Because I have been in contact with Oculii for a long time, today I will share with you about Oculii 4D radar
The first one is: Innovation and Application Report of Aoku 4D High-Definition Imaging Radar on April 15, 2021.
Report entry: https://apposcmf8kb5033.pc.xiaoe-tech.com/live_pc/l_606f1355e4b0d4eb038fabc9 =
The other is a sharing from Aoku at this year's Suzhou Millimeter Wave Conference.
Special annotations and statements: It is not Ruanguang. The records and explanations in this article are from two reports, both of which are the level claimed by Mr. Qie.
-----------------------------------------------------------------------------------------------------------
Here is the text:
Both reports are made by Qi Jianjun, and Mr. Qie will do it. Let me briefly introduce Mr. Qie (the introduction is from the introduction of the Suzhou Millimeter Wave Conference):
Mr. Qie Jianjun has more than 30 years of experience in the automotive, Internet, mobile communications and public security industries. He has held senior management positions in Motorola, AutoNavi, and the Ministry of Public Security. In 2017, he joined Oculii as CMO and President of the Asia Pacific Region, and promoted 4D imaging millimeter wave radar in China for the first time. Facilitated the designation of mass production pre-installation projects for many OEMs, and the company was finally successfully acquired by Ambarella. Since joining Ambarella in 2021, he has led the Chinese market team and is committed to rooting and developing Ambarella's smart driving chip solutions in the Chinese market. Mr. Qie Jianjun holds a bachelor's degree in wireless communication from Jilin University and a master's degree in mobile and network communication from Harbin Institute of Technology.
Report one:
The report is made by Mr. Qie. The beginning of the article talked about his background, which is a high-precision map background. He also took Xiaomi as an example and mentioned some cooperation with Xiaomi. Similar to the demand for mobile phones, key components have become an urgent need for the industry and may become a decisive "inflection point technology".
The 4D corner radar can be made so small, the power consumption is about 2W.
The report is divided into 4 parts: industry status, virtual aperture imaging technology, Aoku's latest 4D products, applications in L3/L4 and other fields .
Mr. Qie mentioned that a certain OEM has already used their 4D radar on a fixed-point basis .
Mr. Qie started from the characteristics of the sensor and then mentioned Musk's views on the sensor at that time:
Apparently Musk believes that vision is better, while radar tends to cause problems.
It is not sensitive to traditional radars that cannot measure height, have low angular resolution, traverse objects and stationary objects.
The technical route of 4D radar:
Aoku’s hardware is similar to other companies. The angle radar mentioned above uses TI’s 1843. Its algorithm + unique antenna design makes it have excellent performance and low material cost.
Waymo has actually installed 6 4D radars and formed a high dependence (waymo adopts 4 cascading methods, and has contacted Aoku to optimize its radars from an algorithm perspective):
SAIC's ZF 4D radar adopts four cascaded TI2243 + FPGA detection range of 300m
Continental 540
Huawei: Four-cascade solution
It can be seen from the above that the power consumption is relatively high and the size is relatively large .
Arbe's special chip: (48 transmissions and 48 receptions) There are many antennas and the noise is relatively large.
Mobileye：
Next, Mr. Qie explained Aoku's technology stack:
Ordinary radar technology stacks are all in this way, adding antennas; (millimeter wave on the aircraft is almost the most important sensing device, and it can achieve a hundred shots and a white harvest)
The way Aoku uses the algorithm can increase the resolution by 100 times, claiming that its waveforms are different (AM, PFM, FM, all three are modulated, and adaptive). The previous frame and the next frame will generate a closed loop and correlation, Varies according to environmental changes. Then demodulate and extract the information. The software virtualizes 20 times or even 100 times the antenna.
As can be seen from the above picture, they are also using their algorithm in the Raptor, which sends out three times and receives four times.
The power consumption is very low, and the four-cascade can even reach the level of the laser (0.1 degree, the field of view is very large, and no FPGA is used).
product comparison:
4 corner radars + 1 forward (Robotaxi scene)
High-speed scene:
It can be seen that all stationary objects with height information can be detected. Lasers can also be added to form a sensor group.
Marking laser: (compared to Velodyne's 16 lines)
For lower computing power requirements, it can be completed directly on the SoC.
In high-speed scenarios, it relies more on millimeter waves. When the cost is high, they can be installed to obtain better perception. If the cost is low, only the millimeter-wave imaging radar can be installed.
Next, the parking scene is mentioned , especially for small targets such as ice cream cones and triangular cones, which are difficult in the industry :
In indoor scenes, multipath is very serious and also a difficulty. Its effect on Amazon is as follows:
Point cloud construction:
Positioning can achieve 10cm accuracy. Radar SLAM, very novel.
Unmanned logistics car - SLAM mapping, positioning and navigation (a project for Great Wall Motors), the right is the processed map, swinging 15cm left and right.
Complicated intersections: Mr. Qie emphasizes the cooperation with vision
Fusion with vision:
Team introduction: (There is even a former Continental CEO)
cooporating company:
In the Q&A session after sharing, Qie always responded to individual questions in the form of voice. Here I will select some parts to sort out:
1. What chip is used?
TI's 1843, 2243, etc. are all in use, as are Infineon and NXP, and they don't depend on a certain chip.
2. What is the anti-interference performance, and how is it guaranteed to deal with vibration?
There are mainly three kinds of modulation, each chirp is different, and the radar is different, the chance of interference is less, and related algorithms have been done; vibration and impact have been tested to ensure it.
3. Was SAR imaging used?
did not use. There are certain limitations on the vehicle, such as real-time (requires accumulation), and no motion compensation .
4. How to solve multipath interference? How to do target classification?
Phase modulation will cause comparative multipath problems, but they have related patents and algorithms to solve; they have more radar points, but they also need to do (deep) learning to get better results.
5. How to integrate with IMU and camera?
The industry has made a lot of efforts, and post-fusion is the mainstream practice in the industry. However, there are a lot of radar point clouds, so you can try to do pre-fusion. Blend before general use.
6. Has automatic parking been done?
It is a good direction that is being explored. Corner radar has great promise, and Zongmu Technology has already done related projects.
7. What are the dimensions of 4D?
xyz+speed or several angles+speed
8. Aoku's virtual aperture array technology greatly improves the angular resolution. However, the general aperture expansion technology generally increases the physical aperture. I would like to ask, did Aoku construct virtual array elements by sacrificing time or Doppler, or did it use some prior information?
None sacrificed, dynamically intelligent.
9. Waymo cooperation, environmental perception and adaptive transformation, according to the analysis of waveform and reflection intensity, the point cloud has 36000 points.
10. Is there any increase in reliability verification?
Some indicators have been added to the traditional mmWave.
11. What is the relationship between 4D point and laser?
Not to replace the laser, 4 corner radar + forward + laser. It is easy to arrange without changing the structure, and the occlusion problem is not serious. The low wire harness has a certain relationship (can be replaced), but the high wire count still needs to be complementary. Although the price of ultrasound is cheap, it is technically feasible to replace it.
12. Price?
The cost is ordinary millimeter-wave chips, and the BOM price is low. Separate calculation for software
13. Refresh rate?
10~20Hz adjustable, can be adjusted according to needs. (accepted the sports car project)
14. What about the latency?
Like standard normal radar, no additional latency is sacrificed.
15. What is the bandwidth?
76~77,500M; 77~79 1G is also available; problems can be solved within one G,
16. Are there publicly available datasets?
nothing now.
17. Advantages compared with Arbe and Senstech?
As mentioned above, the complete virtual aperture and algorithm empowerment are effective and cost-effective.
18. Is there any interference from multiple radars for oncoming vehicles?
Specially tested, the radar does not interfere with each other, the signal of each chirp is different, and the two radars are not the same as each other. For other people's radars, such as Continental or Bosch's, there is basically no impact, only when the slope is similar, there will be a strong impact; in the vehicle-road coordination, it has also been tested, and it does not interfere.
19. How to achieve 20~100 times the virtual aperture without time accumulation?
The uniqueness lies in the phase modulation, which is also the different phase in each chirp, and relevant information can be extracted during demodulation. You can refer to its patents, which are mainly horizontal and vertical, and the things released separately are different. Horizontal and vertical 10 times, 4 cascades can reach 100 times.
20. What is the number of points per frame for falcon as an example?
2000 points per frame, in fact, the original point cloud is 100,000, which can be achieved by using a single TI 1843, without DSP and FPGA processing.
------------------------------------------------------------------------------------------------
Report 2: Suzhou Millimeter Wave Conference
Topic: 4D Imaging Radar enters AI Era ——Ambarella Semiconductor Technology (Shanghai) Co., Ltd.
First, Mr. Qie gave a brief introduction to Ambarella:
Ambarella is a semiconductor chip company that started as a video chip company. In 2015, it acquired VSLAM and so on.
CV3 chip: Domain control chip with high computing power
(Supplementary reading: A single chip can build an automatic driving system? Ambarella CV3 series chips, the first batch of samples will be launched in the first half of 2022 )
The next thing to talk about is 4D radar: (can you "squeeze" the sensor technology to the extreme)
At first, we talked about the classic problem of strong and weak light in rain, snow, haze, and distance measurement and speed measurement. And visual + radar accidents (x Peng car, x Tesla, etc.)
Reason analysis: mainly the classic defects of two kinds of vision and the problems of traditional radar.
The main thing is to talk about its virtual aperture technology PPT with the previous report:
Adaptive waveforms (same as last report) are followed by only some new additions:
Basement height display:
Rainstorm weather: You can see 300~400m
Separate two angle counter-tests: up to 0.6 degrees
It can do the pre-fusion of the original point cloud level, Ambarella and Aoku have complementary advantages:
Summary: The report mainly expounds the advantages and technology stack of 4D radar, as well as the mainstream method of implementation, the current solutions of different manufacturers, and the status of their products in the current industry. It can be seen that the current 4D radar has entered the fields including mass-produced vehicles, map building and positioning, etc. It can be said that it has a very promising prospect!