Towards 3D lidar point cloud registration improvement using optimal neighborhood knowledge

https://doi.org/10.1016/j.isprsjprs.2013.02.019Get rights and content

Abstract

Automatic 3D point cloud registration is a main issue in computer vision and remote sensing. One of the most commonly adopted solution is the well-known Iterative Closest Point (ICP) algorithm. This standard approach performs a fine registration of two overlapping point clouds by iteratively estimating the transformation parameters, assuming good a priori alignment is provided. A large body of literature has proposed many variations in order to improve each step of the process (namely selecting, matching, rejecting, weighting and minimizing). The aim of this paper is to demonstrate how the knowledge of the shape that best fits the local geometry of each 3D point neighborhood can improve the speed and the accuracy of each of these steps. First we present the geometrical features that form the basis of this work. These low-level attributes indeed describe the neighborhood shape around each 3D point. They allow to retrieve the optimal size to analyze the neighborhoods at various scales as well as the privileged local dimension (linear, planar, or volumetric). Several variations of each step of the ICP process are then proposed and analyzed by introducing these features. Such variants are compared on real datasets with the original algorithm in order to retrieve the most efficient algorithm for the whole process. Therefore, the method is successfully applied to various 3D lidar point clouds from airborne, terrestrial, and mobile mapping systems. Improvement for two ICP steps has been noted, and we conclude that our features may not be relevant for very dissimilar object samplings.

Introduction

Lidar systems provide 3D point clouds with increasing accuracy and reliability. When the same area of interest is acquired twice or more, over time or from several points of view, depending of the application, the registration problem arises. Using for instance a hybrid IMU (Inertial Measurement Unit)/GPS georeferencing system on airborne and mobile platforms introduces 3D shifts between distinct point clouds. For airborne surveys, this is mainly due to drifts of the IMU. Therefore, the so-called strip registration issue has to be overcome: planimetric and altimetric discrepancies exist between two overlapping strips at a level higher than the sensor noise (up to 2 m in altimetry). When dealing with mobile mapping systems in urban corridors, the problem is increased by GPS signal gaps, resulting in significant shifts in platform trajectory estimation, as well as by the presence of moving objects such as cars and pedestrians. Finally, registration is also required in static terrestrial devices, when several points of view of the same object are acquired, facing the issue of putting them in correspondence with few overlapping areas and varying point densities (Lichti and Skaloud, 2010). Consequently, alignment of several point clouds remains a prerequisite for subsequent analysis and processing steps.

In the literature, a plethora of papers have addressed the problem of 3D data registration, alternatively in terms of point cloud alignment or range image matching. Both rigid and non-rigid methods exists.

One can roughly divide the algorithms into three main families. First, the alignment using feature-based methods is achieved through the correspondence of feature primitives or keypoints, that, in addition, may have interesting properties with respect to the issue of interest (e.g., invariant to rigid-motion). Therefore, such methods rely heavily on the primitive extraction step. Features can be keypoints (Huber and Hebert, 2003b, Stamos and Leordeanu, 2003, Barnea and Filin, 2008, Weinmann et al., 2011), corners (Thirion, 1996), segment or curves (Stein and Medioni, 1992), local planes (Dold and Brenner, 2006), specific patterns (spheres, cylinders) or higher-level shape descriptors (Frome et al., 2004). An exhaustive search for corresponding feature pairs can be improved with pruning or selection techniques (such as RANSAC), or efficient hierarchical optimization techniques.

Secondly, in surface-based approaches, 3D datasets are represented by a surface model (usually, a mesh model). Then, the registration is performed directly on models rather than on the 3D point cloud. The introduction of surfaces allows to take into account holes, potential deformation between them and the introduction of well-established techniques such as thin-plate splines (Szeliski and Lavallée, 1996, Allen et al., 2003, Chui et al., 2004, Mitra et al., 2004).

Finally, one can find non-focused point-based methods, requiring neither feature extraction nor pre-modeling step. They may work on the full set of points or on specific subsets. Their aim is (1) to find correspondence between the two point sets and (2) to estimate the transformation. These methods can classified according to:

  • The performance of these two steps, simultaneously or sequentially.

  • The type of the underlying optimization method: global or local. Some authors even tried to mix both levels of information (Breitenreicher and Schnörr, 2011, Papazov and Burschka, 2011).

Simultaneous methods are very robust since the errors are distributed among all the points of the sets to limit distortion while preserving the geometry (Huber and Hebert, 2003a, Myronenko and Song, 2010). Although, their major shortcomings are the computation time, and the potential loss of small details owing to error accumulation. Sequential methods may produce imprecise results since errors can be propagated more easily, unless one can guarantee initial correct alignment (Chen and Medioni, 1992).

Global, deterministic or stochastic, optimizations using for instance branch-and-bound methods, genetic algorithms, or evolutionary methods also exhibit significant computing times (especially deterministic ones, but with guaranteed convergence), and are generally regarded as providing coarse registration (Silva et al., 2005). Therefore, they can be coupled with local methods, even if global minima can be reach without any initialization (Li and Hartley, 2007). The landmark contribution in local family is the Iterative Closest Point (ICP) algorithm. We have selected it for our study since this is one of the most widespread methods to compute registration of two point clouds. Other methods exist, based for instance on the Least-squares procedure (Gruen and Akca, 2005, Grant et al., 2012), the Random Sample Consensus algorithm (Chen et al., 1999), kernel correlation (Tsin and Kanade, 2004) or the Normal Distribution Transform (Ripperda and Brenner, 2005) exist. ICP iteratively minimizes the mean square error between points in point set and the closest points in the other one (Chen and Medioni, 1992, Besl and McKay, 1992). It is extensive used for a large variety of datasets and contexts. Nevertheless, due to sensitivity of the iterative method to noise and poor iteration, many variants have been developed to improve the five consecutive steps: selecting, matching, rejecting and weighting comprise the correspondence finding process, whereas transformation estimate consists in minimizing a given function. Authors often focus on specific issues, mainly convergence speed versus accuracy (Lu and Milos, 1997, Gelfand et al., 2003, Segal et al., 2009, Bae, 2010). ICP is only valid for pair-wise registration, and other methods are required for simultaneous registration of multiple point clouds (Craciun et al., 2010). According to Rodrigues et al. (2002), no optimal solutions exist. For the time being, the ICP method is still the state-of-the-art algorithm (Salvi et al., 2007).

Since ICP is an iterative descent algorithm, it requires a good initial estimation so as to converge to the global minimum. Besides, the ICP matching step is the most time-consuming part of the registration phase. Thus, improving the rate of convergence is crucial in making registration faster. To reduce the matching time, effective features of interest should be found. Such attributes may also be relevant to coping with erroneous associations between nearest points. This can frequently happen in case of objects acquired with different point densities or different points of view (Salvi et al., 2007). Consequently, two main solutions have been proposed in the literature: working at object level (Douillard et al., 2012) or computing, for each point, local features providing neighborhood information. Thus, several interesting local descriptors, have been developed successfully (Bae and Lichti, 2008). Indeed, the introduction of features of interest allows to focus on the most reliable regions in the registration process. The “reliability” may be evaluated according to planar criteria or with scale-space analysis (Sharp et al., 2002, Ho et al., 2009). For more complex environments, other primitives may be introduced, for example shapes like spheres, cylinders and tori in industrial areas (Rabbani et al., 2007).

Recently, several authors have focused on multi-scale local 3D point analysis for several purposes: dimension filtering for suitable operator definition (Unnikrishnan and Hebert, 2008, Digne and Morel, 2012), line extraction (Pauly et al., 2003), normal vector estimation (Unnikrishnan et al., 2010) and propagation (Digne et al., 2011) or 3D model compactness analysis (Novatnack and Nishino, 2007). In (Demantké et al., 2011 and Brodu and Lague, 2012), the multi-scale analysis of 3D lidar points, based solely on the geometrical information allows to retrieve for each point the optimal neighborhood size and the prominent behaviour of the vicinity (linear, planar, or volumetric).

The aim of this paper is to propose a general and automatic method for 3D point cloud alignment, applicable on the three kinds of topographic lidar datasets, mentioned in Section 4. We focus on pair-wise registration of datasets that do not exhibit large changes (especially in rotation), while coarse 3D registration issue is beyond the scope of this article: we assume the existence of a good a priori alignment before the two point sets of interest. If not, when scan orientations are unknown (e.g., terrestrial surveys), methods specific to the areas of interest are required (Makadia et al., 2006, Barnea and Filin, 2008, Theiler and Schindler, 2012).

Indeed, our goal is to assess how the introduction of the multi-scale features of (Demantké et al., 2011), may improve a standard fine-registration algorithm, namely the Iterative Closest Point method. This paper is an extension of the work we presented in Gressin et al. (2012) by giving more details, improving the geometrical features selection, and by presenting new results.

The geometrical features of interest are introduced first (Section 2). Then, the five steps of the ICP algorithm are described in Section 3. For three steps, the introduction of the proposed features is discussed. After a short presentation of the datasets in Section 4, the different variants of the ICP algorithm are evaluated and compared in Section 5. This allows to propose an optimized combination of ICP variants. Finally, conclusions are drawn in Section 6.

Section snippets

Finding features of interest

The method proposed in Demantké et al. (2011) aims to find, for each 3D point, the optimal neighborhood size. The primary goal was to find the most suitable local point set facing the interdependence problem: geometrical features depend on the choice of the neighborhood, whereas a good neighborhood choice should rely on the local geometry, and thus on geometrical features.

This is a two-step approach. At first time, three dimensionality features (1D, 2D, 3D) are proposed for a given neighborhood

ICP steps

The purpose of the ICP algorithm is to perform the registration of two coarsely aligned point clouds (a mobile point cloud registered on a fixed reference point cloud). It finds the best correspondence between two point sets by iteratively determining the translation and rotation parameters of a 3D rigid body transformation. In each step, the algorithm computes the closest point in the mobile scan for each point in the reference one. Then, the retrieved transformation is applied to the mobile

Datasets

Three kinds of lidar datasets are exploited in order to assess the relevance and performance of each proposed variant of the algorithm. These datasets have various point densities, point distribution, and points of view since they have been acquired with different lidar systems: airborne (ALS), terrestrial static (TLS), and mobile mapping systems (MMS).

Experiments

ICP Variants can be analysed and compared through several criteria: speed, overall accuracy, stability, tolerance to noise or outliers, and maximal initial misalignment. Since we deal with real lidar datasets, we will not tackle the noise/outlier tolerance issue. The problem of initialization will also be put aside since this is beyond the scope of the paper. Finally, ICP variant effectiveness will be assessed by speed and geometrical accuracy performance, considering the stability issue

Conclusion

In this paper, we have demonstrated how the standard and well-known Iterative Closest Point algorithm can be improved by using geometrical features which optimally describe the local shape around each 3D lidar point. Our method, which takes into account both the neighborhood shape and how confident in the estimate of this shape we are, allowed to improve two of the five steps of the method, namely the selection and rejection issues. Since the computation of the features of interest only

References (58)

  • J. Salvi et al.

    A review of recent range image registratioon methods with accuracy evaluation

    Image Vision Computing

    (2007)
  • M. Weinmann et al.

    Fast and automatic image-based registration of tls data

    ISPRS Journal of Photogrammetry and Remote Sensing

    (2011)
  • B. Allen et al.

    The space of human body shapes: reconstruction and parameterization from range scans

    ACM Transactions on Graphics

    (2003)
  • Bae, K.-H., 2010. Automated Registration of Three Dimensional Unorganised Point Clouds from Terrestrial Laser Scanners....
  • R. Bergevin et al.

    Towards a general multi-view registration technique

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (1996)
  • P. Besl et al.

    A method for registration of 3-D shapes

    I EEE Transactions on Pattern Analysis and Machine Intelligence

    (1992)
  • D. Breitenreicher et al.

    Model-based multiple rigid object detection and registration in unstructured range data

    International Journal of Computer Vision

    (2011)
  • C.-S. Chen et al.

    Ransac-based darces: a new approach to fast automatic registration of partially overlapping range images

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (1999)
  • H. Chui et al.

    Unsupervised learning of an atlas from unlabeled point-sets

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2004)
  • J. Demantké et al.

    Dimensionality based scale selection in 3D lidar point cloud

    The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences

    (2011)
  • Digne, J., Morel, J.-M., 2012. A Numerical Analysis of Differential Operators on Raw Point Clouds. Tech. Rep....
  • J. Digne et al.

    Scale space meshing of raw data point sets

    Computer Graphics Forum

    (2011)
  • C. Dold et al.

    Registration of terrestrial laser scanning data using planar patches and image data

    International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences

    (2006)
  • B. Douillard et al.

    Scan segments matching for pairwise 3d alignment

  • A. Frome et al.

    Recognizing objects in range data using regional point descriptors

  • N. Gelfand et al.

    Geometrically stable sampling for the icp algorithm

  • Godin, G., Rioux, M., Baribeau, R., 1994. Three-dimensional registration using range and intensity information. In:...
  • A. Gressin et al.

    Improving 3d lidar point cloud registration using optimal neighborhood knowledge

    ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences I-3

    (2012)
  • H. Gross et al.

    Extraction of lines from laser point clouds

    The International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences

    (2006)
  • Cited by (0)

    View full text