After Tesla, why did Xiaopeng also abandon it?
Latest update time：2024-09-05
Reads：
"Only fools install LiDAR on cars. It's like the appendix on a human body. Its existence itself is meaningless." Musk said this in 2019, which made the whole world think that he was "demented". After all, the price of LiDAR continued to be reduced to the thousand yuan range, and even some car companies said "If you don't have four (LiDARs), please don't talk."
However, five years later, the whole world is in a state of confusion. Recently, automakers and Tier 1 companies have been lining up to abandon LiDAR and choose the pure vision route.
Xiaopeng launched a new generation of Xiaopeng AI Eagle Eye vision solution. At the same time, Xiaopeng's new car P7+ is also equipped with this technology, abandoning the lidar; the basic version of Wenjie M7 PRO intelligent driving has removed the lidar; Jiyue released the "pure vision + end-to-end large model" intelligent driving solution ASD; Weilai's second brand Ledao is equipped with a pure vision solution on its first model L60; Baojun Yunhai realizes city memory navigation and high-speed navigation under the binocular pure vision solution.
Some people say that domestic car companies have finally realized the difficulty of applying lidar, while others say that in most cases lidar may only be a backup for vision.
In short, the market trend has changed again, and intelligent driving players have begun to switch to pure visual solutions.
Fu Bin | Author
Electronic Engineering World (ID: EEWorldbbs) | Produced
Give cars different eyes
The so-called "pure vision" is easy to understand. It means installing cameras around the car and inside the cabin to achieve 360° environmental perception without blind spots. Only cameras are used for perception, without using any radar-related hardware.
However, the visual route also has another fork - the "full vision solution", which means removing the laser radar while retaining the millimeter wave radar and ultrasonic radar.
For example, Ledao L60 uses a Scion 4D millimeter-wave radar; Huawei's pure vision solution includes 3 millimeter-wave radars and 12 ultrasonic radars in addition to 10 visual perception HD cameras, and is mass-produced in the Wenjie M7 Pro, Zhijie S7 Pro, and Deep Blue S07 models; Jiyue's pure vision solution includes 11 HD cameras, 5 millimeter-wave radars and 12 ultrasonic radars, and Jiyue 01 and Jiyue 07 are equipped with this hardware configuration.
The eyes installed on the car are not the consumer-grade cameras we know. In-vehicle cameras must not only pass vehicle regulations, but also have high requirements in terms of field of view FOV, detection distance, resolution, signal-to-noise ratio, frame rate and dynamic range. The hardware structure includes optical lenses (including optical lenses, filters, protective films, etc.), image sensors (CCD or CMOS), image signal processors ISP, serializers, connectors and other devices.
At present, vehicle-mounted cameras are mainly divided into five categories according to their installation position: front-view cameras (installed on the front windshield, including front-view main cameras, front-view wide-angle cameras, and front-view narrow-angle cameras), surround-view cameras (installed around the vehicle body), rear-view cameras (installed in the trunk), side-view cameras (installed on the B-pillar or rearview mirror), and built-in cameras.
Among them, the price of front-view cameras is relatively high, with a market price of 300 to 500 yuan; the prices of other cameras are in the range of 150 to 200 yuan.
Depending on the application, vehicle-mounted cameras are divided into extra-cabin applications (parking assistance, driving assistance, CMS, DVR) and in-cabin applications (DMS, OMS).
In addition, as technology continues to advance, binocular vision technology has also begun to be used in cars. This technology is like human eyes and calculates the distance of objects based on the parallax of the left and right cameras. Compared with lidar, binocular cameras are cheaper and the point cloud is denser, but the algorithm is more difficult.
Who makes the camera module and what is in the camera?
Upstream:
-
Optical lenses: Asia Optical (South Korea), Kanto Chenmei (Japan), Largan Precision, Yujing Optoelectronics, Sunny Optical, O-Film, Lianchuang Electronics, Jinghua Optics, Yutong Optics, Fulan Optics, Lenswang, Ruifeng Optoelectronics, Zhongrun Optics, Sanjing Technology, Raydin Optoelectronics, Lens Optoelectronics, Xinxu Optics, Ningbo Fengmei, etc.
-
Filters: Asahi Glass, Da Vacuum, Nippon Electric, Tanaka Giken, Ottolen, O-Film, Crystal Optech, Wufang Optoelectronics, Jingji Optoelectronics, etc.
-
Protective film: 3M, LG, Nice, Zeiss, Crystal Optech, Tianjin Hitech Environmental Protection, etc.
Midstream:
-
Lenses: Sunny Optical, O-Film, Lianchuang Electronics, Hongjing Optoelectronics, Teles Optics, Yutong Jiuzhou Optics, Phoenix Optical Advanced Optoelectronics, Zhongshan United Optoelectronics, Asia Optics, Largan, Yujing Optoelectronics, Largan, Maxell, Fujifilm, Nidec Sankyo, Sekonix, Kyocera, Ricoh, etc.
-
CMOS image sensors: ON Semiconductor, OmniVision, Sony, BYD Electronics, GalaxyCore, SmartSens, pixelplus, Samsung, SK Hynix, Panasonic, and Canon.
-
Adhesive materials: European and American companies such as Henkel, Dow Corning, DowDuPont, BASF, 3M, etc.; Japanese companies such as Nitto, Nippon Seiko, Epson, etc.
-
Serializer chips: TI, ADI, Longxun Co., Ltd., Kangzhi Technology, Ruifa Technology, Kun High-tech Xinxin, Yutai Micro, Ruitai Micro, Jinglue Semiconductor, Feiang Innovation, Renxin Technology, etc.
-
Computing and processing chips: Mobileye, Ambarella, Horizon Robotics, Flink, NVIDIA, Qualcomm, etc.
Downstream:
-
Optical background camera module suppliers: Panasonic, Sony, Hitachi, Samsung, Sunny Optical, OFILM, Yinchuang Technology, Huaruijie, etc.
-
Tier 1: Magna, Continental, Bosch, ZF, Valeo, Denso, Desay SV, Hirain, Haun Automotive, Huayang General, Tongzhi Electronics, Hikvision, Zhihua Technology, Zongmu Technology, Furuitech, etc.
-
Cross-industry enterprises in the mobile phone camera module industry: Q Technology, Sunwin, Truly International, Holitech, etc.
Solutions from different car companies
Tesla is a representative of pure vision, and has repeatedly stated that it "insists on using pure vision automatic assisted driving solutions." Tesla's explanation for this is that the iteration efficiency of multi-sensory solutions is low, it is difficult to simulate human driving, and it is even more difficult to deal with complex and changing road conditions and multiple information. Tesla's pure vision solution for automatic assisted driving uses cameras to observe and simulate human driving habits, bringing a safer, more cost-effective and comfortable car experience.
Specifically speaking, in terms of cameras, Tesla's FSD (full automatic driving system) HW2.0/2.5/3.0 versions are equipped with 8 different cameras, namely three front cameras (1 main camera, 1 wide-angle camera, 1 narrow-view telephoto camera), 2 front side cameras, 2 rear side cameras, and 1 rear camera.
Tesla's cameras have always been criticized for their low pixels. Compared with the domestic new forces that often have 8-megapixel and 15-megapixel camera modules, Tesla's mainly hovers around 1.2-megapixel. Under the HW4.0 hardware and software, Tesla's forward perception camera has been upgraded from 1.2 million pixels to 5 million pixels.
In terms of algorithms, Tesla has experienced a transformation from "feature extraction network RegNet" to "BEV+Transfomer" and then to "BEV+Transfomer+Occupancy Network". The decision-making and planning level has shifted from "Rule-based" to "Machine learning-based" since 2021. In January 2024, Tesla launched FSD V12 Beta, becoming the world's first company to mass-produce "end-to-end" neural networks, realizing the integration of perception, decision-making, and planning. Now, Tesla places great emphasis on the role of AI in visual recognition.
In terms of chips, HW1.0, HW2.0, and HW2.5 hardware system chips come from different suppliers, including Mobileye; starting with HW3.0, Tesla began to equip its self-developed and produced FSD chips. The FSD chip is responsible for graphics processing as well as data processing and deep learning. The current FSD chip has been developed to the second generation, namely the HW4.0 chip.
Xiaopeng is a representative of new forces and released a new pure vision intelligent driving solution - AI Eagle Eye on August 27.
It is said that the world's first Lofic architecture is adopted, the camera accuracy and visual distance are upgraded, the color resolution ability is further enhanced, and it can see clearly in low light, backlight, large light difference and other environments, and can clearly identify at night, rainy and foggy days, tunnels, etc. The AI Eagle Eye Vision solution releases 20% of the computing power, the real-time perception distance is 1.25 times higher than the previous generation, the recognition speed is increased by 40%, and the delay is reduced by 100ms.
The Xiaopeng P7+ will be the world's first model equipped with the AI vision Eagle Eye solution. At the same time, models previously equipped with lidar solutions will be updated synchronously via OTA to maintain a consistent experience.
The overall architecture of Xiaopeng P7+ is similar to that of Xiaopeng P7. P7 is equipped with 14 cameras, including 3 front cameras, all of which are 2 megapixels, namely long-range perception camera, medium-range perception camera, and long-range perception camera. The use of so many cameras makes P7 highly scalable, which is convenient for further OTA upgrades in the future. Now P7+ further uses AI technology to improve the visual system, and the overall approach is very similar to Tesla.
In terms of chips, Xiaopeng P7+, like Tesla, has begun to use self-developed chips. It is said that Xiaopeng's Turing chip is the world's first AI chip that is used in AI cars, robots, and flying cars at the same time, and is customized for AI large models. For L4 autonomous driving, it uses a 40-core processor and integrates a 2xNPU self-developed neural network processing brain. It is designed for L4 autonomous driving and has three times the computing power of existing chips. The end-to-end large model has increased the amount of effective visual perception information by 8 times, improved spatiotemporal sequences, and improved environmental understanding and prediction capabilities, greatly improving visual perception capabilities.
Mercedes-Benz S-Class is a representative of traditional OEMs, and the binocular stereo camera solution is the biggest advantage of Mercedes-Benz S-Class. Mercedes-Benz S-Class is equipped with 7 cameras. The front camera module consists of 2 binocular stereo cameras, each with a pixel of 1.3MP, and the supplier is Continental Automotive Electronics; a forward fisheye camera is installed on the bumper, with a field of view of about 120°; a side fisheye camera is installed on each of the two rearview mirrors; and a rearview camera is installed on the trunk, with a field of view of about 120° and 2 million pixels.
In short, judging from the practices of the above three types of manufacturers, they are currently using low- to medium-pixel cameras. With the development of technology, more car models will be able to use 8-megapixel, 15-megapixel and above car cameras in the future. In terms of chips, they are turning their attention to self-development.
Pure vision or lidar?
There are two reasons why pure vision and LiDAR have entered the whirlpool of public opinion: one is cost, and the other is technology:
First, LiDAR cannot compete with cameras in terms of cost. Even though the price of LiDAR has dropped significantly, and some manufacturers even claim that it will drop below 1,000 yuan, the cost of a camera is as low as 200 yuan, while the total hardware cost of Tesla's eight cameras is also 200 US dollars (about 1,417 yuan), which is equivalent to the hardware price of a LiDAR. The cost difference is huge.
High-end cars have been using very expensive sensor devices and multi-functional integrated chip systems from the beginning, but this is unsustainable. OEMs themselves are under great pressure to reduce prices. If they want to reduce costs, eliminating lidar and sensors will be the first priority, but at the same time, they must maintain the necessary functions that bring them core value and ensure the safety of the vehicle, so better pure vision systems are needed to reduce OEM price pressure.
Even if lidar is used, it is also used in conjunction with a visual recognition system. The use of visual recognition technology already requires a large amount of information to be processed. If lidar is added, the data that the computer needs to process will be even larger, and the performance of the computer on the vehicle is limited. In this case, using a single technology can reduce computing pressure and help reduce costs.
Secondly, LiDAR itself has certain technical deficiencies. For example, on rainy days, the LiDAR signal will also be reflected by raindrops. In this case, there is no "difference" between vision and LiDAR, and visual recognition may even be safer.
Mobileye previously told EEWorld that a pure camera system is an independent, mainstream, and safe solution. Many Chinese automakers have chosen a pure vision camera solution. For ADAS-level intelligent driving models, there is no need to add lidar or radar. A camera system with sufficiently powerful chips and AI algorithms is sufficient.
They said that LiDAR is only an alternative, it only provides more safety and comfort for the whole vehicle; when it comes to L3 and L4, LiDAR becomes a must, and at that time, not only the front LiDAR must be installed, but there must also be three long-range LiDARs around to assist in pursuit of higher safety.
In short, car companies are now paving the way for cost-effectiveness. What they are considering is how to balance safety and performance within limited costs. After abandoning high-precision maps, abandoning lidar is undoubtedly a very good solution.
However, although pure vision has recently returned to its peak, its future is still unknown. After all, the price of lidar continues to fall below the bottom line, the computing power of car computers continues to increase, and in the future we will have to pursue higher levels of autonomous driving. At that time, it may be difficult for pure vision to continue to make breakthroughs.
References
· END ·
What else do you want to see about automotive electronics? Please leave a message below to communicate!
You are welcome to set us as “
Star
” so that you can receive push messages as soon as possible.
Scan the QR code to follow: Automotive Development Circle, reply "
driving
"
Get Get free information packages on autonomous driving, assisted driving, etc.!
Scan the QR code to add the assistant and reply "join the group"
Exchange experiences with automotive developers face to face
京公网安备 11010802033920号