure of a certain manufacturer's LiDAR, as shown in the figure below, has a scanning structure consisting of a layer below the orange circle at the bottom, and the transceiver module occupies the main volume of the LiDAR.
Point cloud regularity

The regularity of point clouds affects the adaptation difficulty of point cloud processing algorithms. In Figure 8, the score of this indicator is 5, indicating that it is mainly determined by the scanning components.

In terms of point cloud regularity, the scanning method that can form a horizontal and vertical matrix effect is the best for the algorithm, such as mechanical and one-dimensional rotating mirrors; Many LiDAR manufacturers believe that for two-dimensional mirror rotation schemes, on the one hand, multiple lasers need to be used to splice the field of view angles, which may cause deformation at the edges. On the other hand, scanning formed by two-dimensional motion is difficult to achieve an absolutely flat matrix effect, ultimately affecting the regularity of point clouds.

reliability

Reliability, also known as passing regulations in the industry, is determined jointly by the transceiver module and scanning module. After all, a certain component has a "weakness" that cannot pass the regulations, and the assembly is also very difficult to pass the regulations.

Generally speaking, the electronic components of the transceiver module are relatively easy to pass the vehicle regulations, especially the 905nm used by mainstream manufacturers. The upstream supply chain is relatively mature, while the supply chain of devices based on the 1550nm transceiver module is still relatively early, posing significant challenges in passing the vehicle regulations.

However, compared to the transceiver module, the scanning module has a greater impact. Some manufacturers believe that in addition to scanning methods such as mechanical, MEMS, and two-dimensional rotating mirrors that have been validated by Fareo and have lower risks, other scanning methods such as mechanical, MEMS, and two-dimensional rotating mirrors still need time to prove their reliability.

cost

The cost is determined jointly by the transceiver module and the scanning module, with a coordinate value of 3. Based on different scheme choices, both may have a significant impact on the cost.

Based on the above discussion, it is not difficult to draw the following conclusion:

The performance indicators of LiDAR mainly depend on the transceiver module;

Reliability is mainly determined by the scanning module;

The cost is determined jointly by both parties.

So, on the basis of selecting a reliable and stable scanning module, continuously optimizing the transceiver module will be a necessary path for an excellent LiDAR.

The essence of the scanning module is mechanical, and the reliability issues it brings need to be solved through engineering methods, mainly relying on time. In contrast, the barriers to transmission and reception modules are the evolution of electronic technology, consisting of technical barriers and engineering barriers, in which technological innovation and research and development capabilities are crucial.

Overall, the transceiver module is the "engine technology" in the LiDAR industry.

What efforts have LiDAR manufacturers made to address performance, reliability, and cost issues? Let's review the evolution path of the industry.

As the pioneer of the industry, Velodyne mechanical LiDAR has been selected by many Robotaxi projects due to its excellent performance parameters, which can achieve 360 ° horizontal field of view angle scanning, and excellent point cloud quality. The representative is HDL-64.

However, mechanical LiDAR also has its obvious problems.

The first issue is the expensive price. The HDL-64 was priced at $80000 in 2016, but it was not intentional selling by Velodyne. At that time, the transceiver modules were assembled from discrete components, and it is said that the material cost of each transceiver channel (a total of 64 channels) exceeded $100. In addition, the complex debugging and assembly process brought about high labor costs, ultimately resulting in its high price.

The second discrete component, which is manually assembled and adjusted, is difficult to ensure consistency.

In order to solve the above problems, the industry has begun to evolve from mechanical LiDAR to hybrid solid-state LiDAR, and in this process, two technological routes have emerged.

Route 1: Few channels, 2D scanning

This technology route reduces costs by reducing the number of lasers, and uses two-dimensional scanning (such as MEMS micro mirrors, two-dimensional rotating mirrors, prisms, etc.) in scanning components to achieve the effect of equivalent multi line scanning.

The obvious benefit of this solution is that the cost of the transceiver module is significantly reduced. For example, the first generation Innoviz lidar using MEMS micro mirrors is said to be priced at around $1000, which is 1/80 of the $80000 price of the Velodyne HDL-64E.

However, while compensating for the reduction in transmission and reception channels through innovative scanning structures, complex scanning structures also bring some shortcomings.

The first is reliability, which is widely reflected in the industry. Currently, manufacturers adopting this technology route have not clearly proven that they can pass the vehicle regulations, and are more in a "standard vehicle state".

The typical representative of this route is the MEMS scheme. Although MEMS is not a new technology and has applications in other sensors in vehicles, the size of MEMS mirrors used in other sensors is very small; In order to achieve a long detection distance, LiDAR requires the detector's receiving aperture to be as large as possible, and correspondingly, the size of the MEMS mirror is also as large as possible. However, the larger the size of MEMS, the higher the requirements for materials and processes, and the greater the reliability challenge.
Hesai publicly released the PandarGT based on MEMS technology in 2018, and later switched to a one-dimensional mirror scheme. It is said that the main reason is that the MEMS scheme has a small aperture and is not conducive to ranging. Another company in the MEMS industry, Innoviz, has not had smooth progress in its mass production project at BMW, and there were even rumors of "poor development progress" at one point.

Another representative of this route is the two-dimensional rotating mirror scheme, which, as the name suggests, involves scanning two rotating mirrors in different directions.

The scanning component of Luminar includes two scanning mirrors. In the above figure, 12 is the horizontal scanning mirror and 14 is the vertical scanning mirror; The Tudatong LiDAR planned to be installed on the NIO ET7 is also of a similar design. In fact, the prism scheme used by Livox is essentially a two-dimensional scanning.

This two-dimensional scanning scheme requires a scanning mirror to rotate at high speed in order to cover the entire field of view. In fact, in a two-dimensional scanning scheme, the scanning frequency of the scanner often needs to reach several hundred Hz, and the motor speed can reach several thousand revolutions per minute. High speed has a direct impact on the lifespan of the motor, making it difficult to pass the vehicle regulations. At the same time, the noise generated at high speed also needs to be overcome.

Next is the impact on performance.

As discussed earlier, the point frequency is determined by the number of lasers and the emission frequency of a single laser. The two-dimensional scanning scheme significantly reduces the number of lasers and can only compensate by increasing the transmission frequency of a single laser. However, there is an upper limit to the point frequency of a single laser, so its total point frequency will still be affected. While ensuring the refresh frame rate, some sacrifices can only be made in FOV or resolution.

In addition, due to the size of MEMS micro mirrors, the detection distance of MEMS will also be affected.

Route 2: Multi channel, one-dimensional scanning

The second technical route is achieved through multiple laser channels and one-dimensional scanning, and its typical technical route is the one-dimensional mirror rotation scheme. The scanning structure principle of this scheme is relatively simple: keep the transceiver module stationary, allowing the motor to reflect the beam of light to a certain range of space during the process of driving the rotating mirror, thereby achieving scanning detection.

The scanning frequency of the one-dimensional mirror rotation scheme is generally not high (not exceeding 10Hz), and this scheme has a precedent of being validated by car regulations, with good reliability - the SCALA jointly developed by Fareo and Ibeo was mass-produced on the Audi A8 in 2017.

However, the one-dimensional mirror rotation scheme is not perfect either. The most typical problem is that the number of lines that can be achieved depends on the number of laser transceiver units. According to the traditional path, the only way to increase the number of lines is to stack a large number of laser transceiver units. However, this not only leads to a proportional increase in cost, but also makes the system large and complex. Because of this, Fareo's SCALA 2 only achieves 16 lines.

The best solution to this challenge is to integrate hundreds of laser emitters and receivers onto several chips, and then significantly reduce costs and system complexity through Moore's Law. The rotating mirror scheme of Hesai is a combination of one-dimensional scanning and chip technology.

How to continuously reduce costs in the industry

After discussing these two options, let's discuss how hybrid solid-state radar can reduce costs.

Overall, there are two ways to reduce costs: scale effect and structural cost reduction.

scale effect 

Scale effect refers to the fact that LiDAR manufacturers share research and development costs by expanding their production scale, and reduce material costs through large-scale procurement. The transceiver and scanning modules of LiDAR follow this pattern.

Structural cost reduction

The meaning of structural cost reduction is the cost reduction brought about by simplifying and integrating module design.

Small cost reduction space for scanning modules

The two-dimensional rotating mirror requires two scanning structures, and there is not much room for further optimization in the structure. The mirrors made by MEMS technology are already several millimeters in diameter, and if the size is even smaller, it will further reduce the ranging performance. Since the size cannot be reduced, the space for cost reduction is also limited.

The one-dimensional rotating mirror structure is inherently simple, and there is not much room for cost reduction.

The trend of chip based electronic components for transceiver modules

The transceiver module includes various optical lenses, lasers, detectors, laser drivers, analog front-end, etc. Among them, the optical lens no longer has room for cost reduction. The founder of a certain LiDAR manufacturer said, "In recent years, the price of optical lenses has not changed much. Nowadays, buying a good lens still costs tens of thousands to tens of thousands, and the performance is not much different from a decade ago."

Therefore, the cost reduction of the transceiver module is mainly achieved through electronic components such as lasers, detectors, lasers, laser drivers, analog front-end, etc., and the cost reduction of these electronic components also relies on chip technology. Of course, the chipization of these electronic components will also drive optical lenses to evolve towards smaller sizes.

The trend of chip based electronic components for transceiver modules is already very obvious. In addition to the self-developed transceiver module chips mentioned earlier, other laser radar manufacturers have also developed their own chips. Velodyne stated that they have "completed the development of 8-channel laser radar ASIC chips" in their solid-state products, and Ouster has also customized and designed chips to integrate VCSEL lasers and SPADs into a single ASIC chip, in order to improve integration.
The trend of chip based technology can make Moore's Law effective in the field of LiDAR - by integrating lasers and other components onto chips, it can significantly reduce installation and debugging costs while reducing material costs. In the future, by continuously improving the semiconductor process, chip costs can be further reduced.

Prior to this, the cost reduction effect brought about by Moore's Law had already appeared in the field of cameras. Taking mobile phone cameras as an example, in the past decade, while the total cost has remained basically unchanged, the pixel count of mobile phone cameras has increased from 1 million to 100 million, and the average cost per pixel has decreased to 1/100 of the original.

Similarly, the trend towards chip based electronic components in LiDAR transceiver modules has made it possible for the number of lasers to continue to increase, ultimately resulting in higher point frequencies for LiDAR without a significant increase in cost.

At present, some manufacturers independently develop transceiver chip processes at the tens of nanometers level. With the improvement of integration and process technology, there is still huge cost reduction space in the future.

For the "few channels+two-dimensional scanning scheme", although cost reduction can be achieved through chip based implementation, the laser channels themselves are few, the cost proportion is not high, and the reduction space is limited.
For the "multi-channel+one-dimensional scanning scheme", due to the multiple laser channels, there is a lot of room for cost reduction through chip based implementation.

At present, the front-end mass production project is the easiest scenario for LiDAR production, and it is also a battlefield that major LiDAR manufacturers are competing for. To achieve front-end mass production, it is necessary to meet the vehicle's specifications.
In the industry, people are talking about car grade, and many manufacturers may even self label themselves as having reached car grade. So what exactly is car grade?

Car grade refers to the ability to pass a series of certification tests by car companies, obtain project targets, and start mass production. Car companies will have complete and systematic testing and verification, especially for global major brand car companies. Their testing system is complete and the testing conditions are strict. They have passed their certification and have strong industry credibility. Currently, it is widely recognized in the industry that the only car that has passed the standards is the SCALA series from Fareo.
Under the trend of chip based transceiver modules, compared to previous discrete devices, chips are more integrated, with significantly improved consistency and reliability. Overall risk is controllable and certainty is high.

For the scanning module, the one-dimensional scanning scheme has been validated by Fareo, and the scheme is mature and reliable; Two dimensional scanning, due to its more complex design and involvement of high-speed scanning components, still faces considerable controversy regarding its reliability.
Let's return to the question at the beginning of this article, what are the core technological barriers in the field of LiDAR, such as engine technology in the automotive industry?

The true core barriers will not easily fail with the evolution of technological routes. As LiDAR moves from mechanical to hybrid solid-state, and ultimately to pure solid-state solutions, although scanning schemes have been constantly changing, the electronic architecture of the transceiver module can match different scanning schemes with just minor adjustments, meeting the needs of different scenarios for FOV.

For example, the Robotaxi market requires a 360 ° horizontal FOV, and the transceiver module can be combined with mechanical radar to perform 360 ° scanning. The ADAS market requires a compact LiDAR with a certain field of view angle, which can be combined with MEMS, one-dimensional or two-dimensional mirror scanning schemes to achieve a horizontal 120 ° scanning. But no matter how it is combined, the underlying transceiver system is the foundation, just like passenger cars are being upgraded every year, but engine technology remains the core technology that continues to play a role.

In addition, after the reliability of a certain scanning scheme has been verified in the market, a deep accumulation of laser radar manufacturers in the transceiver module can quickly switch to this scanning scheme. In fact, major LiDAR manufacturers have technical reserves in different scanning schemes: Hesai has developed mechanical and explored MEMS schemes, and currently has chosen a one-dimensional rotating mirror scheme for hybrid solid-state; Although DJI Livox ultimately chose the prism scheme, it also explored the MEMS scheme in the middle; Sagitar Juchuang has worked on mechanical and MEMS systems; Although Huawei uses a one-dimensional mirror rotation scheme, it also has patents and technological reserves in the MEMS field.


The transceiver electronic system that determines the performance of LiDAR is the core barrier of LiDAR, and the chip based transceiver system is the concentrated embodiment of the core barrier of LiDAR.

The chip based transceiver module can continuously improve consistency and reliability through automated processes under the influence of Moore's Law, continuously reduce material costs exponentially, and thus assist in the large-scale production of LiDAR, accelerating the commercialization process of the autonomous driving industry.