Shulou(Shulou.com)11/24 Report--

There has been a strange accident in the history of autopilot: driver Joshua Brown turned on Tesla's Autopilot mode on a highway in Florida in 2016 and left his hands off the steering wheel.

At an intersection, a white truck turned left and happened to meet Tesla, who was driven by Brown.

According to witnesses at the time, Brown's Tesla was still driving at full speed and crashed straight into the bottom of the truck when the white truck appeared in front of him. The survey showed that Tesla was going at a speed of 119km / h at that time.

Brown died on the spot.

But Tesla continued to drive and gradually deviated from the road until he hit two fences and a telephone pole.

Wikimedia Commons by Jzh2074 official explanation of the accident is: "neither the autopilot nor the driver noticed that the white side of the trailer set off the bright sky, so there were no brakes." In other words, autopilot recognized the white van in the sun as the sky.

In the five years after that, at least three more Tesla rushed towards the white truck at full speed-it seems that autopilot still can't tell the difference between the white giant and the glowing sky.

The world is always different in the eyes of autopilot.

How does autopilot see the world? The eyes of autopilot are mainly served by sensors such as millimeter wave radar, lidar and cameras.

Millimeter wave radar uses the reflection of electromagnetic waves to measure the distance, velocity and other physical information of objects. It has strong anti-interference ability and long detection distance, so it is suitable to be used in bad weather such as rain, snow, wind and sand. It can also be installed concealed without affecting the overall appearance of the vehicle. But it cannot measure height, and the world is flat in its eyes.

Millimeter wave radar on shopping platform | Taobao screenshot lidar can see the 3D world. Similar to millimeter wave radar, lidar also calculates the distance and velocity of objects by sending pulses and receiving reflections. But the lidar has a shorter wavelength, which means higher resolution and more accurate measurements. Lidar emits thousands of pulses per second. By calculating the reflected light waves at different angles, lidar can restore the three-dimensional image of the target object.

Lidar scan schematic map | Wikimedia Commons but want to know what is in the eye or give it to the camera. With the blessing of the algorithm, the camera can identify whether it is a person, an animal or a building in front of it, but it is still in 2D form, like looking at a picture.

At present, the mainstream recognition scheme is multi-sensor fusion, like some representative models in China, which will have 5 high-precision millimeter wave radars, 13 out-of-car cameras and 1 in-car camera.

Tesla, on the other hand, insists on using only the camera (in the latest update, Tesla even gave up the most basic ultrasonic radar in the car, really realizing pure vision without radar in the whole car), because he was worried that when the radar and the camera perception were inconsistent, the car wouldn't know who to trust. But the more basic reasons are: first, the cost of the camera-only solution is lower; second, the rapid development of the algorithm can indeed make up for the disadvantage of cameras compared to radar.

Tesla displays | unsplash in the car, but no matter which route wins in the end, the sensor can only see. To understand what these "see" means, it depends on the perceptual algorithm. With the information from different sensors, the current algorithm can know where the car is and what the road looks like, understand traffic lights, track obstacles on the road, and read the text on the road signs.

"honest autopilot is better than novice drivers." to a certain extent, this is true.

When a Pikachu pops out in the middle of the road, but when a Pikachu jumps into the middle of the road, the autopilot must be confused.

When a Pikachu jumps into the middle of the road | Giphy is a "rare and sudden" accident, which is called a corner case (corner case, CC) in the self-driving industry. This is a term from system testing, which originally refers to a situation where multiple parameters are in extreme cases. The accident at the beginning of the article is a typical CC.

CC is usually divided into five levels, corresponding to the different complexities of autopilot cognitive accidents: a pixel is not clear; the scene changes from "habitual" to "unaccustomed"; unknown objects appear (but the location and target are clear); location and goal are not clear; until the overall "accident" of the goal, scene, and environment.

But any small CC can have serious consequences.

A lovely yellow-haired electric mouse appears on the road, which is a kind of "unknown object" level CC. The result of failure in recognition is that you either get an electric shock of 100, 000 volts or are hammered by an elf trainer. What could happen in the real world can only get worse. Volvo's Australian technical manager revealed in 2017 that his company's autopilot didn't know kangaroos.

Kangaroo: "you don't know me"? | Giphy "cars use the ground as a reference point to detect animals to determine the distance of objects," the technical manager said. But when a kangaroo is bouncing across the road and jumping in the air, the autopilot will recognize it as an object in the sky and determine that it is much farther away than it really is; and when the kangaroo lands, the autopilot will recognize it closer than it actually is.

Autopilot is difficult to recognize the jumping of kangaroos Giphy We can imagine an Australian coastal road, next to a few happily jumping muscle kangaroos, a self-driving car is at a loss in the face of an unknown object in its eyes, big, small, far and near, neither stop nor walk.

I'm scared for it when I think about it.

An autopilot with a laser pen can be seen in the following picture, the current mainstream sensors all have their shortcomings, which can cause inaccurate judgment of road conditions.

_ identify 3D information

Easy to interfere with

Direct velocity measurement

Detection range

Millimeter wave radar

No

No

Yes

15-250m

Lidar

Yes

Yes

Yes

Less than 300m

Camera

It can be realized with the help of software

Yes

No

Greater than 500m

Under the limitation of the sensor, the algorithm becomes easy to be deceived.

It has been experimented that when an ordinary laser beam is illuminated on a streetcar, the autopilot will regard the tram as a frog; when the laser shines on the sea turtle, the turtle becomes a jellyfish in the eyes of the autopilot; if the laser changes color again, it will also identify a snake as socks, microphones, jackfruit, corn, or hot dog intestines.

Reference [1] has autopilot safety expert analysis, this may be because the algorithm will identify the color of the laser beam as part of the original object, for example, when the wavelength of 400nm laser beam shines on the "hedgehog", the hedgehog's thorn combined with the purple introduced by the laser beam will be finally identified as "prickly thistle" by the algorithm.

Under laser, hedgehog changes into thistle. Reference [1] the success rate of such a confusing algorithm is very high. In indoor and outdoor tests, the success rate can reach 100% and 77.43%, respectively. And this kind of laser pen can be bought for 20 or 30 yuan.

Another important reason for the emergence of CC comes from the inexhaustible number of current image recognition data. The training methods of image recognition are mainly human data stuffing, machine memory and classification. The algorithm, like a child with a good memory, can remember any data and scenes crammed in, but it is not yet a mature and flexible master. A leaf glued to a traffic sign and an insect on a camera may make the algorithm output completely different results.

Why is CC so important? The field of self-driving uses the "law of 28" to describe the importance of solving the CC problem-it takes 80 per cent of the time to solve the seemingly "rare" 20 per cent problem-which they say is the only way to true autopilot.

Accurate identification of CC and correct handling of CC is actually a matter of balancing safety and efficiency (or driving experience). People expect autopilot to be absolutely safe, but no one is willing to accept the self-driving traffic experience of five kilometers per hour and frequent brakes from time to time (and low speed can cause safety problems as well as braking).

No one is willing to accept the autopilot experience with frequent brakes | the danger and popularity barrier of Giphy autopilot comes not only from "yes, but unrecognized", but also from "none, but misjudged". Tesla's "ghost brake" incident is a typical example.

When a Tesla owner was driving in San Francisco, there was a plastic bag floating on the road. "suddenly, the car seemed to be locked," the owner recalled. Its Tesla suddenly slowed down automatically from about 40 kilometers per hour to 24 kilometers per hour, "but it immediately loosened because the plastic bag was removed."

Common plastic bags on the road can also affect autopilot | Wikimedia Commons by Ivan Radic is lucky that at least his car didn't stop suddenly on the San Francisco road. From February to June 2022, the National Highway Traffic Safety Administration received 404 complaints about Tesla's inexplicable sudden braking. Car owners use "ghost brakes" to describe the encounter-Tesla, who turns on autopilot assistance mode, suddenly brakes or slows down when there is no danger, which increases the risk of a rear-end collision.

Phil Kupman, a professor at Carnegie Mellon University in the United States who focuses on the safety direction of self-driving cars, analyzed that this may be due to Tesla's developers failing to correctly set the decision threshold for cars. In particular, Tesla's information recognition almost only depends on the camera, without the assistance of radar and other sensors.

Road traffic is already one of the most complex systems we have created. Giphy the good news is that today, Tesla has solved the problem of smart braking through algorithm optimization. But on our way to true autopilot, there are more strange cases to be solved-road traffic, which is one of the most complex systems we have ever created.

reference

[1] Duan R, Mao X, Qin A K, et al. Adversarial laser beam: Effective physical-world attack to dnns in a blink[C] / / Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021: 16062-16071.

[2] http://jst.tsinghuajournals.com/CN/rhhtml/20180417.htm#outline_anchor_8

[3] https://new.qq.com/rain/a/20220818A00WJL00

[4] https://www.youtube.com/watch?v=PRg5RNU_JLk

[5] https://equalocean.com/analysis/2022091918917 [6] https://thedriven.io/2021/07/13/it-can-see-dogs-the-big-reveal-from-teslas-radar-free-version-of-full-self-driving/

[7] https://www.jianshu.com/p/5ab134804d4c

[8] https://www.theguardian.com/technology/2016/jun/30/tesla-autopilot-death-self-driving-car-elon-musk

[9] https://blog.csdn.net/maopig/article/details/107961922

This article comes from the official account of Wechat: fruit Shell (ID:Guokr42), author: Rui Yue, Editor: sleeping worm, Shen Zhihan

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.