Researchers have developed an automated system that allows cars to perceive around corners and discover oncoming traffic and pedestrians.
The system can be easily integrated into today's vehicles, using Doppler radar to reflect radio waves from surfaces such as buildings and parked cars. The radar signal hits the surface at a certain angle, so its reflection bounces back like a cue ball hitting the wall of the pool table, and the signal continues to hit objects hidden in the corner. Some radar signals are reflected back to the detector mounted on the car, so that the system can see the objects around the corner and judge whether they are moving or stationary. Felix Heide, an assistant professor of computer science at Princeton University, said: "This will enable cars to see obscured objects that cannot be recorded by today's lidar and camera sensors, for example, to allow autonomous vehicles to see danger. The intersection." The researchers also said, "The cost of radar sensors is also relatively low, especially compared to lidar sensors, and can be mass-produced."
In a paper published at the CVPR2020 conference on June 16, the researchers described how the system distinguishes objects including cars, bicycles, and pedestrians, and measures its direction and oncoming speed.
The author writes: "The proposed method allows pedestrians and cyclists to provide collision warning in real-world autonomous driving scenarios, and then use existing line-of-sight sensors to see pedestrians and cyclists."
In recent years, engineers have developed various sensor systems to enable cars to detect other objects on the road. Many of them rely on lidar or cameras that use visible or near-infrared light, and such collision-preventing sensors are now common in automobiles. However, it is difficult to use optical sensing to identify objects out of sight of the car.
In an earlier study, Hyde and his colleagues wondered whether it was possible to create a system that uses imaging radar instead of visible light to detect hazards beyond the car's line of sight. For radar systems, the signal loss of smooth surfaces is much lower, and radar is an effective technology for tracking objects. The challenge is that the spatial resolution of the radar (used to photograph objects around corners such as cars and bicycles) is relatively low. However, the researchers believe that they can create algorithms to interpret radar data to make the sensors work properly.
To enable the system to distinguish objects, the team processed part of the radar signal, that is, the standard radar considered background noise rather than available information. The team applied artificial intelligence technology to improve the processing and read the images. Fan Fanwei, a graduate student in computer science and one of the main authors of the paper, said that the computer running the system must learn to identify cyclists and pedestrians from very sparse data.
"First of all, we have to detect whether there is something. If something is important? Is it a cyclist or a pedestrian?" she said. "Then we must find it."
She said that the system is currently able to detect pedestrians and cyclists because engineers believe that these objects are small in size and have different shapes and movements, so they are the most challenging objects. She said the system can also be adjusted to detect vehicles.
The researchers plan to further improve on radar and signal processing, so that the technology can be applied in multiple directions. This system has the potential to fundamentally improve vehicle safety. Since it relies on existing radar sensor technology, it should be possible to prepare for the deployment of radar systems in next-generation vehicles.
