You would never have guessed that a roadside car ad almost caused a major car accident!
Recently, Ms. Li, a car owner in Xiangyang, Hubei Province, shared a dashcam video. The video shows that the ideal L9 Pro version of the new car purchased by her family suddenly and urgently braked on the highway, causing the rear car to avoid and collide with it. At first glance, this may seem like an ordinary rear-end collision, but it exposes a glaring loophole in autonomous driving technology.
The accident originated because Ms. Li's husband was using the All-Scene Intelligent Driving (NOA) assistance system at the time. Her husband just sat relaxed in the driver's seat, leaving full control of the car's autopilot system. Just when they were driving on the Xiaogan Expressway, the system suddenly detected that there seemed to be a car across the road in front of them, and immediately braked wildly, but the car behind had no time to react and collided with them.
After analysis, the source of this dangerous "fake car accident" turned out to be a car billboard next to the highway! The Ideal L9, which is equipped with 10 high-definition intelligent driving cameras, actually recognizes the vehicle image on the print advertisement as a real car, so the system judges that it needs to brake immediately to avoid it.
Many people are surprised by the "myopia" of autonomous driving technology. Even a 6-year-old wouldn't think of a flat image as a three-dimensional physical object, right? Even AI systems, once considered "incredible", are now able to tell the difference between the virtual world and the real world. But why does AI driving technology repeatedly go wrong on this "pediatric" problem?
In this regard, Li Auto's official response seems "helpless": at present, the assisted driving function is still at the L2 level, which cannot completely replace the driver's judgment of the surrounding environment, so manual intervention is still required.
This is not the first time Li Auto has been in danger due to misidentification. Last year, their customer service explained a similar miscalculation when the lidar mistook the "jump-ready portrait" on the billboard for a real person, so it braked sharply to avoid it.
Level 2 driver assistance systems do have the limitation of not being able to focus 100% on complex and changing road conditions. But it is also questionable why autonomous driving technology, which is a representative application field of artificial intelligence technology, performs so poorly in basic object recognition capabilities. To make the system distinguish the difference between a flat and a three-dimensional object is not a great "black technology".
Analysts believe that the reason for this low-level mistake is likely to be because the autonomous driving system simply pursues the strategy of "preferring to kill by mistake rather than release". They consciously amplify identification errors to minimize false negatives. Although this can ensure driving safety to the greatest extent, it also increases the "accidental injury" rate.
Another potential reason is that autonomous driving systems rely more on detection equipment such as lidar and millimeter-wave radar when analyzing the environment, rather than relying less on visual imaging. However, sensors such as radar often only recognize the contours of objects, and cannot clearly distinguish between images and real objects.
Industry insiders tell us that it will not be easy to solve this problem completely. At the same time, it is necessary to improve the accuracy of the sensor, optimize the recognition algorithm, and increase the training samples, so that the system can "grow a brain" and learn to distinguish between virtual and real. However, in the short term, automakers may first strengthen the human-computer interaction part, reminding drivers to pay more attention to the appearance of "fake vehicles", so as to avoid the out-of-control autonomous driving system from "stabbing" the innocent again.
In addition to technical reasons, it is also worth reflecting on the use of exaggerated techniques and realistic effects by advertisers. If everyone is creating a "fake big empty" business environment, it will inevitably interfere with the judgment of the machine vision system.
Although it is ridiculous that self-driving cars mistake advertisements for real cars, they also reflect the many difficulties in the application of artificial intelligence technology. Are we ready to fully embrace this "imperfect" new technology that is rapidly evolving? Perhaps everyone should remain rational and sober.