This invention relates generally to processing three-dimensional (3D) data, and more particularly to extracting planar features from 3D sensor data.
Low-cost depth cameras and 3D sensors, such as Microsoft Kinect®, can be used in robotics and computer vision applications. Typically, the sensors acquire a set of 3D points, i.e., a 3D point cloud, of a scene. Those 3D point clouds are generally noisy and redundant, and do not characterize high-level semantics in the scene.
Typically, primitive are extracted from the 3D point clouds to model the scene compactly and semantically. In particular, planes are one of the most important primitives because many man-made structures have planar surfaces.
Plane Extraction
A typical, method used for extracting planes from a point cloud is based on RANdom SAmple Consenus (RANSAC). That method first hypothesizes several candidates of planes, each of which is generated by randomly sampling three points from the point cloud and determining a set of points (referred to as inliers) that are on the plane defined by the three points. That method then selects from the candidates an optimal plane that is supported by the largest number of inliers. After removing the inliers from the point cloud, that method iterates the process to extract multiple planes. Because that method requires a relatively long time, several variants are known, e.g., Hough transformation and connected component analysis can be applied to the point cloud for pre-segmentation and then RANSAC can be applied to each of the segments. RANSAC can also be applied to local regions of the point cloud, followed by finding connected points on the plane from each of the local regions, called region growing.
A similar but much slower variant is voxel growing. Regions can also be grown from line segments extracted from individual scan lines. Another method clusters the points in a normal space and further clusters according to distances to an origin. To avoid normal estimation per point, graph-based segmentation using a self-adaptive threshold can be used.
Applications Using Planes
Planes extracted in this manner are used in various applications, such as robotics, computer vision, augmented reality, and 3D modeling. Compact and semantic modeling of scenes provided by planes is useful in indoor and outdoor 3D reconstruction, visualization, and building information modeling. Extracting major planes is a common strategy for table-top manipulation using robots, because planes help to segment objects placed on the planar table top. Planes have also been used for simultaneous localization and mapping (SLAM) and place recognition systems, where planes can be used as landmarks of the scene.
Although features in the form of planes are generally more accurate than point features, at least three planes whose normals span a 3D space are required to determine a 6-degrees-of-freedom (DoF) camera pose in such SLAM and place recognition systems. To avoid the degeneracy due to the insufficient number of planes, points and planes can be used together as landmarks.
Embodiments of the invention provide a method for concurrently extracting multiple planes from a 3D point cloud in real time. Our method can be classified as a region-growing-based method, to contrast with the prior methods, which generally use points or lines as primitives in region growing, our method uses regions as primitives.
Point clouds can be classified into organized point clouds and unorganized point clouds, in organized point clouds, the neighborhood relationship between 3D points is known, for example, because the points are generated from a depth map where we can define a 2D grid describing the neighborhood relationship. In unorganized point clouds, on the other hand, the neighborhood relationship is unknown, i.e., points are stored in a random order.
In one embodiment, a region is a small 2D patch of a depth map, from which an organized point cloud is generated. In this case, the 3D points are first partitioned into a set of disjoint regions according to the 2D grid structure defined on the depth map. A graph is defined, where regions correspond to nodes in the graph and the neighborhood relationship between regions correspond to edges in the graph. Uniform or non-uniform partitioning of the 2D grid structure is possible.
In another embodiment, a region is a voxel in a 3D space, in which an unorganized point cloud is defined. In this case, the 3D space is partitioned into a set of disjoint voxels, and each of the 3D points are assigned to a voxel according to the position of the point. Voxels including at least one 3D point are defined as nodes in a graph. Then edges in the graph are defined if two voxels are next to each other in the 3D space. Uniform or non-uniform partitioning of the 3D space possible.
After the graph is defined, agglomerative hierarchical clustering (AHC) is applied to the graph. AHC constructs a hierarchy of clusters in a “bottom-up” manner. That is, each node is defined as a cluster at the beginning, and then pairs of clusters are merged as one moves up the hierarchy. Each cluster includes a set of 3D points, to which a plane is fitted and the mean squared error (MSE) of the 3D points from the fitted plane is determined. The clustering systematically merges neighboring clusters in the graph such that the merged cluster has the minimum MSE.The clustering is repeated until no merge can produce the MSE that is smaller than a predefined threshold.
The boundaries of the planes generated by the clustering are coarse, because the planes are represented by a set of regions. If needed, the boundaries of the planes can be refined via region growing by using individual 3D points as primitives. We refer to planes after the clustering that have coarse boundaries as coarse planes, and planes after the refining that have fine boundaries as fine planes.
The method can extract planes at an effective frame rate of 30 Hz or higher for organized point clouds with 640×480 points, which is considerably faster than prior art methods, and enables real time applications.
As shown in
The point cloud is partitioned 110 into a set of disjoint regions 111. A selected size for the regions can vary depending on the application need, e.g., in the range of about 10 to 1000 points per region.
A graph 121 of nodes and edges is constructed 120 for the regions. The nodes represent the regions and the edges represent spatial relationships of the regions; regions that are next to each other are connected by an edge. Agglomerative hierarchical clustering (AHC) 130 is applied to the graph. At the beginning, each node is a cluster, and then pairs of clusters are merged systematically. Clusters of nodes after the clustering correspond to coarse planes 131. The coarse planes can be refined 140 using point-based region growing to obtain the fine planes 102.
The steps of the method can be performed in a processor 100 connected to memory and input/output interfaces by busses as known in the art.
Method Overview
We define an organized point cloud to be a set of 2D indexed 3D points pi,j, here 2D indices (i, j) and (i±1, j±1) reflect a 3D proximity relationship between points pi,j and pi±1,j±1 when the points are on the same surface. This index space can be called the image space, usually obtained from a depth map produced by devices such as a time-of-flight camera, structural light scanning system, or even rotating a scanning plane of a laser range finder.
2D Line Segment Extraction
We first describe our method using an example of extracting 2D line segments from a set of 2D points. Later, we extend our description to extracting planes in the 3D space from a set of 3D points.
As shown in
As shown in
For example, in one of the iterations of the AHC shown 603, the node g is found that has the minimum line fitting MSE. The thick line indicates that merging the node g with its left neighbor ef gives a smaller line fitting MSE than merging with its right neighbor thus the node g is merged with the node ef, producing the node efg 604.
If the minimal merging MSE is larger than a predefined threshold, which can usually be decided by the noise characteristics of the sensor, then the merging is canceled and the node g can be extracted as a line segment. Merging two nodes and calculating the resulting line fitting MSE can be done as a constant time operation using an integral image. Integral images are described in, e.g., U.S. Pat. No. 7,020,337.
Differences When Generalizing to Plane Extraction from 3D Point Clouds
Realizing the significance of neighborhood information hidden in the order of the points in the sequence, we generalize from 2D line extraction to 3D plane extraction in an organized point cloud, where the neighborhood information is stored in the 2D indices. However, this generalization is nontrivial, and not at all obvious. There are two major differences in the 3D case not found in the 2D line extraction:
Node Initialization: As opposed to the fine segment extraction, initial nodes, and thus any two nodes during and after merging, should have no identical points shared among the nodes.
Node Merging: In the line segment extraction, merging a node with its neighbor is a constant time operation with at most two merging attempts, either to left or to right neighbor. In the 3D case, a larger number of merging attempts might be required, because nodes are connected to form a graph and have a varying number of neighbors after the merging steps.
The reason for the first difference is because there are two possible directions for nodes to be merged with each other in the image space of the organized point cloud. After a few merging steps, the 3D points belonging to a certain node Bs form an irregular shape instead of maintaining an initial rectangular shape in the image space. Thus, if different nodes are allowed to have identical points, it is very difficult to efficiently handle the overlapping points when merging two nodes, even with the help of the integral image.
In the line segment extraction, merging two neighboring line segments still result in a line segment represented by a starting and ending index in the point sequence, which makes overlapping nodes feasible. It is important to notice that the overlapping nodes enable the line segment extraction to automatically split line segments at their boundaries, because nodes containing points at different line segments tend to have larger line fitting MSE than others, nodes c, d and h in
The second difference, i.e., the varying number of neighbors of a node, increases the number of merging attempts to find the minimum merging USE. However, we empirically found that the average number of merging, attempts stays at a small number and thus the clustering can be performed efficiently.
Fast Plane Extraction Method
Our plane extraction method includes the following main steps: 1) graph initialization; 2) agglomerative hierarchical clustering (AHC); and 3) boundary refinement, if the application only requires rough segmentation of planar regions, e.g., the application requires the normals of planes but not their boundary information, then the boundary refinement step can be omitted, which can increase the processing frame rate to more than 50 Hz.
Graph Initialization
As described before, our method generates disjoint regions and non-overlapping nodes during initialization. Let us consider the scene shown in
Because of inherent limitations in 3D sensors, some regions of the scene cannot he sensed, leading to missing data, e.g., a glass window next to a window shutter. The regions containing missing data points should not be considered. Also, some regions can have a very large plane fitting MSE, which is usually because the region violates the planarity assumption, and thus should not be considered.
Two other difficult scenarios are regions including depth discontinuities and regions at a boundary between two planes. In the former ease, a region contains two sets of points lying on two surfaces that do not connect in 3D but are close to each other in the image space because of the view direction, usually one surface partially occlude the other, e.g., a monitor occludes the wall A behind the monitor.
If principle component analysis (PCA) is performed on points belonging to this region for plane fitting, then the fitted plane is nearly parallel to the view direction and still has a small MSE. Merging this “outlier” region with its neighbor will have the bad effect on the plane fitting result because of the well-known issue of over-weighting outlier in least-squares methods. In the latter case, a boundary region contains points close to each other in 3D but lie on two difference surfaces e.g., the corner of the room, which decreases the plane fitting accuracy after merged. Thus, we discard such regions.
Agglomerative Hierarchical Clustering
Instead of operating on a double linked list as in the case of line extraction shown in
If this minimum merging MSE exceeds some predefined threshold TMSE, then a plane segment v is found and extracted from the graph. Otherwise, the merged node umerge is added back to the graph by edge contraction between v and ubest.
Implementation Details
Several implementation details improve the speed and accuracy for the coarse plane extraction:
Boundary Refinement
For many applications, coarse planes may be insufficient because of artifacts such as sawtooth edges at plane boundaries, unused data points at boundaries of occlusions, and oversegmentation.
Such artifacts can be reduced by erosion of boundary regions followed b region growing using individual points as primitives. The region growing starts from all new boundary points to assign all unused data points to the closest plane that was extracted previously.
Applications
Planes extracted by our method can be used in many applications. Planes can be used for registering a set of 3D point clouds defined in different coordinate systems. The registration can be done when there are three or more planes whose normals span a 3D space; otherwise, the 3D points can be used in conjunction with planes for the registration.
Planes are less affected by noises of 3D sensor measurements than individual points. Thus, by using planes, we can determine more precise coordinates of 3D locations than using the original 3D points in a 3D point cloud.
For example, after the planes are fitted to the point cloud shown in
Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention. Therefore, it is the object of the appended s to cover all such variations and modifications as come within the true spirit and scope of the invention.
Number | Name | Date | Kind |
---|---|---|---|
5961571 | Gorr | Oct 1999 | A |
5978504 | Leger | Nov 1999 | A |
6650779 | Vachtesvanos | Nov 2003 | B2 |
7129942 | Martin | Oct 2006 | B2 |
7643966 | Adachi | Jan 2010 | B2 |
8004517 | Edelsbrunner | Aug 2011 | B1 |
8005767 | Cassella | Aug 2011 | B1 |
8144921 | Ke et al. | Mar 2012 | B2 |
8199977 | Krishnaswamy | Jun 2012 | B2 |
8243988 | Buddemeier | Aug 2012 | B1 |
8289321 | Gutmann | Oct 2012 | B2 |
8340400 | Lukas | Dec 2012 | B2 |
8355552 | Schwartz | Jan 2013 | B2 |
8660365 | Lukas | Feb 2014 | B2 |
20020059042 | Kacyra | May 2002 | A1 |
20030038802 | Johnson | Feb 2003 | A1 |
20040108999 | Martin | Jun 2004 | A1 |
20050166413 | Crampton | Aug 2005 | A1 |
20060034545 | Mattes | Feb 2006 | A1 |
20060052943 | Ramani | Mar 2006 | A1 |
20060182327 | Mundy | Aug 2006 | A1 |
20060269144 | Kubota | Nov 2006 | A1 |
20070031064 | Zhao | Feb 2007 | A1 |
20070052706 | Martin | Mar 2007 | A1 |
20070070226 | Matusik | Mar 2007 | A1 |
20080130978 | Neemuchwala | Jun 2008 | A1 |
20080294332 | Levanon | Nov 2008 | A1 |
20090016564 | Ke | Jan 2009 | A1 |
20090226060 | Gering | Sep 2009 | A1 |
20090268214 | Lucic | Oct 2009 | A1 |
20100027892 | Guan | Feb 2010 | A1 |
20100092075 | Lee | Apr 2010 | A1 |
20100207936 | Minear | Aug 2010 | A1 |
20100277571 | Xu | Nov 2010 | A1 |
20100316257 | Xu | Dec 2010 | A1 |
20110054783 | Kishikawa | Mar 2011 | A1 |
20110134225 | Saint-Pierre | Jun 2011 | A1 |
20110274343 | Krishnaswamy | Nov 2011 | A1 |
20110310088 | Adabala | Dec 2011 | A1 |
20130011069 | Quan | Jan 2013 | A1 |
20130156262 | Taguchi | Jun 2013 | A1 |
20130166255 | Chang | Jun 2013 | A1 |
20130208098 | Pujol Alcolado | Aug 2013 | A1 |
20130282208 | Mendez-Rodriguez | Oct 2013 | A1 |
20140003705 | Taguchi | Jan 2014 | A1 |
20140072213 | Paiton | Mar 2014 | A1 |
20140093160 | Porikli | Apr 2014 | A1 |
20140193076 | Gardiner | Jul 2014 | A1 |
20140254941 | Schwartz | Sep 2014 | A1 |
20150117708 | Guigues | Apr 2015 | A1 |
20150127650 | Carlsson | May 2015 | A1 |
Entry |
---|
Boykov et al—An experimental comparison of min-cut max-flow algorithms for energy minimization in vision—Sep. 2004—IEEE trascation on PAMI vol. 26-9—pp. 1124-1137. |
Boykov et al—An experimental comparison of min-cut max-flow algorithms for energy minimization in vision—Sep. 2004—IEEE trascation on PAMI vol. 26-9-ppl 124-1137. |
McKinnon, et al. Masters thesis: Point—line segment and region based stero matching for mobile robotics—Master of science thesis—Aug. 2009. |
R. Schnabel, R. Wahl, and R. Klein, “Efficient RANSAC for point cloud shape detection,” Computer Graphics Forum, vol. 26, No. 2, pp. 214-226, Jun. 2007. |
J. Poppinga, N. Vaskevicius, A. Birk, and K. Pathak, “Fast plane detection and polygonalization in noisy 3D range images,” in Proc. IEEE/RSJ Int'l Conf. Intelligent Robots and Systems (IROS), 2008, pp. 3378-3383. |
K. Georgiev, R. T. Creed, and R. Lakaemper, “Fast plane extraction in 3D range data based on line segments,” in Proc. IEEE/RSJ Int'l Conf. Intelligent Robots and Systems (IROS), 2011, pp. 3808-3815. |
Number | Date | Country | |
---|---|---|---|
20150154467 A1 | Jun 2015 | US |