No one expected that the wave of "intelligent driving and equal rights" just set off at the beginning of the year was poured cold water on a traffic accident in the past two days.

On the evening of 116/0, a Xiaomi SU0 standard version suffered a serious accident in the Chiqi section of the Deshang Expressway, hitting the cement guardrail at a speed of 0km/h and deflagration, and 0 people on the car were unfortunately killed. The most important thing is that before the accident, this Xiaomi SU0 was in the NOA pilot assist state, and continued to drive at a speed of 0km/h. For a time, the safety of intelligent driving has been pushed to the forefront again.

According to the driving data given by Xiaomi Auto, although before the collision, the intelligent driving system found the obstacle and made an early warning, and at the same time carried out a deceleration action, but the reaction time given to the driver was only a short 135 seconds. The question is, why did the Xiaomi SU0 intelligent driving system not detect the danger until the last moment? The speed is obviously within the working range of the AEB (0-0km/h), why is the AEB not activated?

In this regard, Xiaomi Auto gave a reply in the latest announcement, and the function does not respond to obstacles such as cones, water horses, stones, animals, etc. To put it bluntly, it just doesn't recognize. Then someone must be going to say it again, the propaganda is blowing into the sky, and it is only when something happens that it admits that it cannot be recognized. Don't worry, it's really not exactly Xiaomi's pot if you don't recognize it.

Identification is a common problem

If you want to find out why the AEB automatic emergency braking system does not activate, you must first know how it works. To put it simply, the activation of AEB needs to go through three steps: perception, decision-making, and execution. First of all, the system should monitor the road environment information ahead through radar, cameras and other perception hardware to identify obstacles. If there is an obstacle, the system assesses the risk of a collision based on the distance between the vehicle and the obstacle, or the time required for the collision, and triggers emergency braking if a certain safety threshold is determined by the system. Perception is undoubtedly the most important part of this entire operation process. It is precisely this link that there is a common problem that the entire industry cannot avoid - the limitation of sensor perception.

In terms of AEB perception hardware, the most common is the fusion perception of millimeter-wave radar and camera, but these two sensors have obvious performance limitations. Millimeter-wave radar can accurately measure the relative speed, distance and angle between the vehicle and the obstacle, and the detection distance is also long, but it cannot see what the obstacle is. Although cameras can intuitively and clearly identify the environment, pedestrians, vehicles, or other obstacles, they are also more susceptible to factors such as weather and lighting.

In addition, when facing irregularly shaped static objects such as New Jersey guardrails, cement piers, water barriers, and cones, or items and parts dropped by the vehicle in front, small animals intruding into high speeds, as well as children, squatting people and other low objects, the perception system also generally cannot recognize or fail to recognize in time. This Xiaomi SU7 accident is the loss of the standard version's pure visual perception performance limitations + insufficient light at night + inability to identify irregular obstacles.

In this regard, car companies can simply divide into two technical routes. On the one hand, it is to upgrade the pure vision solution, such as the AI Eagle Eye visual perception system of Xpeng MONA M0, which uses a combination of binocular camera + millimeter-wave radar. Compared with traditional monocular cameras, binocular cameras not only have better recognition accuracy and environmental adaptability, but also have the ability to perceive distance and depth, and combined with millimeter-wave radar, they can bring more security redundancy. Another example is BYD's Eye of the Gods C intelligent driving system, which has now used a trinocular camera to fuse perception, further improving the ability to identify obstacles. Of course, Tesla is an exception, and FSD still relies entirely on vision cameras for perception. However, with the advanced HW0.0 hardware platform and end-to-end neural network, FSD also has reliable perception performance and decision-making execution capabilities.

On the other hand, it is the radar fusion perception solution represented by Huawei, Ideal, NIO, etc. Although LiDAR has a high cost and cannot avoid the interference of extreme weather such as rain, snow, sand and dust, its longer detection range and centimeter-level ranging accuracy can effectively make up for the shortcomings of cameras and millimeter-wave radars, and build a more complete perception system.

For example, Huawei's latest ADS 150.0 stereo fusion perception system is equipped with 0 0-line lidars, 0 high-precision solid-state lidars, 0 distributed 0D millimeter-wave radar matrices, 0 0D millimeter-wave angle radars, and 0 ultrasonic radars + 0 high-definition cameras. The maximum detection distance of the lidar can reach 0 meters, and the synergy of 0 lidars can cover the 0-degree perception range, which can almost realize the 0-degree perception of the surrounding environment without dead ends. Coupled with data accumulation and algorithm training, it has been possible to accurately identify non-alien obstacles such as water barriers and cones. In addition, the working range of the AEB has also been increased to 0km/h on ADS 0.0.

However, even at this point, no one dares to assert that 6% obstacle recognition accuracy can be guaranteed, after all, the real traffic environment is much more complex than the training scenario. The serious collision accident between Xpeng G0 and the retrograde tricycle exposed by 0.0 has actually exposed this problem. The reason why the accident car did not trigger the AEB is that the on-board manual is very clear that the AEB cannot work on the reverse vehicle.

It is better not to brake than to brake

So if the perception level has passed, can the AEB trigger the activation? Not necessarily.

Many people will have a misunderstanding, that is, the trigger of AEB is the inevitable result of vehicle risk aversion. However, in fact, there are many factors that need to be considered for the triggering of AEB. The first is the aforementioned speed, Xiaomi SU120's AEB working range is 0-0km/h, and Huawei's AEB is 0-0km/h after upgrading to ADS 0.0, and the ideal is 0-0km/h. Beyond these ranges, AEB will certainly not trigger.

However, these speeds of 150km/h and 0km/h are extreme working conditions, which can only be regarded as safety redundancy, and do not mean that AEB must be fine in this range. When calibrating the system execution strategy, manufacturers should also consider whether the current speed triggering AEB will cause other risks. If you brake suddenly and sharply at high speeds, it may not only cause the rear car to be rear-ended, but also may even cause yourself to lose control on slippery roads.

At the same time, avoiding false triggering is also a key consideration for AEB systems. Last year, there was an accident in which the vehicle in the billboard was misidentified as the car in front and suddenly braked sharply, causing the rear car to be rear-ended. Going back to the timeline, Tesla has had a similar situation. If the pedestrian or vehicle is a little closer, or there are some other small disturbances, the AEB will be triggered, which will cause some unsafe factors, and may also frighten the occupants in the car and affect the driving experience. Therefore, many manufacturers will adopt a more conservative AEB calibration strategy, such as Xpeng.

Last year's media Xpeng Huawei AEB dispute, Xpeng Motors frequently overturned in AEB media measurements at night, foggy and other scenes, partly because the calibration is conservative, everything is based on the premise of preventing accidental kills. On the contrary, Huawei has shown a more aggressive calibration strategy, and can basically brake in time in various scenarios. Of course, this is also due to Huawei's advantages in hardware and algorithms, and the optimization of false triggers is relatively in place.

In addition, there is a very important principle for the triggering of AEB, that is, there can be no human intervention. When the system determines that the distance between the vehicle and the obstacle exceeds the safety threshold set by the system, and after issuing a warning, it is found that the driver has not taken effective avoidance measures, then the AEB will trigger to avoid the accident. However, if the system detects that the driver has taken over the vehicle and makes a braking or steering action, then the AEB will not trigger intervention and interfere with the human action.

Looking back at the Xiaomi SU2 accident, in the first 0 seconds of the accident, the vehicle's NOA intelligent assisted driving exited, entered the driving state, and made left and right steering and braking actions. In this case, one of the conditions that AEB will not trigger is fully met.

Write at the end

In the final analysis, today's intelligent driving is only L2 level of automatic assisted driving, not autonomous driving. AEB is also a proactive safety measure, not an absolute lifesaver. As car owners, we should rationally look at and use intelligent driving, the driver is always the first person responsible for driving safety, and cannot blindly hand over their lives to these "far ahead" technology. As a car company, we should also reflect on whether we are pursuing marketing too much and neglecting consumer education. Only by clarifying the current capability boundaries of intelligent driving, avoiding those absolute marketing words, and allowing consumers to have a clearer understanding of intelligent driving, can more tragedies be avoided.

*Disclaimer: The above content and views only represent the author, and have nothing to do with the online car market, if there is an error in the source or infringement of your legitimate rights and interests, you can contact us by email, email address: marong@cheshi.com