AEB mistook the high-speed billboard vehicle for the real thing, why does the "ghost brake" appear frequently

CBN

2024-05-13 22:07Published on the official account of Shanghai Yicai

Recently, an ideal L9 driving on the highway mistakenly recognized the billboard as a real car and caused a rear-end collision.

An Ideal L9 owner in Xiangyang, Hubei Province, said on social media that he was driving an unlicensed new car Ideal L9 and turned on the assisted driving system on the highway. The vehicle involved was an ideal L9 Pro model. After the accident, the traffic police determined that the owner of the car bears full responsibility for the accident.

Li Auto said that the conclusion of the analysis of the accident is that the visual misdetection of the truck on the billboard in front of it led to the abnormal deceleration of the car, and the assisted driving currently has certain limitations.

AEB misjudged

According to official information, the ideal L9 Pro version of the model is equipped with intelligent driving AD Pro, which can achieve safe and comfortable high-speed NOA (automatic assisted navigation driving), and the model is equipped with 10 cameras, 12 ultrasonic radars and 1 millimeter wave radar. During this year's Beijing Auto Show, the capabilities of the AD Pro and AD Max platforms for Ideal Intelligent Driving have evolved in an all-round way. Among them, AD Pro 3.0 was pushed with OTA 5.2 in early May, realizing high-speed NOA at the level of 1,000 kilometers of takeover, urban LCC (lane centering assist) that supports starting and stopping at traffic light intersections, and intelligent parking in complex parking spaces.

In May last year, a car owner in Ganzi, Sichuan Province drove an ideal L7 on the road normally, and suddenly the speed of the vehicle dropped from 80km/h to 20km/h, and it braked directly. The owner said there were no obstacles on the road at the time, only a billboard with a person ready to start. The owner of the car contacted Li Auto, and the official said that the incident was because the lidar identified the person on the billboard who was about to start as a real person in the middle of the road, so it stopped urgently.

When the driver turns on the driver assistance function, the vehicle suddenly brakes at high speed without any obstacles in front of it or without colliding with the vehicle in front, which is called "ghost braking".

The "ghost brake" phenomenon occurs due to an incorrect recognition of the AEB function.

Ma Jing, a R&D personnel of Huixi Intelligence, said that AEB is a function that many vehicles have been equipped with, through the auxiliary driving system installed on the car - it can be a camera, millimeter wave radar, lidar, or a combination - real-time monitoring of the state of the car and the surrounding traffic environment, and analysis and calculation, to determine the appropriate timing of braking intervention.

The TP (correct triggering) and FP (false triggering) of AEB are a pair of contradictions, and probably speaking, the larger the TP, the greater the FP. In practice, the danger of false triggering is self-evident, and with the popularity of AEB, the shock and injury caused by false triggering have become more and more the focus of complaints.

Ma Jing believes that it is difficult to really do a good job in AEB. "It requires strong perception and low system link latency, and it is not easy to do these two things well. In addition, the AEB functional algorithm itself will also cause vastly different performance. ”

"At present, the mainstream driving assistance system mainly uses cameras and millimeter waves for perception, but the camera recognition accuracy is low, and it is difficult to distinguish objects that are close to the color of the sky; Millimeter-wave radar has higher recognition accuracy, but it is less sensitive to stationary objects and cannot accurately identify objects made of plastic, cloth, and other materials. An industry source said in an interview with the first financial reporter. In addition, assisted driving technology needs to further improve its vision and perception capabilities, and be equipped with multiple safety assurance systems, including sensors, cameras, radars, lidars, etc., to perceive the surrounding environment in real time in order to make correct decisions. In addition, it is necessary to realize online and offline linkage supervision, including monitoring systems, data transmission systems, intelligent control systems, etc., to monitor and feedback vehicle status and driving conditions in real time, including vehicle speed, direction, road conditions, meteorology, etc., and deal with emergencies and failures in a timely manner to ensure safety and stability.

Autonomous driving functions still need to be optimized

Previously, Tesla also broke out a similar accident. In July 2023, Tesla was investigated by the California Attorney General's Office for self-driving safety issues. The cause is that in August 2022, a 2018 Tesla Model 3 owner filed a complaint with the Federal Trade Commission about the "ghost brake" problem with the vehicle. The owner said that he spent thousands of dollars to buy Tesla's FSD software, but in the actual use of it, he felt misled by Tesla. When the autopilot function is turned on, the vehicle will mistake the roadside sign for a speed limit or stop sign during normal driving, and then slam on the brakes, bringing personal danger to the owner and the risk of a rear-end collision.

There is an opinion that Tesla may have adopted a purely visual self-driving solution, which caused the problem. Tesla Motors relies entirely on cameras to collect data, and compared with the hybrid vision solutions of domestic car companies, its data collection may not be comprehensive, especially in the presence of occlusions.

"The image formed by lidar is called a point cloud, which is composed of thousands of points, which cannot see the image itself, but can shape the outline of an obstacle like a blind man touching an elephant. Therefore, no matter what the billboard is drawn, it will not affect the presentation in the system, it is just a square. Liu Gang (pseudonym), an analyst in the automotive industry, told reporters that a few years ago, Tesla also had a situation where it was directly hit by a container truck parked on a white side, and it understood the white car as a white cloud under a clear sky. However, after years of evolution, I believe that Tesla has done targeted key training and has already solved this bug.

It is worth noting that Luminar, a listed company in the United States, recently revealed in its financial report that Tesla became its largest customer in the first quarter of 2024, contributing more than 10% of the company's revenue. According to the data released in the financial report, Tesla purchased more than $2.1 million in lidar.

In terms of intelligent driving technology route, Tesla has always been firm in its vision solution, while most car companies have chosen the technical route of lidar. However, since the beginning of this year, GAC and Huawei have released pure vision solution intelligent driving products. Among them, Huawei announced that it will adopt a pure vision autonomous driving solution for the first time on the Zhijie S7 model.

CITIC Securities pointed out in a research report that the biggest advantage of the pure vision solution is the overall cost, the hardware cost of Tesla's eight cameras is only about 200 US dollars, and the cost of a set of lidar ranges from 3000 ~ 10000 US dollars. For multi-perceptron fusion schemes, lidar is indispensable. Although the cost of lidar has been decreasing in recent years, a single one is still around 600~2000 US dollars, and the whole automatic driving system generally needs 3~5 lidars, so the total cost is also 3000~10000 US dollars. Although the cost of pure vision solutions is lower, it relies more on advanced algorithms.

"Corner cases (anomalies that rarely occur in intelligent driving but can lead to danger) can be said to be endless, so Chinese car companies can only better solve this problem through big data training and improved algorithms. I also think this shows the indispensability of lidar. Pure vision is bound to have defects, and I believe that autonomous driving done by pure vision raises the lower limit of autonomous driving, not the upper limit of capabilities. In other words, it can be free in everyday scenes, but it will still make mistakes if it encounters scenes that it has not seen and understood incorrectly. Liu Gang told reporters.

It is worth noting that, in addition to the "ghost brake", which can easily lead to danger, there are many problems with misjudgment of assisted driving in recent years. Previously, a Xpeng Motors owner was mistakenly judged by the system to drive to sleep because of his small eyes; Auto KOL Chang Yan also said that because of his small eyes, he has triggered real-time warnings from many car brands' assisted driving systems. Mr. Tsunei said that GM's Super Cruise would judge him to be overtired; VOYAH FREE will turn on the cold wind in the winter to keep him from sleepy; The NIO ET7 determined that he was tired and distracted.

The reason for this miscalculation is the vehicle's Driver Monitor System (DMS). While technology has come a long way, no company has yet been able to achieve full autonomous driving, and drivers still need to monitor the vehicle's operating status, even with advanced driver assistance features enabled. However, a significant number of users, either misled by the hype or driven by curiosity, rush to try the advanced driver assistance system as autonomous driving. In the past two years, there have been many news reports of accidents due to the irrational use of driver assistance systems by users.

(This article is from Yicai)

View original image 85K

• 

•