Utilization of Multisensor in the Initial Registration of Multiview Point Clouds

Abstract

The three-dimensional(3D) scanning is one of the basic technologies to support the autonomous vehicle, the virtual, augmented, mixed reality as well as the shipbuilding or the construction industry. That's why it is important to do research on processing the three-dimensional point clouds obtained by scanning the objects. In the data processing, it is necessary to merge several point clouds taken by laser scanner at the different positions into one huge point cloud. It is called the registration, which the multiple point clouds in each local coordinate are combined to become a big one in the global coordinate.

The registration consists of the initial registration and the fine registration. The former one is the task of aligning the point clouds approximatively using the position and the orientation of the scanner. The latter one is that of searching the corresponding points in one point cloud to match with the points in another point cloud to move them as closely as possible. Meanwhile, the types of registration are pairwise and multiview. The pairwise registration is to carry out the registration process between two point clouds, and the multiview one is to do it among more than three ones.

In this thesis, the whole framework for the registration of the point clouds obtained by scanning a building without marker is proposed, which the building has the repeated pattern such as windows or walls. This framework includes from the fabrication of a mobile laser scanning system to the multiview registration. The overall process is as follows. First, a 3D laser scanner module is fabricated by connecting a 2D laser range finder and a servomotor. Also, a position tracking module is integrated with GPS (Global Positioning System) and IMU (Inertial Measurement Unit) sensors. Second, the 3D scanning is performed using the mobile laser scanning module, and at the same time, the position and orientation data from the position tracking module are obtained. Third, the scanned point clouds are aligned roughly by using the sensor data (MSIR). Fourth, the fine registration of two adjacent point clouds is performed sequentially (SPFR). Fifth, the multiview registration (MVR) is done by considering not only adjacent point clouds but also not-adjacent ones with the overlap region to each other to reduce the global error (GE).

Two registration algorithms are utilized to complete the above registration process. K. Pulli's pairwise registration algorithm is utilized for the sequential pairwise fine registration (SPFR), and the optimal estimation algorithm (OEA) of D. Borrmann's is utilized for the multiview registration (MVR). Furthermore, by substituting the point-to-plane (P2L) method for the point-to-point (P2P) method in the step of the computation point correspondences in OEA, the results from these methods are compared.

One of the contributions of this thesis is to carry out the entire registration process, which is including from the fabrication of the 3D scanner to the multiview registration, by using the own mobile laser scanning (MLS) system and position tracking (PT) system without any noticeable marker. The other is to change the P2P method (P2P-OEA) into the P2L method (P2L-OEA) in the original OEA for the multiview registration. However, there are limitations. One is that the constraint parameter values for the registration should be adjusted heuristically through a number of trial and error in the pairwise fine registration (PFR) because each point cloud has different characteristics. The other limitation is that the users cannot check whether the results of 3D scanning and registration are good or not until they come out in the laboratory. Therefore, in the near future, the proposed framework will be integrated on a platform so that the whole process of the registration will be done right after the 3D scanning and its result can be examined on-site.