An efficient registration algorithm based on spin image for LiDAR 3D point cloud models

Abstract

Spin image is a good point feature descriptor of the 3D surface and has been used in model registration for many applications from medical image processing to cooperation of multiple robots. However, researches show that current Spin-Image based Registration (SIR) algorithms present disadvantages in computational efficiency and robustness. Thus in this paper, aiming at 3D model acquired from LiDAR sensor, a new SIR algorithm is proposed to solve these problems. The new algorithm is on the basis of a new-constructed three-dimensional feature space, which, composed of the curvature, the Tsallis entropy of spin image, and the reflection intensity of laser sensor, is combined with the concept of KD-tree to firstly realize the primary key point matching, i.e., to find the Corresponding Point Candidate Set (CPCS). After that, spin-image based corresponding point searching is conducted with respect to each CPCS to precisely obtain the final corresponding points. The most absorbing advantages of the proposed method are as the following two aspects: on one hand, due to the introduction of the extra features, the fault corresponding relation introduced by spin image based method can be effectively reduced and thus the registration precision and robustness can be improved greatly; on the other hand, the CPCS obtained using low-dimensional feature space and KD-tree reduces extraordinarily the computational burden due to spin-image based correspondence searching. This greatly improves the computational efficiency of the proposed algorithm. Finally, in order to verify the feasibility and validity of the proposed algorithm, experiments are conducted and the results are analyzed.

Introduction

Registration of 3D model is of great importance in many applications in which the final information is gained from the combination of various data sources, such as, medical image processing [1], remote sensing [2], environment mapping in autonomous robotics [3], [4], and so on.

In general, 3D model registration can be implemented through solving two sub-problems¹: one is feature description and detection, the other is feature matching. The feature often refers to the distinctive and easily detectable local descriptor of 3D models, and it is usually used to denote a piece of information which is relevant for solving the computational task related to a certain applications. Typical feature description and detection methods include “Point Signature”[5], “Principle Curvature”[6] and “Point Feature Histogram”[7], etc. Actually, the feature often determines the basic performance which the whole registration algorithm can finally achieve. While during the procedure of feature matching, the correspondences between the detected features in different models are established, and then transformation map is constructed based on the correspondences if necessary. In order to implementing feature matching, there should be some methods or operators to quantitatively evaluate the difference and similarity of any two features, based on which, searching or optimization algorithms, such as “Iterative closest point”[8], “Chen”[9] and “GA algorithm”[10], are used to find the best “correspondence”.

Most recently, the rapid development of the 3D model registration technique results in its more extensive applications. Nevertheless, applications of the existing 3D model registration algorithms have also encountered some problems, such as algorithm robustness. Here the robustness means that the performance of registration algorithms should be ensured in different conditions (e.g., observing point of view and scale) and insensitive to the uncertain factors (e.g., measurement noises).

As far as the robustness is concerned, spin-image based registration (SIR) algorithms have shown great priority compared to the other existing registration algorithms since it is robust with respect to the data noise, as well as the observing point of view and scale [11]. For example, in reference [12], SIR is used in unstructured 3D point clouds acquired from photos to recognize objects. The experimental results show that it is robust with respect to the scene Gaussian noise. In reference [13], the authors use SIR to register different 3D CT images with large pose differences. Also, the experimental results show the acceptable performance in the aspect of accuracy and robustness. While in reference [14], SIR algorithm is used to test the geometric tolerances in manufacturing processes. Most recently, in reference [15], it is utilized in heavily vegetated environments to realize the relative localization between an air robot and a ground robot.

Spin-image is actually a feature descriptor denoted as a grid based image. Its main idea is to map a 3D model into a 2D plane to describe the local shape of the 3D objects. In order to ensure the robustness, the dimension of the spin image (i.e., the number of the grid) should generally be selected large enough. This, unfortunately, will make the process of constructing a spin-image and similarity evaluating of two spin-images very time consuming (the time complexity linearly relates to the dimension of spin image). Aiming at this problem, there also appear some researches recently. For example, in reference [16], PCA method is used to reduce the dimension of the spin-image feature, so that the total computation requirement can be cut down. In reference [17], Dinh and Kropac propose to use multiple resolution method to reduce the computational burden of spin image based registration algorithm. While in reference [18], SIFT operator is used to reduce the key-point that needs to be considered, and thus the computational burden of the registration can be decreased greatly. Generally, the computational burden of SIR algorithm highly relates to its robustness and the accuracy, and most of the preceding algorithms improve the time efficiency at the cost of robustness and accuracy. That means, problem on how to design new efficient spin-image based registration with ensured robustness and accuracy is still an important and open problem in the related fields.

Thus, in this paper, a new efficient SIR algorithm is proposed. The main ideas of this paper are as follows: Firstly, multiple different feature descriptors (including a simplified spin-image descriptor) are utilized to construct a low-dimensional feature space. In this paper, the curvature, the Tsallis entropy of the spin image feature, and reflection of the LiDAR sensor are considered. Then, the 3D model is resampled and reconstructed using KD-tree structure in the low dimension space. Subsequently, coarse matching is conducted in the low-dimensional space through searching the KD-tree to obtain a so-called candidate correspondence key point set (CCKS). Finally, in the obtained CCKS, fine matching searching is done based on the spin-image feature descriptor. Since the candidate corresponding point set after the coarse matching can be reduced, the computational burden due to spin-image searching can be decreased greatly and the time efficiency of the proposed algorithm can be improved. What׳s more, due to the introduction of the extra features, the fault corresponding relation introduced by spin image based method can also be effectively avoided and thus the registration robustness and precision can be improved obviously.