Why don't China car companies let go of the technology eliminated by Tesla? Analyze BEV perception

With Tesla HW4.0 completely radar-free hardware as the hardware, Tesla's new generation of pure vision-assisted driving software and hardware will also be accelerated to land in China after being declared by the Ministry of Industry and Information Technology. Starting from the later period of HW3.0, Tesla began to adopt the technology of occupying the network, and put the vehicle itself in a 3D world, which can better solve some corner cases of assisted driving.

Tesla FSD Beta has adopted BEV technology since 202 1, and has used professional network technology developed by BEV since last year. Last month, domestic suppliers of new power vehicles and autonomous driving solutions began to launch BEV technology in unison.

Weilai will complete the switching from new NOP+ to BEV in the first half of 2023. The latest BEV technology is also integrated into the urban assisted driving function recently released by Tucki P7i, and other manufacturers such as Mimo Zhixing, Baidu Apollo, Ideality, etc. have also released the recent progress related to BEV.

BEV, a visual technical solution eliminated by Tesla, was taken over by domestic manufacturers in a short time. Why? It is said that Tesla's pure vision is unsafe, but is its visual technology really fragrant?

In order to reduce the hardware cost, Tesla had to eliminate BEV?

The full name of BEV is bird's eye view, which means bird's eye view. Let's take Tesla as an example: Tesla models use eight cameras to collect images and fuse them effectively. Of course, this is different from our most common 360-degree circular car image mosaic. The system corrects all the images collected by the camera and inputs them into the neural network, thus directly extracting features. Then, these features are correlated by using a transformer based on self-attention mechanism, and then projected into a vector space. Tesla, which has not removed the radar before, will add some radar data, and finally get a bird's eye view reflecting the surrounding environment.

This bird's-eye view seems to open a god's perspective, allowing the vehicle to put the nearby perception on a plane, increasing the scope and redundancy of perception as much as possible. However, there is an inherent BUG in this technology, that is, the bird's eye view is still a 2D image, just like when you look down vertically as much as possible when flying, you can't tell the ups and downs, and it is difficult to perceive the depth or gap, just like buildings or mountains and the surrounding flat land.

Therefore, before and after using BEV technology, Tesla models will still encounter the problem that static objects cannot be identified or ghost brakes, because although the system can see the objects, it may still be unable to identify what the objects are, and it still relies heavily on the pre-calibration and classification of car companies in the R&D process, so it can be avoided in cognition, and it is too late to react if it is unclear, so it will directly hit.

However, only by studying and calibrating in advance, the system will always be on crutches, and it can't be foolproof in the face of unexpected situations on the road, such as temporary construction and scattered objects. Of course, if the driver is the last guarantee in the field of assisted driving, it is good to say, but this technology is still difficult to apply to autonomous driving.

So you will see that Tesla has evolved from BEV to occupying network. Although occupying network is also an extension of BEV technology, the biggest difference between them is that the perception of the system has changed from 2D to 3D.

Tesla put himself in 3D space, so that all obstacles can be displayed in 3D space with a large number of objects. The system can output the occupancy probability of each 3D position around the vehicle to the calculation unit within 10 millisecond, and can predict the blocked obstacles in real time. Tesla doesn't have to worry about what an object is, nor does it have to identify and classify it. As long as it knows the approximate shape of the object, it will know whether to hide. Of course, let's take Tesla for example. Mobileye's latest super vision also adopts a similar 2D-to-3D technology model, because only two mainstream autopilot manufacturers are still insisting on visual solutions.

Wei Xiaoli relies on heap hardware to make up for technical shortcomings, but in the end, is Tesla right?

So why did Wei Xiaoli turn to BEV eliminated by Tesla at this stage? Tesla was eliminated precisely because it did not have some advantages of Wei Xiaoli. Wei Xiaoli dared to use this technology precisely because the hardware was piled high enough.

Whether it is BEV or network occupation, or the more advanced technology that HW4.0 may bring, Tesla is constantly involved in visual algorithms because of the lack of radar sensors, especially sensors that can scan 3D space. If we simply understand Tesla's occupation of the network, it can be regarded as a new route, and we don't need lidar to force ourselves out, because the visual plane perception can't have 3D effect, and we can't bring 3D glasses to the vehicle, so we can only roll it up on the algorithm architecture.

However, Xiao Wei ignored them, as well as most domestic suppliers of autonomous driving solutions. Everyone has chosen the lidar route, and lidar can really cut corners for everyone. Whether you have the first-Mover advantage or not, using lidar will speed up the development of autonomous driving. Although BEV technology can only give a bird's-eye view in 2D, lidar can give a 3D perception effect. Many manufacturers' lidar is arranged in a higher position, so that it can have a better view. Moreover, many models use more than one lidar, which can actually have 3D perception in all directions.

The laser beam emitted by lidar can draw an approximate image of an object through a point cloud. Some lidar with high equivalent beam can almost have a certain imaging ability, and it is 3D rendering. The visual advantages of BEV technology can be absorbed, and the lack of perception accuracy can be supplemented by laser radar or 4D imaging millimeter wave radar.

Although BEV looks a little outdated, BEV technology has become more and more important since smart driving began to enter the city. The aerial view clearly reflects the surrounding objects, but its cost is still high. Li Xiang said at the communication meeting some time ago that if you want to be a city-assisted driver of BEV, you may need to invest more than 65.438+0 billion yuan, so don't think that autonomous driving has been rolled up to the end and there is no money to burn. With such a large investment, it is almost impossible for the new forces to make a profit in a short time.

abstract

Although the technology eliminated by Tesla is used, domestic manufacturers should continue to make BEV technology bigger and stronger, because Tesla rolled up software to reduce costs, and domestic manufacturers piled up enough hardware and computing power to make up for Tesla's weaknesses, so they have the ability to overcome some bottlenecks faced by BEV technology.

Of course, there is also an unavoidable problem, that is, the perception priority judgment of multi-sensor fusion scheme. Tesla is purely visual, and there is no need to ponder the perceptual conflict between radar and vision, while other car companies and suppliers have to face this problem. If the perception of 2D BEV bird's-eye view conflicts with that of 3D lidar or 4D imaging millimeter-wave radar, who should you listen to?

To solve the above problems, we still need to work hard on visual perception. The crutches of lidar are likely to become the shackles of the future development of autonomous driving.

This article comes from Luca Auto, the author of Easy Car, and the copyright belongs to the author. Please contact the author if reproduced in any form. The content only represents the author's point of view and has nothing to do with the car reform.