China and Africa News, June 17th, being able to quickly detect other vehicles or pedestrians on the road is essential for autonomous vehicles. Researchers at Carnegie Mellon University have shown that they can help vehicles identify Invisible objects to significantly improve the “detection accuracy”.

Obviously, objects in the human line of sight can obscure or obscure objects further ahead. But Peiyun Hu, a PhD from the Institute of Robotics at Carnegie Mellon University, said that self-driving cars usually don’t judge surrounding objects in this way.

Source: Carnegie Mellon University

Unlike a person, the presence of objects around the self-driving car allows him to work more accurately. These tools use data from LIDAR sensors to define objects as “clouds” and then try to match these clouds with objects in the 3D database, but the part that Hu needs to pay attention to is that the 3D data from the sensors may not be true 3D. According to the doctoral student’s description, the vehicle’s sensors may not be able to see the facade outside of the object’s field of view, and existing algorithms cannot make inferences in this situation.

Hu’s research enables self-driving car perception systems to consider visibility when inferring what the sensors are seeing. In fact, reasoning about visibility has been used by some companies to build digital maps.

Carnegie Mellon University Robotics Associate Professor Deva Ramanan explained, “Map construction fundamentally explains which places are empty and which places are occupied. But this is not commonly used when dealing with obstacles moving at traffic speed in real time. .”

At the Computer Vision and Pattern Recognition (CVPR) conference, Hu and his colleagues borrowed map-making techniques to help the system infer the visibility of objects when recognizing objects. In standard benchmark tests, their method performed better than the previous technology. The detection of cars increased by 10.7%, the detection of pedestrians increased by 5.3%, the detection of trucks increased by 7.4%, and the detection of buses increased by 7.4%. Increased by 18.4%, and the detection of trailers increased by 16.7%.

The previous system did not consider visibility, and one of the reasons may be the calculation time. But Hu said that his team’s method only requires 24 milliseconds, while the lidar scan time is 100 milliseconds.