"Automatic driving" should be said to be "assisted driving" cars. It is common to turn a blind eye to some behemoths. On the evening of December 12, a Tesla Model s suddenly lost control and entered a residential building at the speed of 80km / h in a community in Beijing. Big trucks, buildings... What will we do to ensure the safety of automatic driving in the future?
Huge objects are the bane of "advanced assisted driving"
Who could be the Savior?
In October, Tesla software hacker "green" broke the news and found a new function prompt in Tesla software update - a new radar option called "phoenix", which is the name of Arbe radar. This has something to do with Musk's recent remarks:
Now image recognition still comes from isolated images. In fact, these images are very related in time. Then, when we transition to 4D and add the time dimension to the 3D space, it will be like video. This architecture is what is really needed for fully automatic driving.
Musk's words have been further confirmed. The source said that Tesla is preparing to update model 3, and the new sensor will be part of the update. Previously, Arbe CEO and co-founder Kobi Marenko had confidently told the industry: "super high resolution 4D imaging radar is expected to upgrade the radar from auxiliary parts to automatic driving system, upgrading from redundancy to automatic driving core."
Phoenix is Arbe's radar chipset solution. It identifies, evaluates and responds to common to special challenge scenes through 4D ultra-high resolution real-time imaging, serving the driving needs of the real world. Regardless of speed, altitude, distance, size or surrounding weather and lighting conditions, Phoenix can distinguish between real threats and false alarms to ensure the safety of drivers, pedestrians and other vulnerable road users.
Musk: it suits me
Because to the point, marenko's words musk likes to hear:
Marenko said: "after the difficult years of many car accidents, automatic driving is facing reasonable concerns. The discussion around these tragic events focuses not only on the ethics of artificial intelligence (AI), but also on whether the technology behind it is ready. "
"Today's sensors are not mature enough to support tomorrow's automatic driving," he said. However, 4D imaging radar can achieve the required level of security. "
To sum up, 4D imaging radar has the following advantages:
Real time obstacle detection:
Under all weather and lighting conditions, provide highly detailed environmental images with a wide field of vision, find various obstacles in real time, determine whether they are moving and in which direction, and provide real-time situation data and alarms to vehicles.
Remote detection:
To realize the longest distance detection among all sensors, it is most likely to become the first equipment to find danger. Then, it can guide the camera and lidar to the region of interest, greatly improving the security performance.
Path planning:
Provide true path planning because it can create a detailed image of the road within more than 300 meters and capture the size, position and speed data of objects around the car.
Object height separation:
Identify whether the object directly in front of the car (such as a bridge) is stationary, whether it must stop or can drive past safely.
Reduce processing and server requirements:
Since only the camera and lidar are aimed at the region of interest, the use of high-quality radar post-processing will solve the main problem of the current prototype - power consumption.
Significantly reduce production costs:
Even above L3, it is not necessary to equip each vehicle with more than one lidar unit. The mass production cost of autonomous driving sensor kit should be less than $1000, and the cost of components and systems used in some vehicles today is 100 times that price.
So, how is it different from the current radar on the road? Listen to marenko.
Nowadays, radar plays an important role in safety systems such as adaptive cruise control, blind spot detection and automatic emergency braking. However, the current radar technology on the market must choose between the medium resolution of limited field of view and the low resolution of wide field of view.
In order to realize L4 and L5 vehicles, the main engine factory must enter the next level of sensing technology, use high-resolution imaging radar to sense the environment in a wide-angle range of 100 degrees with high resolution at an azimuth of 1 degree and an elevation of 2 degrees.
In marenko's view, optical sensors, such as cameras and lidar, are certainly needed in the automatic driving sensor kit. However, 4D imaging radar solves the following problems to enable the vehicle to achieve the required safety performance:
Highest reliability in all weather conditions, including fog, heavy rain, dark nights and air pollution
According to the requirements of the automotive industry, detect obstacles up to 300 meters
Measure Doppler (radial velocity) per frame
Leave the POC (confirmatory testing) stage and enter mass production
The disassembly results of system plus consulting show that Tesla currently uses Continental's 2D Radar. Phil Magney, founder and President of VSI labs, said: "Tesla has turned to Bosch from the mainland in order to obtain greater range. But both are 2D radars and have limitations. " "In my opinion, Tesla's future shift to imaging radar will be a smart move," he said
Iteration of automotive radar technology
The semiconductor technology used in automotive radar is still developing after several generations of iteration.
The first generation of products used GaAs, which was originally a bare chip installed on a circuit board. The next generation is SiGe, with more on-chip integration functions and high carrier mobility. It can be used for high-frequency radar even at large lithography nodes (such as 130nm). Later, with advanced packaging, the technology developed from cob (chip on board packaging) to fowlp (fan out wafer level packaging).
At present, many companies are developing silicon CMOS and SOI technologies, using 40nm technology nodes, but also reducing them to below 28nm. Shorter channels can also support high frequencies in the case of low carrier mobility. With CMOS technology, higher function integration can be realized in the chip. At present, the latest generation radar chip not only integrates transceiver and chirp, but also integrates microcontroller and digital signal processing (DSP) unit. Turning to silicon technology will also better maintain the cost reduction route and be more suitable for mass production.
Development of semiconductor technology in automotive radar
With the progress of packaging, various forms of wafer level products have emerged. The circuit board has also evolved into a top RF layer composed of special low insertion loss materials, such as ceramic filled polytetrafluoroethylene or similar materials. When the small antenna array is enough to meet the requirements, the encapsulated antenna (AIP) design has been adopted, and some have been proved to be suitable for short-range automotive applications.
The traditional radar has the ability of two-dimensional space measurement. The 2D radar uses a special rotating antenna to monitor the echo signal to obtain two coordinates and determine the target position. After the 3D radar appears, it will rotate like the 2D Radar, but after each scanning rotation, the antenna elevation will change for the next detection. In this way, 3D radar can detect three dimensions: azimuth, elevation and velocity.
Where is the new 4D imaging radar?
At present, the automatic driving industry is still in the proof of concept stage, and the sensors it relies on can not meet 100% of the requirements. High resolution imaging radar seems to be the only sensor that can meet the requirements.
In principle, 4D imaging radar is very different from traditional radar and lidar. From a physical point of view, time is the fourth dimension. When applied to radar, it will become Doppler frequency to show whether an object moves towards itself or to a distance.
4D radar integrates the measured fourth dimensional data into 3D radar to better understand and draw environmental map. Even if the specific technology is not novel, its integration also has new ideas.
Magney explained: "I think time is the fourth dimensional space, because the time element is obtained from Doppler. Imaging radar essentially creates an array, so the measurement density increases sharply. Traditional 2D Radar is rough, and each object produces only one point. However, imaging radar can provide many points and produce vertical resolution, which can better understand what is being tracked. "
In other words, the time factor has always been the key to radar function. When asked the same question, Lehmann pointed out that the fourth element of 4D imaging sensor is "lateral resolution". "4D imaging radar can recognize not only the horizontal plane, but also the vertical plane. For example, a car can decide whether to pass under or above an object," he said
He continued: "imagine a car traveling at 80 kilometers per hour on the highway and a motorcycle (small object with low reflectivity) coming from behind at 200 kilometers per hour. Unlike cameras and lidar, 4D radar can detect motorcycles that were originally far away and recognize that the two objects are moving at different speeds. "
With the emergence of high-resolution imaging radar, many radar suppliers are eager to upgrade the radar to the only high-speed sensor that can work in bad weather and light conditions.
4D technology erodes lidar Market
Perhaps, in his early attempts at automatic driving, musk made a clear decision to use cameras supplemented by radar and computer vision.
Now, there is an imaging radar that produces point clouds. Although its resolution is lower than lidar, it is much better than traditional radar. In Magney's view, lidar now has more advantages than radar. With the emergence of new radar, this gap may be narrowing“ One of the key applications of lidar is relative positioning based on basic maps. This is essential for autonomous taxis and shuttle buses, so lidar is the first choice for this task. Even if 4D radar appears, lidar will not disappear. "
From the technical trend, the point cloud of radar is becoming more and more dense, and various methods have emerged to enable radar to detect, classify and track many objects in 3D space under all weather and lighting conditions. With the emergence of larger artificial or semi-automatic labeled training data sets and the development of fusion technology, these efforts are expected to be significantly improved.
Indeed, the capabilities of radar technology are expanding rapidly. First, the antenna array is getting larger and larger. Some startups have designed and demonstrated radar chips that support 6500 virtual channels. This radar can achieve 1DEG (elevation unit) azimuth and 2DEG elevation resolution. This trend will enable a wealth of 4D data points per frame to provide accurate information about speed, distance, azimuth and elevation. Therefore, such radars will erode the areas currently occupied by lidar, although the latter may retain the advantages of angular resolution and potential target classification.
Dense high-resolution point clouds will make target detection, classification and tracking based on radar data possible. In order to achieve this goal, people are developing deep learning technology. A major challenge is the lack of extensive marker training radar data. The manual labeling process requires the input of experts, so it is costly and time-consuming. Now, although some companies have deployed camera, lidar and other radar data acquisition fleets with data synchronization, and are developing semi-automatic label technology, this technology still depends on the later data fusion between camera, lidar and radar. These efforts and techniques will accelerate the development of training data sets.
Application of thunder, heavy rain point is small?
In the whole process of transition to a higher level of automation, the main engine plant should be able to reuse software and hardware using a single platform in order to solve the problems of different brands and models. Only a highly integrated packaged solution can facilitate the transition to 4D imaging radar.
There are not many start-ups and manufacturers developing 4D imaging radar. After combing, we can see the context and trend of this technology.
NXP processor and transceiver
In order to promote L3, L4 and L5 vehicles with safety as the core goal, NXP has released a series of new radar sensing technologies.
The second generation 77GHz RF CMOS radar transceiver series tef82xx of NXP is the first automotive radar transceiver using 40nm RFCMOS in the industry.
Compared with the first generation, the RF performance of the new product is doubled, the output power is 13.5dbm and the noise figure is 11.5db. The device can also reduce the phase noise by four times around a given target.
Arbe chipset Tesla favorite
Providing Tesla with new 4D sensor technology is the start-up Arbe robotics, whose 4D imaging radar chipset won the 2020 Edison Award.
CEO Kobi marenko said that the physical resolution provided by this radar chipset is 2 to 10 times that of the synthetic or statistical resolution enhancement methods used by competitors.
In addition, the radar chipset provides real-time management of 48 receiving channels and 48 transmission channels. While maintaining the equivalent processing throughput of 3 megabits per second, it can generate a complete 4D image of 30 frames / second.
Rfisee phased array technology
Because phased array is an expensive technology, its adoption in the automotive industry has always been a worrying problem, which is why this radar was initially limited to advanced military systems such as F-35 fighters.
In September, rfisee released the industry's first phased array 4D imaging on-chip radar. It is a high-resolution, low-cost radar sensor that can generate real-time 3D position and velocity maps of objects around cars.
Using rfisee's radar on-chip solution, the receiver can ensure significant improvement of radar image, better signal-to-noise ratio and greater obstacle detection range such as vehicles and pedestrians.
Xilinx and continental are ready for mass production
Also in September, Xilinx and continental announced the launch of the first "mass production ready" 4D imaging radar in the automotive industry, which caused a shock in the industry.
Xilinx will pass zynq ® UltraScale+ ™ The MPSoC platform supports continental to develop a new advanced radar sensor ars540 and jointly build the first mass production 4D imaging sensor in the automotive industry.
The cooperation between the two sides will help carry out the new model of ARS540 to realize the SAE J3016 L2 function and pave the way for the development of L5 automatic driving system.
Write at the end
The development of automatic driving has come to a crossroads. Based on some concerns, the advanced mobile travel industry needs to re-examine the role of 4D imaging radar as an indispensable element in the automatic driving sensor suite, and provide more sensitive ears and eyes for autopilot cars, thus forming a safer automotive mass market.
Editor in charge: Tzh, read the full text“
Our other product:
