Fast 3D Reconstruction of Large Objects with ICP-based Extrinsic Calibration
This paper presents a novel method for fast, scaled, and accurate reconstruction of large objects (such as vehicles) using a fixed array of RGB-D sensors. The process is divided into four stages. First, an extrinsic calibration of the sensors is generated by registering fiducial markers with direct, linear least-squares optimization techniques as used in the Iterative Closest Point algorithm. Next, RGB-XYZ point clouds of an object to be reconstructed are collected from the sensors. A coarse registration of these point clouds is obtained by reprojecting them into a common coordinate frame given by the extrinsic calibration. Third, the coarse registration is refined using the Generalized Iterative Closest Point (GICP) algorithm. Finally, the aligned clouds are consolidated and filtered, yielding a dense point cloud reconstruction. The pipeline is validated quantitatively and qualitatively through simulated and real-world reconstructions of several large objects. Our method can scan a 2.5m x 3m x 2m object in less than 10 seconds with 2% dimensional error on average, enabling an assembly-line style reconstruction of objects.
PA Approval 88ABW-2023-0141