Deciphering the Uber Autopilot System: Why does multi-sensor protection still kill people?

According to foreign media reports, in the early morning of March 19, US local time, Uber, a self-driving car, killed a woman crossing the road in Tempe, Arizona. This is also the first accident in the world that caused a self-driving car to crash on public roads. Therefore, this has also triggered widespread concern about this autonomous car accident. Many people have questioned the safety of self-driving cars.

Currently, Uber has stopped all autopilots in all of its ongoing road test cities, including Pittsburgh, Toronto, San Francisco and Phoenix. At the same time, Uber issued a statement saying that they will make up for the victim's family to the greatest extent and fully cooperate with the local authorities' investigation.

The responsibility for the accident is not in Uber?

At a time when outsiders accused Uber and autonomous driving technology, the director of the Tempe Police Department in Lisanna said that according to preliminary investigations of the accident, Uber may not be at fault in the traffic accident.

From the video taken by the camera equipped with Uber's self-driving car, the way the woman crossed the road when the accident occurred, regardless of whether Uber's self-driving car was in manned or autonomous driving mode, it was difficult to avoid the accident. occur.

The driver on the front seat of Uber's self-driving test car also said that because the victim suddenly appeared in front of them, the reaction was not over. The first traffic accident alert he received was the sound of the collision between the two.

At the same time, Mark Rosenk, former chairman of the US National Transportation Safety Board, also said that the public should not overreact to Uber autopilot car deaths. He cautioned that the United States had an average of more than 6 million traffic accidents per year, killing 6,000 pedestrians and nearly 40,000 people. “But this accident will still be a hindrance, and we must manage to regain public trust in self-driving cars,” he said.

But it is worth emphasizing that the speed limit for the accident was 35 miles per hour (about 56 kilometers), but the Uber was at a speed of 38 miles per hour (about 61 kilometers). Obviously, speeding may also be one of the causes of this accident.

Secret Uber Autopilot System

In fact, people or things that accidentally enter the driving path are almost the first thing that autonomous car engineers consider. There are many such situations, such as a car that suddenly stops in front, a deer, and a pedestrian passing by. Engineers will find them as early as possible when designing an automated driving system, confirm and take appropriate measures, and may slow down, stop, and return.

In general, autonomous vehicles are equipped with multiple safety systems. Even if one system fails, another system should work. Unless all systems fail, it is generally difficult to have such a fatal accident (except, of course, in unexpected situations, such as pedestrians who normally walk on the sidewalk suddenly rush into autonomous vehicles, whether human Drivers driving cars or self-driving cars simply do not have enough time to take measures to avoid accidents.

Uber self-driving car is equipped with a complete set of sensors

Uber self-driving cars are equipped with a number of different imaging systems that perform common tasks (monitoring nearby cars, markings and lane lines) as well as special tasks such as the above-mentioned measures to slow down the car. In this accident, the image system capable of saving the victim should be no less than four.

Top lidar

This barrel-like object at the top of the car is a lidar system that generates 3D imaging around the car at multiple times per second. Lidar uses an infrared laser pulse that bounces back when it encounters an object and returns to the sensor to detect stationary and moving objects in the day or night.

Heavy snow and thick fog can blur the laser of the lidar, and the accuracy will decrease as the distance becomes longer. However, within a few feet to a few hundred feet, the top lidar is still a very important imaging tool, almost exist in Every self-driving car.

Laser generated image

If the top lidar unit is functioning properly, it should be able to distinguish the victim without being completely occluded. At the same time, the car is more than 100 feet away (more than 30 meters) from pedestrians, and can pass pedestrian imaging to the "brain" for verification.

Front-end radar

Like a lidar, the radar sends a signal waiting for it to bounce back, but uses radio waves instead of lasers. This makes the radar more immune to interference, because the radio can penetrate snow and fog, but it also reduces the resolution and changes its distance image.

Tesla's autopilot system Autopilot relies mainly on radar

Depending on the number of radars deployed by Uber, the range of distances it provides may vary greatly. Uber is likely to deploy multiple radars in front of and behind the car to achieve 360-degree coverage. If Uber is to supplement the lidar, there may be significant overlap between radar and lidar imaging, but more radar can be used to detect other vehicles and larger obstacles.

Short and telephoto optical camera

Lidars and radars are easy to determine the shape of an object, but are not good at reading marks, understanding the color of objects, and so on. This task is handed over to the visible light camera, which incorporates sophisticated computer vision algorithms that enable real-time analysis of the image.

These cameras on Uber's self-driving cars pay close attention to the indicators that mark the car's brakes (suddenly red lights), traffic lights, and pedestrians crossing the intersection. In particular, the front end of the car uses multi-angle, multiple types of cameras to get a complete picture of the car's way forward.

Detecting pedestrians is one of the most common problems engineers are trying to solve. This algorithm is very mature. This process is often referred to as "segmentation" of images and generally involves marking, trees, human actions, and the like.

Image taken with an optical camera

For the camera, working at night is difficult, which is an obvious problem. However, this problem is left to the first two systems to solve, Lidar and radar can work at night. Even in an environment where you can't reach your fingers, people wearing black clothes can be identified by lidar and radar. These radars then warn the car that it should be slowed down or confirmed by the headlights of the car. This may be why autonomous vehicles usually do not have a night vision system.

Safe driver

Humans are very good at finding obstacles, even if there is no laser in the human eye. Human reaction time is not the best, but the car will not respond or make a wrong response, and the trained safety driver will respond correctly.

It is worth noting that there is also a central computing unit in the self-driving car that collects the information provided by each sensor and then produces a more complete picture of the surrounding environment of the car.

Autonomous vehicles in road test are generally equipped with safety drivers.

summary:

Although the accident is still under investigation, regardless of the final result, it will have a huge impact on the future of autonomous driving tests and the use of autonomous vehicles on the road. Perhaps, due to the accident, unmanned road test supervision around the world will be further strengthened.