The author | Mizuki Gong
The source | engaged in car zero distance
What comes to mind when you talk about autonomous driving?
For now, autonomous driving is no longer a high-tech technology far away from us. On the contrary, it should be one of the fastest landing areas and scenarios in artificial intelligence at present. #汽车科技要闻 #
Huawei ADS ▲
Today, L2 assisted driving is still a bad street, but there are still two distinct genres on the market: "pure visual algorithm school" and "good use of lidar school".
Sharon Mech Dragon with 4 lidars
The pure visual algorithm faction led by Tesla as a whole conceives a very simple concept of smart cars: simulating the behavior of human drivers, and replacing human eyes and brains with cameras and chip algorithms;
In the face of congestion Tesla chose to overtake on the double yellow line
Companies that make good use of lidar, such as Google, Xiaopeng, Weilai and Ideal, are equivalent to installing a "superhuman vision system" on a smart car, and adding a clairvoyant buff to the driver under the premise of imitating human drivers.
Weilai ET5 with "horns" on the top of the head
So who are the two schools of thought that is the ultimate form of the development of autonomous driving technology? Today's car jun will start from Tesla, take you to talk about how the current companies are doing.
It takes three steps to put an elephant in the refrigerator; it also takes three steps to train a Tesla to be a "driver":
First, the surrounding traffic participants are modeled by passing cameras. Then add the created photo data to the neural network training set, and let the vehicle learn through the annotation method.
After this cycle is repeated and a certain scale is formed, the "brain" will have its own judgment ability, such as what car, what is a bicycle, what is a person, how far away they are from us...
This learning is not limited to Tesla's own internal testing, but also applies to the millions of Teslas speeding on the road around the world.
See watermark ▲
It is precisely because of the support of a large user base that Tesla can continue to use pure visual solutions without fear; but even with so much data, Tesla's autopilot system will still misjudge:
For example, identifying slow-moving trucks as clouds; identifying human projections as real people; USing STOP signs on billboards as road signs...
One day in May 2016, Tesla made a big news: due to the error in the judgment of Tesla's Autopilot system, a Model S with autopilot on autopilot burrowed under the container without slowing down, the roof was cut, and the owner died on the spot.
It's not hard to imagine the tragic situation on the spot, so it's easy to understand why to this day, there are still many people who can't fully accept the automatic driving of vehicles.
So much so that later even Tesla itself expressed the expression of automatic driving, from "automatic driving" to "assisted driving" to now "automatic assisted driving".
It is precisely because of the drawbacks of pure vision that many manufacturers have turned to the solution of visual algorithm + lidar. So what exactly is lidar, and what exactly can it bring to the vehicle?
Cameras are ultimately suitable for image recognition, but it is difficult to work in bad weather and is not suitable for building 3D environments;
Millimeter wave radar is suitable for use in various climatic conditions, and is also suitable for speed and distance judgment, but the resolution is very low, and it is not suitable for the identification of the type of item;
Lidar can make up for the shortcomings of cameras and millimeter-wave radar: it can build a more realistic 3D environment than cameras, and it has more accurate object recognition capabilities than millimeter-wave radar.
Lidar can be divided into the following three types: mechanical lidar, hybrid solid-state lidar, and pure solid-state lidar.
The research and development progress of mechanical lidar is relatively early, and the maturity is relatively early, but the disadvantage is that it is too expensive, and it costs tens of thousands of dollars to spend a radar.
Google's earliest proposed self-driving car was based on a mechanical rotating lidar product placed on the roof of the Velodyne, which cost more than $70,000 at the time;
The cost of hybrid solid-state lidar is low, and the vibration amplitude of the micromirror is relatively small, but the disadvantage is that the field of view range is too limited, and if a car wants to achieve a better detection effect, it must be equipped with multiple radars, and the cost is still not low;
Pure solid-state lidar scanning speed block, good controllability, smaller size, at present, through the efforts of various manufacturers, the price has been hit down, not as expensive as before.
As for why Musk is reluctant to use lidar, it is likely that he does not want to increase the cost of the vehicle, after all, the huge user system can bring more data to Tesla.
So why are our new car-making forces using lidar? Aren't they afraid of rising costs?
The Weilai ET7, which has just been delivered, and the ideal L9 that have been exposed one after another, are all models with high attention recently.
Weilai ET7 is equipped with a Tudtong Falcon series radar, from the radar parameters, this radar is very strong.
The detection distance is 250m, and the farthest detection distance is 500m: The longer detection distance means that real-time data can be obtained faster, so that the system has more time to react and help the driver make more reasonable decisions.
ROI angular resolution of 0.06°*0.06*: Ultra-high angular resolution ensures the detection progress of small objects, and can detect 20-40cm-sized projectiles 100 meters away, which also leaves the system with increased reaction time.
To put it simply, the LIDAR of THE NIO ET7 can see far and very clearly.
The Ideal L9 is equipped with a Hesai AT128 radar, and the position is the same as the NIO ET7, and both choose the watchtower type. But from the parameter point of view, it is slightly inferior to weilai ET7.
Of course, the parameters belong to the parameters, and how the actual experience depends on the actual measurement of the car; at present, the lidar industry is still in its infancy, and the parameters and experience are not very linked.
As for the demo video that was released, it was a left turn at the intersection of the city section, very confident.
At this stage, whether it is high-speed or garage, it is actually not interesting, and now the true direction of the new forces is actually the urban auxiliary driving.
And we can also see the ideal side and side rear cameras, the outflow of the angle of view overlap is small, without tesla-like double-sided ultra-wide angle lens.
Tesla side lens ▲
Here, by the way, Xiaopeng's lidar processing method: Xiaopeng's lidar use is different from the two mentioned above.
From P5 to G9, Xiaopeng's strategy is to install one on the left and right of the front of the car, abandon part of the field of vision, and do the obstacle avoidance in front of it to the extreme, so Xiaopeng can buy Sagitar M1, which has high-grade parameters and relatively low cost products.
This just happens to illustrate a problem: people are no longer new to car-making who only imitate Tesla.
At present, it seems that Wei Xiaoli's assisted driving solutions and technologies have gradually opened up a gap with other brands, even Tesla in China.
In the past two years, with the start of mass production of lidar products, the reliability brought by the design of lidar redundant sensors, the safety of vehicle driving, and the computing power of the ADAS platform have been greatly improved.
Regardless of whether the regulations allow it, regarding automatic driving, the next year is another fishy (internal) wind (roll) blood (severe) rain (heavy). What are your thoughts on autonomous driving? Welcome to discuss in the comment area~ together