Registration of Point Clouds based on Global Super-Point Features using Auto-Encoder Deep Neural Network

Registration of scanned point clouds is the process of integrating two separate local point clouds into one global coordinate system. This process is a key stage in robotic vision SLAM[1], [2], 3D scan to model matching[3] and precision navigation with noisy GPS input. New data acquisition technology such as LIDAR laser scanners mounted on vehicle or aircraft enables for the capture of high quality 3D point clouds over very large areas, as can be seen in Figure 1. In this research we propose a new algorithm for solving the registration problem between a small local point cloud and a large global point cloud. By solving this problem, the exact location and orientation is found, from where the local point cloud was scanned. This effectively acts as a highresolution localization method, independent of GPS. In the proposed approach, the registration transformation matrix is computed through the following algorithm: super-point division of the point cloud, super-point normalization, depth map projection, global descriptor extraction, dimension reduction, saliency detection and feature point matching. The super-points are computed by applying over-segmentation method VCCS[4] on the 3D point cloud. The dimension reduction is computed on a Deep Auto-Encoder neural network (DAE) [5]. The feasibility of this approach is demonstrated on large outdoor 3D point clouds. Figure 1: Demonstration of aerial point cloud collection using LIDAR technology [9]. INTRODUCTION There are numerous 3D registration methods[6], which perform well on small point clouds with large overlapping areas. One scenario being focused on, is the registration between a large global point cloud of an entire city or area, and a small local point cloud. This data can reach ~100 million 3D points and has small overlapping area. A new approach is needed to deal with this problem. The classic pipeline of registration algorithm for 3D point clouds is based on methods developed originally for image registration[7]. Illustrated in Figure 2 a common 3D point cloud registration: Figure 2: Standard registration of point clouds The standard approach today has major limitations that can be eliminated with the proposed new algorithmic pipeline. APPROACH In this research, our solution for registration of small to large point clouds builds upon state-ofthe-art super-point and deep learning technologies. It includes the following steps: a. Over-segmentation of a point cloud which labels semantically uniform segments within the point cloud, called super-point cloud. The VCCS[4] algorithm has been applied here. b. Normalization of super-point cloud location and direction. In addition the projected convex hull shape is computed, in order to create Proceedings of the 34 Israeli Conference of Mechanical Engineering Faculty of Mechanical Engineering, Technion I.I.T Haifa, 21-22 November 2016 baseline for geometric comparison. These computations include: a) Cropping a sphere from the center of the super-point cloud; b) Finding the main directions of the cloud covariance; and c) aligning the cloud to a local coordinate system. c. Global descriptor creation is used for depth map projection of the super-point cloud on hyperplane with minimal data variance. Using DAE, it is important to keep the reduced data as close to the original point cloud as possible. d. Dimension reduction is based on a Deep Auto-Encoder Neural Network [5]. The global descriptor has been reduced in a compact, unique and Euclidean space comparable form. e. Saliency detection is applied to filter out the common descriptors that may lead to false connections. They are filtered by thresholding the amount of eigenvectors that are needed to closely recreate the descriptor. f. Match proposal creation [8] of super-point descriptors is applied between point clouds. By finding matching super-points in both clouds, potential connections are found. Then the RANSAC algorithm finds the optimal transformation while disregarding outliers. The result of this method is a 3D rigid transformation matrix. Figure 3: The proposed registration of 3D point clouds based on super-points, global descriptors and deep auto-encoders. Currently the algorithm is in progress. Preliminary results are demonstrated in Figure 4. ACKNOWLEDGEMENTS This work was supported by Magnet Omek Consortium, Ministry of Industry and Trade, Israel. Figure 4: Registration demonstration of small to large super-point labeled aerial scan of urban environment (Data from Elbit Systems Ltd [10]). CONCLUSIONS In order to register a small local point cloud to a very large global point cloud a new approach is required. The classic method of matching local key point features in large point clouds is like hoping to find a pair of matching needles in two large haystacks. Instead we take a new macro approach, which compresses the raw point cloud data into semantically uniform super-points. By describing the local geometry of each super-point cloud with a unique and compact descriptor, matches can be found in a much smaller pool of options. This is still in progress but our preliminary results show strong potential for this research.