Recently, "car owners say that Xiaopeng Automobile's automatic assisted driving failure" has become a hot topic for a while.
Mr. Deng, the owner of Xiaopeng Automobile, said that after driving more than ten kilometers on the national highway, he suddenly encountered a car that overturned on the road, without any alarm and deceleration, and the car crashed straight into it. Tragedy.
It was another rollover stationary vehicle. Coincidentally, Tesla and Nio have also encountered such a situation.
At the end of January 2021, when the NIO ES8 turned on the L2 level automatic driving function (NOP, Pilot Assist) on the highway, it crashed into a stationary Wuling Hongguang.
In early June 2020, Tesla's Model 3 hit a overturned truck while turning on L2 level autonomous driving (AP, Pilot Assist) on the highway.
Why do things like this keep happening? Why do car companies see that friendly businessmen have such problems and do not do it to avoid it?
Before answering this question, let's first look at how the car works when the L2 level of assisted driving system is turned on.
In the state of opening the L2 level assisted driving system, the car can autonomously complete the operations such as lane change overtaking, automatic driving in, and driving out of the ramp under the driver's monitoring. It's like a person driving a car, and the car is in this state of perception (collecting road information), making decisions (knowing how to drive), and executing (executing a well-planned strategy).
The inability to avoid stationary vehicles is clearly a perceptual problem.
Let's take a look at the assisted driving perception solution adopted by these three car companies!
The NIO ES8 is equipped with 25 sensors, a forward 3-eye camera, 5 millimeter-wave radars, 4 surround-view cameras and 12 ultrasonic radars, and a driver monitoring camera.
The Tesla Model Y is surrounded by 8 cameras, 1 millimeter-wave radar, and 12 ultrasonic radars.
The P7XPILOT 2.5 system is equipped with 22 sensors, including a forward monocular camera, four parking 360-degree surround view cameras, 1 in-car face recognition camera, 1 dashcam camera, 3 millimeter wave radars, and 12 ultrasonic radars.
Although Tesla and Nio are three forward-looking cameras, they do not use stereo vision, and the three cameras are mainly different focal lengths and different viewing ranges. So in general, all three of them use the sensor solution of vision + millimeter wave radar.
Don't look at so many smart sensors, but when turning on L2 autonomous driving systems (Autopilot, NOA or EAP systems), vehicles rely mainly on forward-looking cameras and millimeter-wave radar to detect objects in front of them.
This architecture, in other words, is that vision is the primary sensor, and then supplemented by millimeter-wave radar. Among them, the weight of the vision sensor is high, and the weight of the millimeter-wave radar is low.
Sensor solutions using vision + millimeter-wave radar, as long as it is not binocular stereo vision, whether it is using rules-based vision algorithms or using deep learning technology, has a natural and unchangeable defect, which is manifested as the inability to identify some targets.
Whether it is a rules-based visual algorithm or the use of deep learning technology, there must be a big premise, that is, there is corresponding data, in order to recognize, can be recognized.
For example, if you haven't seen a scenario, the training dataset can't completely cover all the targets in the real world, and it's already very good to cover 10%, so the remaining 90% can't be recognized without seeing, not to mention that the real world is generating new irregular targets every moment. For example, a bad car on the road.
This has been the case with Tesla in many accidents, such as two rear-end sweepers on the highway in China (the first fatal) and several rear-end fire trucks in the United States.
The second is that the image lacks texture characteristics, like a blank piece of paper in front of the camera, which naturally cannot recognize what object it is.
Some large trucks with high chassis are like white paper at a certain moment or a white wall, and the machine vision based on deep learning is like a blind man at this time, and it will directly crash into it without slowing down.
In order to make up for this mistake, the detection results of millimeter-wave radar are introduced on the basis of vision for verification.
But millimeter-wave radar has a flaw when detecting stationary target information, it can detect static targets including buildings, vehicles, pedestrians, etc. But it cannot be well distinguished and recognized. It takes a certain algorithm to distinguish the target from it.
For example, if you are a moving car, millimeter-wave radar can detect the target relatively well. If the truck is stationary or moving slowly, millimeter-wave radar needs algorithms to detect if there is an object in front of it.
Then there is the millimeter-wave radar, the current production car millimeter-wave radar angle resolution is too low, the millimeter-wave radar installation angle is also very low, and it is too sensitive to metal objects.
In order to avoid misoperation, radar companies and some car companies will directly filter out some stationary objects or suspected stationary objects through algorithms after getting the reflection data of radar to avoid incorrect reactions.
All millimeter-wave radars filter out static targets or encounter large trucks with taller chassis, which may also be undetectable.
This is the drawback of using cameras and millimeter-wave radar schemes at this stage. Visually seen can not be recognized, millimeter wave radar also detected, but also because it can not distinguish and distinguish, it was filtered out.
The result is that the target is not recognized and the target is not present, no different from a blind person - the vehicle will think that there is no obstacle in front of it, and then directly hit it without slowing down.
In view of the defects of this technology, some automakers have developed new technical solutions, such as the three-dimensional binocular solutions of Mercedes-Benz and Toyota, and there are also lidar fusion schemes promoted by new automakers!
The solution of lidar and stereo binocular sensors can only solve some problems, but there is no large-scale application now, so the individual cases of problems have not been exposed.
So why install such a flawed technology on the car?
In fact, this is not a defect, but everyone's understanding of the ability boundary of this technology is not clear enough.
This has to mention the classification of automatic driving.
At present, the "driverless" and "automatic driving" promoted by car companies are actually L2-level automatic assisted driving.
According to the CLASSIFICATION OF SAE (Institute of Automotive Engineers), autonomous driving can be divided into six stages of L0-L5.
Among them, L0 represents no automatic assistance function; L1-L2 is a driver assistance system, and the driver is the main body of vehicle operation; L3-L4 is a conditional automatic driving system; only L5 is truly unmanned.
The L2 level can achieve the function of independently completing driving skills such as lighting lane change, overtaking, following the car, and getting up and down ramps on the high speed, but the driver needs to concentrate at all times and be ready to take over.
L2 and below is the primary level of autonomous driving, and the autonomous driving technology in this range does not have "full automatic driving capability", but is a driving assistance system that always requires human supervision - the main person responsible for the accident is always the human driver.
For a car user, the concept of "assisted" driving and "automatic" driving should be the most basic cognitive common sense. Sadly, players in the self-driving industry have long exaggerated publicity and fooled users.
Tesla promotional page
WM W6 promotional page
This publicity situation is after last year's "WEILAI ES8 owner activated the NOP pilot function, a traffic accident occurred in the Hanjiang section of the Shenhai Expressway, and unfortunately passed away." It was only after the incident that things changed.
However, consumers who have become accustomed to and rely on this system are not clear about what kind of system they are opening up and whether there are risks in this system.
Many times, they judge by name. Just like the "automatic assisted driving" that the owner of the Xiaopeng car said, it may be that for the owner of the Xiaopeng car, he does not know what can be done and what cannot be done in the so-called automatic assisted driving mode.
This creates errors, and some people trust the automatic driver assistance system too much, thinking that they can completely free their hands and feet when driving. But encountering some unrecognizable situations can lead to tragedy.
Finally, it is recommended that when you use such functions, read more user manuals, watch the explanatory videos provided by car companies, understand where the boundaries of automatic assisted driving capabilities are, and be responsible for your own safety!
"The Heart of the Car · The connoisseur said" trailer
On Thursday, April 28, 20:00-21:00, Si Zhengmin, Product Director of Zhixing Technology, will be a guest of the heart of the car · Experts say, let's talk about the "evolution of domain controllers."