Millimeter-wave radar is aimed at autonomous driving. Will 4D imaging be the "optimal solution"?

Musk tweeted last month that Tesla's FSD Beta 9.0 version will no longer rely on radar. If it really arrives, it means that Tesla has regrouped and returned to a purely visual autonomous driving solution.

At present, Tesla has updated the Model 3 promotional page on its North American official website. The content about FSD only retains the visual and ultrasonic sensor parts, and the previous information about millimeter-wave radar has been removed. Instead, it is replaced by: "250 meters of powerful visual processing capabilities."

Unlike Tesla's ambiguous attitude towards millimeter-wave radar, some players at all ends of the industry chain regard millimeter-wave radar as an important component of autonomous driving and assisted driving, but it has not been discussed much in the past. When we talk about the battle of perception routes, we discuss cameras vs. lidar, but in fact, both solutions incorporate millimeter-wave radar.

However, in the middle of last year, after Continental launched the world's first 4D imaging millimeter-wave radar, it seemed to become a small outlet that was about to emerge. On the eve of this year's Shanghai Auto Show, Huawei also held a HI new product launch conference and released a new generation of high-resolution 4D imaging radar, which is used as the core sensor of autonomous driving solutions.

Why is there little discussion about traditional millimeter-wave radar in the automotive market? How is the performance of the 4D imaging radar that has appeared in the past two years?

Millimeter-wave radar, the lost vertical and blurred horizontal dimensions

Millimeter-wave radar is still relatively new to people, but in the industry, automotive millimeter-wave radar has more than 20 years of successful commercialization experience and market accumulation in Europe and the United States. Data shows that the top five global giants such as Bosch, Continental, and Denso account for more than 75% of the market share.

To date, millimeter-wave radar is a relatively mature automotive perception sensor among traditional giants. It has a relatively low cost, and its perception fusion with cameras is also the preferred solution for realizing L2 assisted driving in the early stage.

Traditional Tier 1 not only occupies the "cake", but also holds the "knife" to cut the "cake". Therefore, it is difficult for new players to tear a hole in it and affect the long-term interest relationship between them and downstream car companies. Naturally, new players focusing on millimeter-wave radar will not choose to "go head-on" with these Tier 1s, and will prefer to explore new technological innovations to achieve overtaking.

The market structure tends to be stable, which is one of the reasons why traditional millimeter-wave radar has a weak presence in the eyes of the outside world. Another reason is that when autonomous driving moves towards L3 and L4, the technical deficiencies of traditional millimeter-wave radar are gradually magnified. In the view of "Smart Relative Theory", there are two aspects:

1. The “invisible Y-axis”

Traditional millimeter-wave radars are lacking in longitudinal height measurement capabilities, which can be understood as a lack of ability to "understand" the vertical plane.

Because of the lack of this "understanding" ability, the millimeter-wave radar cannot "see" the height of bridges and road signs, for example. In its "eyes", these stationary objects are regarded as being on the ground plane. Based on this premise, if the signals reflected by them are not completely filtered out, the millimeter-wave radar will undoubtedly issue a false warning of obstacles ahead, causing "ghost braking".

However, when there are stationary vehicles under bridges or road signs, traffic accidents may occur.

Tesla has had several accidents where it crashed into trucks, which is a typical case. Among them, Tesla's camera perception failed and could not identify the truck stopped in front. The millimeter-wave radar, as a backup sensor, should have identified the obstacle in front and issued a warning, but the millimeter-wave radar did not work either.

Because the information of the stationary truck is mixed with those information, the former will also be filtered out by the radar's recognition algorithm. The millimeter-wave radar recognizes the stationary object, but therefore "ignores" its existence. In this way, the millimeter-wave radar is invisible, the camera is ineffective, and the self-driving car becomes blind and deaf, and eventually hits the stationary truck.

2. “Fuzzy X-axis”

Another limitation of traditional millimeter-wave radar is its low lateral resolution, which can be understood as its weak ability to "understand" the horizontal plane.

The lateral resolution refers to the angle formed by the two scanning laser points on the left and right. The smaller the angle, the higher the lateral resolution. Compared with laser radar, the lateral resolution of millimeter-wave radar does not have an advantage.

For example, Tesla has had problems in the past: a car in front is parked next to the road, and half of the car body may be in the lane. At this time, Tesla will not be able to identify the vehicle due to insufficient lateral resolution of the millimeter-wave radar, making it more likely to crash into it.

Regarding this issue, Tesla had two accidents in Florida because the millimeter-wave radar could not measure the lateral speed, resulting in the inability to recognize that the vehicle in front was moving, which ultimately caused the vehicle to have no time to brake.

Therefore, in general, traditional millimeter-wave radars have deficiencies in "understanding" vertical and horizontal planes, which also determines that traditional millimeter-wave radars are difficult to adapt to sensors of high-level autonomous driving perception systems. However, the recent active 4D imaging radars seem to give us some new possibilities.

From quantitative change to qualitative change, two routes for 4D imaging radar

In a way, the emergence of 4D imaging radar is driven by the outstanding performance of laser radar and camera. The requirements for perception in autonomous driving have increased, and the camera has been upgraded from 2 million pixels to 8 million pixels. Semi-solid laser radar has begun to be installed in vehicles at an accelerated pace. The original millimeter wave radar has also been upgraded to the current 4D imaging millimeter wave radar.

Judging from the demos or PowerPoints released by manufacturers in the past two years, 4D imaging radar has indeed technically solved some major shortcomings of traditional millimeter-wave radar and magnified the advantages of millimeter-wave radar. Continental and Huawei, which participated in this year's Shanghai Auto Show, are representative manufacturers.

In this process, the general implementation method is nothing more than increasing the number of transceiver channels in hardware, expanding the antenna aperture while meeting the requirements for resolution in the horizontal and vertical directions. In this way, in addition to the traditional millimeter-wave radar, a vertical dimension of information is added, that is, 3D is upgraded to 4D. This is also the mainstream solution for 4D imaging radar at the Shanghai Auto Show.

In fact, from a technical perspective, we see more "quantitative change" rather than "qualitative change" brought about by technological innovation. If manufacturers only focus on highlighting parameters, it is inevitable that they will be suspected of "over-marketing". It is the real use of it to promote the implementation of related products that reflects the "qualitative change". In the view of "Smart Relative Theory", there are two feasible directions:

1. Focus on segmented needs and find new ideas for monetization in commercial scenarios

4D imaging radar has the ability to operate in all weather conditions without fear of heavy rain, strong light and other harsh environments. On the other hand, 4D imaging radar is an "upgraded dimension" of millimeter wave radar and also continues its past advantage, which is cost.

These two features actually give it the opportunity to realize and commercialize unmanned driving technology in commercial scenarios such as closed parks. Among them, "unmanned delivery car" is a suitable option, because at this stage, its mass production still has problems with sensor cost performance and all-weather operation, which can be improved by 4D imaging radar.

Unmanned delivery vehicles travel at low speeds, and currently most use laser radars with fewer wire harnesses. The rapid development of laser radars has led to a reduction in costs, but they are not yet affordable enough. The laser radar-based sensor solution makes the cost of a single vehicle too high, which is the main obstacle to the commercialization of unmanned delivery. This has long been an industry consensus.

Therefore, people are very sensitive to the cost of sensors for unmanned delivery vehicles and would be happy to see an unmanned driving solution that guarantees adequate performance and is more cost-effective. This may be possible through the combination of multiple 4D imaging radars.

For example, Gaogong Intelligent Automobile reported that the unmanned logistics vehicle jointly built by Great Wall Motors, Yihang Intelligent and Oculii provides customized perception solutions for low-speed logistics park scenarios based on 4D imaging radar. The unmanned vehicle can perform point cloud imaging, identification and tracking of 360° low-speed or stationary pedestrians, obstacles and small objects around it around the clock, and its point cloud effect is very close to that of LiDAR.

After further reducing costs, 4D imaging radar can help shorten the mass production cycle of unmanned vehicles and reduce the difficulty of commercialization. Therefore, although 4D is just a "dimensional upgrade" of 3D, it has better "cost efficiency" and higher "added value" in closed scenes. It is foreseeable that, out of considerations of balance, it will replace low-cost lidar in more specific scenarios.