Carnegie Mellon University (CMU) researchers have developed a method to help self-driving vehicles enhance their object-detection accuracy by enabling them to recognize empty space.
CMU's Peiyun Hu said autonomous cars usually reason about surrounding objects by using three-dimensional (3D) LiDAR data to represent objects as a point cloud, and then trying to match those point clouds to a library of 3D object representations.
However, sensors cannot perceive occluded parts of an object, and current algorithms do not reason about such occlusions.
Hu and colleagues used map-making techniques to help the perception system reason about visibility when attempting to recognize objects—and it outperformed a standard benchmark, improving detection by 10.7% for cars, 5.3% for pedestrians, 7.4% for trucks, 18.4% for buses, and 16.7% for trailers.
Said Hu, “Perception systems need to know their unknowns.”
From Carnegie Mellon University School of Computer Science
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA