Indoor laser point cloud registration method based on wall surface structure

Abstract

The invention discloses an indoor laser point cloud registration method based on a wall surface structure. The method comprises the following specific implementation steps: (1) indoor laser point cloud ground and ceiling extraction; (2) wall surface screening based on ground and ceiling constraint; (3) virtual shadow removal based on point cloud density; and (4) NICP registration based on point cloud local features. The method relates to the field of laser point cloud registration, an indoor wall surface structure is obtained through constraint of the ground and a ceiling, and high-precision registration of indoor point cloud is carried out by utilizing wall surface normal direction.

Description

Technical fields:

[0001] This invention relates to the field of laser point cloud registration, specifically a laser point cloud registration method based on indoor wall structure constraints. Background technology:

[0002] With the continuous advancement of 3D data acquisition technology, point clouds have become an important form for describing the geometry and structure of real-world objects, and point cloud registration plays a crucial role in point cloud data processing. Point cloud registration is the process of transforming point clouds from different viewpoints to the same coordinate system. Many scholars have studied this and proposed various registration algorithms, such as registration methods based on local features, global matching methods, and methods combined with deep learning. The most widely used point cloud registration algorithm is the Iterative Closest Points (ICP) algorithm proposed by Besl and McKay in 1992. Its core idea is to continuously optimize the transformation matrix parameters of the point cloud data through iterative processing, and finally obtain the optimal solution. These methods mainly deal with relatively uniform point cloud registration problems. For indoor point cloud registration, in addition to the influence of scene time-varying and observation perspective differences, the shadows produced by illumination and object reflections are not exactly the same at different times of data acquisition. These influences will make the local feature differences of the point cloud obvious, resulting in a significant decrease in registration accuracy. Currently, there is still a lack of effective solutions for fine registration of complex indoor scenes.

[0003] In indoor scenes, there are various types of targets and their structural forms are complex and varied. Walls, on the other hand, have stable and regular structures, usually with a stable height from the ground to the ceiling. Different walls also have regular edges. If the constraints of the walls can be fully utilized, it is expected to improve the point cloud registration effect in complex indoor scenes. Summary of the Invention:

[0004] The purpose of this invention is to address the registration challenges caused by object occlusion and viewing angle differences when registering point clouds in complex indoor scenes. By utilizing the connection relationships between indoor walls, floors, and ceilings, as well as the mutual constraints between walls, laser point cloud features are constructed to provide more stable structural constraints for the registration process. The key is to select the actual walls from the numerous planar structures in the room and use the normal direction of the walls to constrain the point cloud registration, thereby improving the registration effect of indoor scenes.

[0005] This invention follows the following technical solution: an indoor laser point cloud registration method based on wall structure, characterized by comprising the following steps:

[0006] Step 1: Extraction of indoor laser point cloud data from the floor and ceiling;

[0007] Step 2: Wall selection based on floor and ceiling constraints;

[0008] Step 3: Ghost image removal based on point cloud density;

[0009] Step 4: NICP registration based on local point cloud features.

[0010] The advantages of this invention compared to the prior art are:

[0011] (1) The normal direction of the stable wall surface is introduced as a constraint condition into the registration process. Based on the planar extraction, the geometric relationship between the wall surface and the ground and ceiling, as well as the constraint that the normal directions are perpendicular to each other, are used to extract the wall structure in the scene. The wall surface features are used to replace the general point cloud features, which improves the accuracy of indoor point cloud registration.

[0012] (2) It effectively solves the mismatch problem caused by the imbalance of point clouds. When the viewing angle changes, due to the strong occlusion of the indoor scene, the point cloud data collected under different viewing angles will be very different. The registered point clouds have great differences in terms of data volume and target integrity, which is a typical imbalanced point cloud registration problem. After introducing the plane normal direction, point clouds of different densities still have a relatively consistent normal direction, which can effectively solve the influence of the imbalance of point clouds. Attached image description:

[0013] Figure 1 : Flowchart of method implementation.

[0014] Figure 2 : Schematic diagram for determining the direction of the wall normal.

[0015] Figure 3 : Point cloud of lidar to be registered.

[0016] Figure 4 Registration results of indoor point clouds using the NICP method.

[0017] Figure 5 The method of this invention is used for indoor point cloud registration results. Detailed implementation method:

[0018] The implementation process of this invention is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0019] Step 1: Extraction of indoor laser point cloud from the floor and ceiling

[0020] If the ground and ceiling are planar, the ground and ceiling can be directly extracted using modeling methods. However, for actual interior spaces, the point clouds of the ground and ceiling may exhibit curvature and discontinuity, making it difficult to describe with uniform planar parameters. Therefore, it is necessary to separate multiple planar regions.

[0021] In this method, principal component analysis is used to extract planar point clouds, construct the point cloud covariance matrix, and calculate the eigenvalues ​​λ of the matrix. j|j＝1,2,3 (λ1≥λ2≥λ3≥0), then calculate the point cloud features to filter the point cloud for planar structures. Point cloud features include linear L... λ Planarity P λ divergence S λ :

[0022]

[0023] Set a characteristic threshold. When the planarity of a point is greater than the threshold, or the linearity and divergence are less than the set threshold, the point is considered a planar point, and the normal direction of the planar point is calculated.

[0024] Based on the angle between the plane point normal direction and the Z-axis, when the angle is small, that is, when the plane normal and the Z-axis vector are approximately parallel, the plane point is either the ground or the ceiling. Furthermore, based on the height value, the ground point cloud and the ceiling point cloud can be distinguished.

[0025] Step 2: Wall selection based on floor and ceiling constraints

[0026] After separating the ground and ceiling, the planar point cloud contains not only the necessary indoor wall point clouds, but also outdoor point clouds captured through windows, ghosting caused by point cloud reflections, and small planar areas formed by indoor furniture. These all constitute interference in point cloud registration. Outdoor point clouds appear in different areas depending on the acquisition perspective. Point clouds generated by specular reflection usually connect with the real point cloud to form reflections, and the reflected point clouds are sparse. The planar area formed by indoor furniture is small and does not connect with both the ground and ceiling simultaneously. Interfering point clouds in the indoor space can be removed by using planar parameters and indoor structural relationships to filter out stable walls.

[0027] In the wall screening, the wall height constraint is first used. Assuming that the height of the ceiling from the ground is H1, for the planar point cloud to be judged, its upper edge and lower edge are calculated. When the height difference h1 between the upper edge and the lower edge satisfies the formula (2), the plane is not a wall and needs to be removed.

[0028] H1-h1>Th1 (2)

[0029] After height filtering, the normal direction of the plane to be judged is extracted. By utilizing the constraint relationship that the normal direction of the wall is perpendicular to the normal direction of the ground and ceiling, the non-perpendicular planes are removed by the inner product operation of the normal direction.

[0030] Potential wall point clouds are selected by using height and normal direction constraints.

[0031] Step 3: Ghost Removal Based on Point Cloud Density

[0032] Ghost images in point clouds are mainly false point clouds formed by reflections from glass, etc. They cannot be completely removed by the filtering in step 2. However, the point clouds on the wall formed by ghost images are sparser than the point clouds on the real wall. Therefore, the point cloud density can be used to remove ghost images.

[0033] Projecting the plane onto the ground yields a projection line. The length of this line represents the width *w* of the wall, and the distribution range of the wall along the Z-axis represents the height *h* of the plane. Assuming the current plane contains *N* points, the point cloud surface density *ρ* of the wall can be calculated as follows:

[0034] ρ=N / w×h (3)

[0035] When equation (4) is satisfied, the current point cloud is considered to be sparse and is a virtual point cloud that needs to be removed. Th2 is the threshold for the sparseness of the point cloud, which is related to the number of lidar lines and the speed of motion during scanning.

[0036] ρ<Th2 (4)

[0037] Step 4: NICP registration based on local point cloud features

[0038] For the extracted wall point cloud, the ICP variant algorithm NICP, based on local point cloud features, is used for registration. The NICP algorithm first calculates the features of each point in the point cloud, namely the normal direction and surface curvature of its surface, to label each point; then, it finds matching point pairs in two sets of point clouds based on the distance and features of the points; finally, it uses the least squares method to minimize the objective function to solve for the point cloud transformation matrix. Here, the objective function includes the point-to-surface projection and normal direction rotation errors.

[0039] Specifically, for indoor point cloud registration, planar constraints need to be introduced. However, the direction of the normal to the plane has a 180-degree directional ambiguity (pointing to the interior and exterior). To solve the directional ambiguity problem, the direction of the wall normal is limited to the direction pointing to the interior.

[0040] Methods for constraining the direction of the wall normal include: Figure 2 As shown, the center point of the plane is first projected onto the ground to obtain the ground projection point m. p , with m p Using a circle as the center and a set length as the radius, take a neighborhood of the ground point and calculate the geometric center of the ground within the neighborhood as m. f plane normal direction With m p Point to m f vector The included angle should be less than 90 degrees, otherwise Conversely, this gives us the direction of the wall normal pointing into the interior:

[0041]

[0042] In the formula, sign() is the sign operation, and (,) is the inner product operation.

[0043] When the normal direction is defined, using NICP for indoor point cloud registration can yield better registration results than using NICP without defining the normal direction.

[0044] To verify the effectiveness of this invention, point clouds from a measured indoor scene were registered. The two sets of point clouds to be registered are shown below. Figure 3 As shown (using different gray levels to represent two sets of point clouds), the registration result of directly using the NICP method for the two sets of point clouds is as follows. Figure 4 As shown, even when the two point clouds are basically registered, there are still many jagged edges around them. The registration result of the method of this invention is as follows. Figure 5 As shown, compared to Figure 4 The burrs are significantly reduced and the edge overlap is better. The comparison results demonstrate the effectiveness of the method of the present invention.

Claims

1. A method for registering indoor laser point clouds based on wall surface structure, characterized in that Comprising the following steps: Step 1: Indoor laser point cloud ground and ceiling extraction; Step 2: Wall surface screening based on ground and ceiling constraints; Step 3: Ghost removal based on point cloud density; Step 4: NICP registration based on point cloud local features.

2. The method of claim 1, wherein, In step 2, the geometric relationship among the ground, ceiling and the plane to be determined is comprehensively utilized to screen the wall surface. Assuming that the height of the ceiling from the ground is H1, for the point cloud of the plane to be determined, the height difference h1 of the upper edge and the lower edge is calculated, and when formula (1) is satisfied, the plane is not a wall surface, and in addition, the perpendicular relationship between the normal direction of the wall surface and the normal direction of the ground and the ceiling is further utilized to remove the planes that are not perpendicular. H1-h1>Th1 (1) 3. The method of claim 1, wherein, In step 4, the relationship between the wall normal direction and the direction of the ground point cloud adjacent to the wall is used to limit the normal direction of the wall point cloud to point inward. In specific implementation, the center point of the plane is first projected onto the ground to obtain the ground projection point m. p , with m p Using a circle as the center and a set length as the radius, take a neighborhood of the ground point and calculate the geometric center of the ground within the neighborhood as m. f plane normal direction With m p Point to m f vector The included angle should be less than 90 degrees, otherwise Conversely, this method can be expressed as formula (2), where sign() is the sign operation and (,) is the inner product operation.

Images