Scalable 3D perception: from environmental reconstruction to workpiece measurement

Abstract

Three-dimensional geometric perception serves as a primary means for understanding the environment and objects, essential for both industrial and everyday applications. Advances in sensor technology and algorithms have expanded the scope of 3D perception, enabling its use across various fields such as geological surveying, robotic navigation, and industrial manufacturing. However, the practical application of raw point clouds, which consist of discrete 3D points, faces challenges related to robustness and accuracy, especially when derived from diverse sensors with differing characteristics. This paper introduces the "Point Cloud +" approach, incorporating multi-modal sensor fusion, deep feature extraction, active viewpoint planning, and motion priors. The approach enhances raw point clouds with additional features and priors, supporting robust and precise perception for both large-scale environmental reconstruction and workpiece measurement. Key contributions include a geometric-semantic joint mapping framework for outdoor environments, a degeneration-aware place recognition method, and a planning-control-reconstruction system for accurate workpiece measurement.