Mass airborne LiDAR point cloud rapid registration algorithm fusing color features
Through the FCFRA algorithm, downsampling and rasterization processing combined with SuperPoint and SuperGlue algorithms, the problem of high complexity of registration time for massive airborne LiDAR point clouds is solved, and fast and accurate point cloud registration is achieved, and it is suitable for different Data scenario.
Patent Information
- Application Number
- CN202411120751.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-22
AI Technical Summary
When processing massive airborne LiDAR point cloud data, the point cloud registration time is complex and it is difficult to effectively utilize the color information of the point cloud, resulting in low registration efficiency in large-area data and sensitive to noise and local interference. .
Massive airborne LiDAR point cloud fast registration algorithm (FCFRA) using a massive airborne LiDAR point cloud fast registration algorithm (FCFRA) that integrates color characteristics. Through downsampling and two-dimensional rasterization processing, point clouds are converted into raster images, and feature points are extracted and matched using SuperPoint and SuperGlue algorithms. , combine singular value decomposition method and random sampling consistency algorithm to calculate the rotation translation matrix to achieve point cloud registration.
It realizes the rapid registration of massive point cloud data in large outdoor areas, with fast speed, high accuracy, good adaptability, can effectively process point cloud data with different time and space changes, and is suitable for point clouds without color information.
Smart Images

Figure CN119006545A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of laser point cloud data processing, and in particular to a rapid registration algorithm for massive airborne LiDAR point clouds integrating color features. Background Art
[0002] With the rapid development of airborne laser radar technology, it has become more convenient to obtain laser point cloud data in large outdoor areas. It has been widely used in power line inspection, rapid terrain measurement, forestry survey, geological monitoring and other fields. Point cloud registration is an important link in these applications. Point cloud registration is the estimation of rotation and translation transformation between point clouds. Its main purpose is to obtain the rotation and translation matrix when the distance between the corresponding points of the two point clouds is minimized, so as to obtain a new point cloud containing the complete three-dimensional data of the target.
[0003] At present, the research on point cloud registration can be divided into two categories: classical algorithms and deep learning algorithms. The most representative classical algorithm is the iterative closest point (ICP) algorithm. The ICP algorithm alternately executes two steps, namely, finding corresponding points and the least squares optimization algorithm, to update the registration state. However, the algorithm requires a good initial value, otherwise it will fall into the local optimum and is sensitive to noise and local interference. Therefore, many scholars have studied the improved algorithm of ICP. Yang et al. [YANG JL, LI HD, CAMPBELL D, et al. Go-ICP: A Globally Optimal Solution to 3D ICP Point-Set Registration [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38 (11): 2241-2254.] proposed the Go-ICP algorithm based on branch and bound to determine the global optimal solution. In addition, there are convex relaxation, Riemann optimization and mixed integer programming algorithms. However, these methods have high computational overhead and cannot be well used for registration between large-scale outdoor massive point cloud data. Another representative method is the Normal Distribution Transformation (NDT) algorithm based on the probability model and its improved algorithm. This type of method can improve the registration accuracy and has a certain robustness, but it also has a high computational complexity and is inefficient when processing a large amount of outdoor point cloud data. It also requires that the data distribution is relatively uniform and continuous, and the algorithm parameter selection is relatively complex. There is also another type of algorithm that first extracts feature points on the point cloud, then matches the feature points, and finally uses the matched feature points to calculate the transformation matrix between the point clouds to complete the registration. Representative algorithms include PFH (Point Feature Histograms), FPFH (Fast Point Feature Histograms), 3Dsc (3DShape Context), etc. These methods mainly use the spatial geometric features of the point cloud to extract feature points. For massive data, the feature extraction time is long and the registration efficiency is low.
[0004] In recent years, deep learning algorithms have developed rapidly. Scholars have used deep learning methods to conduct research on feature descriptors of point clouds, key point detection, and end-to-end registration networks. The literature [Bai Zhengyao, Xu Zhu, Zhang Yihan. A 3D point cloud registration method using pseudo corresponding point generation [J]. Computer Science, 2023, 50(09): 210-219. [Bo ZR, Xu Z, Zhang Y H. Deep Artificial Correspondence Generation for 3D Point Cloud Registration [J]. Computer Science, 2023, 50(09): 210-219.]] proposed a deep pseudo corresponding point generation network, which takes into account the structural information of the source point cloud and the target point cloud, and exchanges the contextual information between the source and target point clouds in the feature extraction stage to improve the registration performance. The literature [Pham QH, Uy MA, Hua BS, et al. LCD: Learned Cross-Domain Descriptors for 2D-3D Matching. Proceedings of the AAAIConference on Artificial Intelligence. 2020, 34: 11856-11864.] proposed a cross-domain descriptor (Learned Cross-Domain Descriptors, LCD) that can realize the registration between 2D image feature points and 3D point cloud feature points, as well as between 3D point clouds and 3D point cloud feature points, and achieved good point cloud registration results. These deep learning methods have made great progress in point cloud registration. However, there are still many challenges in using deep learning methods to solve the point cloud registration problem. First, some of the generated corresponding points are prone to become outliers, which limits the application of the algorithm in large-scale point cloud scenes. Second, these algorithms are currently based mainly on the structural information of the point cloud, and rarely use information such as the color of the point cloud.
[0005] The current traditional algorithms and deep learning algorithms all have the problem of high time complexity and are sensitive to the number of point clouds. The point cloud acquisition density of current airborne LiDAR (Light Detection and Ranging) equipment is getting higher and higher. The number of point clouds in large areas is massive, reaching tens of millions has become the norm. How to align these massive point clouds is a difficult problem that needs to be solved. In addition, these algorithms mainly use the spatial structure characteristics of point clouds, and the color information of point clouds contains rich grayscale features. For example, the spatial structure characteristics of a flat roof with rich textures are not significant, but the grayscale characteristics are very obvious. If only the spatial structure characteristics of point clouds are used, it will not be possible to effectively extract point cloud feature points and complete the alignment. Summary of the invention
[0006] In response to the above problems, the present invention proposes a Fusion Color Feature Registration Algorithm (FCFRA) for rapid registration of massive airborne LiDAR point clouds that integrates color features. In the registration process, the algorithm fully utilizes the ability of current airborne LiDAR devices to obtain point cloud color information, and completes the registration calculation of different point clouds. The algorithm of the present invention can solve the problem of rapid registration of airborne LiDAR point cloud data in large outdoor areas. This registration can be used in airborne LiDAR power line inspection. By registering point clouds acquired at different times, it can detect spatiotemporal changes such as tower inclination, power line sag changes, and dancing amplitude. The algorithm of the present invention can be used to achieve rapid splicing of point cloud data in large areas, which can be used in the fields of rapid terrain measurement, forestry surveys, geological monitoring, etc.
[0007] In order to achieve the above object, the present invention adopts the following technical solutions:
[0008] A rapid registration algorithm for massive airborne LiDAR point clouds integrating color features, including:
[0009] Step 1: Downsample the airborne 3D laser point cloud data to be registered, perform 2D rasterization on the downsampled 3D laser point cloud using central projection to obtain a raster image, perform median filtering on the raster image, and remove salt and pepper noise on the raster image;
[0010] Step 2: Use the SuperPoint algorithm to extract feature points from the raster image after removing salt and pepper noise;
[0011] Step 3: Use SuperGlue algorithm to match raster image feature points;
[0012] Step 4: Use the matched 2D image feature points to inversely calculate the matched 3D point cloud feature points;
[0013] Step 5: Using the matched 3D point cloud feature points, the singular value decomposition method combined with the random sampling consensus algorithm is used to calculate the optimal transformation parameters between the laser point cloud data to be registered;
[0014] Step 6: Use the transformation parameters to perform transformation operations on the point cloud data to complete the registration of the 3D laser point cloud data.
[0015] Furthermore, in step 1, the downsampled three-dimensional point cloud to be registered is subjected to two-dimensional rasterization processing in the following manner:
[0016]
[0017] Among them, u, v represent the pixel coordinates of the corresponding image point on the grid map, s is the scale factor, K represents the camera internal parameter matrix, R represents the rotation matrix between the camera coordinate system and the world coordinate system, T represents the translation matrix between the sampled point cloud, t represents the translation vector, [X w ,Y w ,Z w ] represents the three-dimensional coordinates of the point cloud, f represents the camera principal distance, d x and d y They represent the actual physical size of a single pixel in the horizontal u and vertical v directions on the camera sensor, respectively. u0 and v0 are the coordinates of the center point of the image coordinate system in the pixel coordinate system.
[0018] When generating a raster image, the corresponding pixel coordinates of each point in the three-dimensional laser point cloud are calculated according to the above formula, and the color value of the corresponding three-dimensional laser point cloud is assigned to the corresponding pixel point.
[0019] Furthermore, the translation vector is determined according to the following formula:
[0020]
[0021] Among them, X max , X min Respectively represent the maximum and minimum coordinates of the point cloud in the X direction, and Y max , Y min They represent the maximum and minimum coordinates of the point cloud in the Y direction, and h is the height of the raster image.
[0022] Furthermore, in step 1, the size of the raster image is determined according to the following formula:
[0023]
[0024] w=x×h,
[0025] Among them, h is the height of the raster image, w is the width of the raster image in pixels, num is the number of points in the point cloud, s is the scale factor, and k is the aspect ratio of the raster image.
[0026] Furthermore, the aspect ratio of the raster image is determined according to the following formula:
[0027]
[0028] Among them, X max , X min Respectively represent the maximum and minimum coordinates of the point cloud in the X direction, and Y max , Y min They represent the maximum and minimum coordinates in the Y direction of the point cloud respectively.
[0029] Furthermore, in step 1, when the raster image is generated, the one-to-one correspondence between each image point and the three-dimensional point in the laser point cloud needs to be recorded to facilitate the subsequent inverse calculation of the corresponding three-dimensional point cloud feature points from the two-dimensional image feature points.
[0030] Furthermore, in step 1, when the raster image is generated, Python is used for encoding, the point cloud data operation is moved into the Numpy array and its vectorized processing is implemented.
[0031] Furthermore, in step 1, for laser point cloud data without color information, a layered coloring method is used to calculate the color value of the corresponding point cloud according to the elevation value of the point cloud:
[0032]
[0033] Where f = 255 × (Z-z_Min) / (z_Max-z_Min), L is a constant, R, G, B are the red, green, and blue components of the point cloud color, Z is the elevation value of the point cloud, and z_Max and z_Min are the maximum and minimum elevation values of all point clouds, respectively.
[0034] Furthermore, the step 5 comprises:
[0035] Step 5.1: Sample n pairs of conjugate points from the feature points of the matched 3D point clouds A and B;
[0036] Step 5.2: Based on the sampled conjugate points, the transformation parameters are calculated using the singular value decomposition method:
[0037]
[0038] [U,S,V]=SVD(H),
[0039] R=VU T ,
[0040] T=-R*centroid A +centroid B ,
[0041] in, Respectively represent the coordinates of the sampling points in the conjugate point cloud A and B, centroid A and centroid B Respectively represent the average center of the sampling points in A and B, H represents the covariance matrix of the sampling points, U, S, V represent the matrix obtained by performing singular value decomposition on the covariance matrix, R represents the rotation matrix between the sampling point clouds, and T represents the translation matrix between the sampling point clouds;
[0042] Step 5.3: Calculate the non-sampling points in A according to the following formula: After transformation Point, if The corresponding point in B If the distance is less than a given threshold, the conjugate point j is considered to be an interior point, otherwise it is not an interior point;
[0043]
[0044] Step 5.4: If the proportion of interior points to all conjugate points reaches the specified value, the algorithm terminates and all interior points are used to optimize the transformation parameters;
[0045] Step 5.5: If the inlier ratio does not reach the specified value, continue to execute steps 5.1 to 5.4 until the inlier ratio meets the requirements or the operation reaches the specified number of iterations. At this time, the transformation parameters corresponding to the highest inlier ratio are used as the optimal transformation parameters.
[0046] Compared with the prior art, the present invention has the following beneficial effects:
[0047] The present invention designs a fast point cloud registration method that uses a point cloud rasterization strategy to integrate the color information of the point cloud into the corresponding feature point extraction, and through experimental analysis, finds the balance point between the downsampling size and the registration effect, and optimizes the efficiency of raster image generation by using the high-speed processing characteristics of the Numpy array; the SuperPoint+SuperGlue combination method is selected to realize the feature point extraction and matching of large-difference point cloud data; the singular value decomposition method combined with the random sampling consensus algorithm is designed to complete the rotation and translation matrix calculation between the point clouds to be matched; finally, a relatively comprehensive evaluation index system is established, and the effectiveness of the algorithm is evaluated and analyzed in detail. The present invention can be effectively used for the registration of large-area massive airborne LiDAR point cloud data outdoors, and has fast speed, high precision, and good adaptability. The algorithm of the present invention can realize the rapid splicing of large-area point cloud data. Even for data with large differences, the algorithm of the present invention can well complete the point cloud registration, and can also complete the registration for point clouds without color information, and the algorithm also has good adaptability. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 This is a flow chart of a rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to an embodiment of the present invention;
[0049] Figure 2 An example of a raster image generated by a point cloud according to an embodiment of the present invention;
[0050] Figure 3 This is an example diagram of the SuperPoint algorithm network structure;
[0051] Figure 4 This is an example diagram of the SuperGlue algorithm framework;
[0052] Figure 5 The effect diagram of the collected data 1 provided by the embodiment of the present invention;
[0053] Figure 6 The effect diagram of the collected data 2 provided by the embodiment of the present invention;
[0054] Figure 7 One of the point cloud registration process result diagrams provided by the embodiment of the present invention;
[0055] Figure 8 The front and side view effect of the original point cloud registration provided by the embodiment of the present invention;
[0056] Fig. 9 The side view effect after the original point cloud registration provided by the embodiment of the present invention;
[0057] Fig.10 The second result diagram of the point cloud registration process provided by the embodiment of the present invention;
[0058] Fig.11 The point cloud thinning experiment results provided by the embodiment of the present invention;
[0059] Fig.12 Four sets of data information provided for the embodiments of the present invention;
[0060] Fig.13 A first set of data registration effect diagrams provided by an embodiment of the present invention;
[0061] Fig.14 A second set of data registration effect diagrams provided by an embodiment of the present invention;
[0062] Fig.15 A third set of data registration effect diagrams provided by an embodiment of the present invention;
[0063] Fig.16 This is the fourth set of data registration effect diagrams provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0064] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:
[0065] The object of the Fusion Color Feature Registration Algorithm (FCFRA) for massive airborne LiDAR point cloud registration is mainly outdoor large-area airborne LiDAR point cloud data. One of the characteristics of this type of data is the large amount of data, and the number of point clouds can reach tens of millions or even hundreds of millions. The existing point cloud registration algorithms are difficult to effectively achieve the registration of such point clouds. The strategy adopted by the FCFRA algorithm is to use the matched feature points to calculate the transformation matrix between point clouds to complete the registration. In this matching strategy, since the point cloud is discretely distributed, when the number of point clouds is huge, the extraction and matching of point cloud feature points becomes very time-consuming. In many algorithms, although the amount of data can be reduced by downsampling, in the face of massive data, if the downsampling is too much, the point cloud data will not reflect the characteristics of the original data, resulting in registration failure. If the downsampling is too little, the problem of too much data will still occur. To address this problem, the FCFRA algorithm proposed in the present invention adopts a rasterization strategy to transform the discretely distributed point cloud into a closely arranged raster image, in which the color information of the point cloud becomes the grayscale value of the corresponding pixel of the raster image. This has two advantages: first, the grayscale feature extraction and matching is performed on the two-dimensional image, so that the feature extraction and matching problem of the three-dimensional point cloud is transformed into the feature extraction and matching of the two-dimensional image, which can greatly reduce the difficulty of feature extraction and matching; second, for the feature extraction and matching of the two-dimensional image, although it is also related to the size of the image, due to the regular arrangement of the pixels, the time consumption will be much less than that of the three-dimensional point cloud, which can greatly improve the efficiency of feature extraction and matching and reduce the time consumption of the entire point cloud registration. In addition, the present invention further combines the point cloud downsampling strategy to control the size of the raster image and finds the balance point between the downsampling size and the feature extraction and matching effect. At the same time, the FCFRA algorithm also makes full use of the high-speed processing capability of the Numpy array in its implementation, optimizes the raster image calculation, and improves the speed of raster image generation.
[0066] There are two different methods for rasterization strategy. One is to use vertical projection, directly map the XY coordinates of the three-dimensional point cloud to the pixel coordinates of the image according to a certain scale, and then use the color value of the three-dimensional point cloud as the corresponding pixel grayscale value; the other is to use central projection. This method uses the camera imaging model to convert the three-dimensional point cloud to the two-dimensional image in the way of central projection, and then use the color value of the three-dimensional point cloud as the corresponding pixel grayscale value. The first method is simple to implement, but it does not conform to the principle of image acquisition. The image is acquired by the camera. In order to facilitate the subsequent feature point extraction and matching, the present invention uses the central projection method to rasterize the point cloud. In addition, it is inevitable that some pixels on the rasterized image do not correspond to the three-dimensional point, which will cause salt and pepper noise on the image. Therefore, the median filtering method is further used in the algorithm to remove the salt and pepper noise. In order to expand the scope of application of the FCFRA algorithm of the present invention, for laser point clouds without color information, layered coloring can be performed according to the elevation coordinate values of the point cloud to obtain color values reflecting the height change information of the point cloud, and a two-dimensional raster image is generated.
[0067] After obtaining the rasterized image, the next step is to extract and match the feature points. Since the sensors, time, height, etc. of the point clouds to be registered may be different, the differences between the point clouds may be large, and the corresponding rasterized images will also have large differences. Therefore, the extraction and matching of feature points must be able to adapt to such large differences in images. After comprehensive analysis and comparison, the FCFRA algorithm uses SuperPoint, which performs well on heterogeneous images, to extract raster feature points, and uses SuperGlue for feature point matching.
[0068] After completing the extraction and matching of the rasterized image feature points, the corresponding three-dimensional point cloud feature points are further used to inversely calculate. With these corresponding three-dimensional point clouds, the FCFRA algorithm uses the singular value decomposition (SVD) method and combines the random sample consensus algorithm (RANSAC) to complete the rotation and translation matrix calculation between the point clouds to be matched, and finally achieves high-precision registration. The main steps of the FCFRA algorithm are as follows:
[0069] (1) Downsampling the airborne 3D laser point cloud data to be registered, performing 2D rasterization processing on the downsampled 3D laser point cloud using central projection to obtain a rasterized image (i.e., a raster image), and then performing median filtering on this basis to remove salt and pepper noise on the raster image;
[0070] (2) Use the SuperPoint algorithm to extract feature points from the raster image after removing salt and pepper noise;
[0071] (3) Use SuperGlue algorithm to match raster image feature points;
[0072] (4) Using the matched 2D image feature points, the matched 3D point cloud feature points are calculated inversely. The inverse calculation method is to record the correspondence between the point cloud and the grid pixels when generating the grid image. After obtaining the 2D image feature points, the 3D feature points can be obtained by using the recorded correspondence.
[0073] (5) Using the matched 3D point cloud feature points, the singular value decomposition method SVD combined with the random sampling consensus algorithm RANSAC is used to calculate the optimal transformation parameters;
[0074] (6) Using the transformation parameters, perform transformation operations on one of the point cloud data to complete the registration of the three-dimensional point cloud data.
[0075] The FCFRA algorithm flow is as follows Figure 1 shown.
[0076] The algorithm specifically includes:
[0077] 1. Point cloud raster image generation
[0078] The amount of airborne LiDAR point cloud data for outdoor scenes is huge, and it is difficult to directly perform point cloud registration. Therefore, the present invention converts the massive point cloud into a two-dimensional grid map, and then uses an excellent image feature point matching algorithm to indirectly perform three-dimensional point cloud registration. The idea of generating the grid map is to convert the three-dimensional coordinates of the point cloud into the two-dimensional pixel coordinates of the grid map based on the central projection principle and the camera imaging model shown in formula (1).
[0079]
[0080] In formula (1), u, v represent the pixel coordinates of the corresponding image point on the grid map, s is the scale factor, K represents the camera intrinsic parameter matrix, R represents the rotation matrix between the camera coordinate system and the world coordinate system, T represents the translation matrix between the sampled point cloud, t represents the translation vector, [X w ,Y w ,Z w ] represents the three-dimensional coordinates of the point cloud, f represents the camera principal distance, d x and d y They represent the actual physical sizes of a single pixel on the camera sensor in the horizontal u and vertical v directions respectively. u0 and v0 are the coordinates of the center point of the image coordinate system in the pixel coordinate system.
[0081] During the raster image generation process, f, d x d yThe three parameters can be selected from a set of conventional camera parameters on the market. It is only necessary to keep the point cloud data to be registered using the same parameters. In subsequent experiments of the present invention, f, d x d y The uniform values are 35mm, 5um, and 5um, and all have achieved good results. The scale factor s is determined as needed. The u0 and v0 in the camera internal parameters are determined according to the size of the generated grid image, and the size of the generated grid image is determined according to the point cloud. The principle of determination is to make one 3D point on the generated grid image correspond to one pixel as much as possible, and the number of blank pixels is as small as possible. Specifically, the size of the grid image is calculated using formulas (2) and (3). num is the number of points in the point cloud, h is the height of the image, and w is the width of the image, in pixels.
[0082]
[0083] w=k×h (3)
[0084] In formula (2) and formula (3), s is a scale factor, which is taken as 2 in subsequent experiments of the present invention, and k is the aspect ratio of the image, which can be calculated by formula (4).
[0085]
[0086] In formula (4), X max , X min Respectively represent the maximum and minimum coordinates of the point cloud in the X direction, and Y max , Y min They represent the maximum and minimum coordinates in the Y direction of the point cloud respectively.
[0087] The modulus of the translation vector t, i.e., the distance from the projection center to the point cloud, is calculated based on the range of the point cloud, the size of the image, and the camera principal distance. The direction of the translation vector t and the rotation matrix R are set according to the general acquisition angle of the airborne LiDAR. The airborne LiDAR is generally collected perpendicular to the ground, so the direction of the translation vector t and the rotation matrix R can be easily given. According to the above principles, when the present invention generates a raster image, R uses the unit matrix E, and the translation vector t is calculated using formula (5)
[0088]
[0089] The meanings of the parameters in formula (5) are the same as those in the previous formula, where the Z coordinate of t is determined based on the ratio of the camera main distance f and the image height when the central projection imaging is equal to the ratio of the flight altitude and the distance in the Y direction of the point cloud.
[0090] When generating a raster image, the corresponding pixel coordinates of each point in the three-dimensional point cloud are calculated according to formula (1), and the color value of the point cloud is assigned to the corresponding pixel point. Figure 2 (a) is a raster pattern generated from a point cloud. In this way, some pixels in the generated raster will inevitably have no corresponding 3D points, resulting in salt and pepper noise on the image. Therefore, it is necessary to further use the median filtering method to remove the salt and pepper noise, such as Figure 2 When the raster image is generated, it is also necessary to record the one-to-one correspondence between each image point and the 3D point in the point cloud, so as to facilitate the subsequent inverse calculation of the corresponding 3D feature points from the 2D image feature points.
[0091] The generation of two-dimensional grid images is a key link in the entire algorithm. However, due to the huge amount of point cloud data, it will take a lot of time to generate a two-dimensional grid image from a three-dimensional point cloud. In order to reduce the time consumption of this link, the present invention uses point cloud downsampling to perform thinning processing without affecting the matching effect. The specific downsampling method uses a more efficient uniform downsampling method to sample the data at equal intervals according to a fixed number of steps. At the same time, when the grid image is specifically generated, the present invention uses Python for encoding, making full use of the high-dimensional array processing capabilities of Numpy, moving the point cloud data calculation into the Numpy array and realizing its vectorized high-speed processing, thereby replacing the very time-consuming loop operation and greatly improving the speed of grid image generation.
[0092] For laser point cloud data without color information, the present invention adopts a layered coloring method to calculate the color value of the corresponding point cloud according to the elevation value of the point cloud. The specific layered coloring formulas adopted are shown in equations (6)-(8).
[0093]
[0094] In equations (6)-(8), f = 255 × (Z-z_Min) / (z_Max-z_Min), L is a constant of 256, R, G, B are the red, green, and blue components of the point cloud color, respectively, with values ranging from 0 to 255, Z is the elevation value of the point cloud, and z_Max and z_Min are the maximum and minimum elevation values of all point clouds, respectively.
[0095] 2 Feature point extraction and matching
[0096] After the three-dimensional point cloud is transformed into a two-dimensional grid image, the next step is to use the two-dimensional image feature point matching method to find the corresponding feature points between the grid images of different point clouds for the calculation of the transformation matrix in the subsequent point cloud registration. There are many methods for image feature point extraction and matching. After comprehensive comparison, the present invention uses SuperPoint to extract grid feature points and SuperGlue to match feature points.
[0097] 2.1 SuperPoint feature point extraction principle
[0098] The SuperPoint algorithm uses a self-supervised network framework to detect pixel-level feature points and their descriptors on the original image. The network architecture is as follows: Figure 3 As shown, a full-size image is input, an encoder is used to first reduce the dimension of the image, and then two decoders are used. The output of the decoder in the upper half of the figure is a feature point, and the output of the decoder in the lower half is a feature descriptor.
[0099] In the entire network, SuperPoint uses a network structure similar to VGG in the encoder to reduce the dimension of the image and extract features to reduce the computational complexity of the subsequent network. The two decoder networks are the feature point decoder and the feature descriptor decoder. The feature point decoder calculates a probability for each pixel of the image, which represents the possibility of it being a feature point. In order to solve the problem of excessive computational complexity of general decoding operations, SuperPoint uses the sub-pixel convolution method to reduce the computational complexity. The feature descriptor decoder first learns a semi-dense descriptor, then performs a bicubic interpolation algorithm to obtain a complete descriptor, and finally uses L2-normalizes to obtain a description of unit length.
[0100] 2.2 SuperGlue feature point matching principle
[0101] SuperGlue is a network algorithm that can simultaneously perform feature matching and filter out outliers. It achieves matching by solving the differentiable optimal transmission problem, and the loss function is constructed by CNN. SuperGlue proposes a flexible content aggregation mechanism based on the attention mechanism, which enables SuperGlue to perform feature matching while perceiving potential 3D scenes. SuperGlue algorithm model (see Figure 4 ) consists of two main modules: attention mechanism CNN and optimal matching layer.
[0102] Among them, the attention CNN structure obtains local feature points through convolution, and uses the key point encoder to map its key point coordinates p and the visual descriptor d obtained by the SuperPoint method into a vector, and then enhances back and forth through self-attention and cross-attention (repeated L = 7 times) to obtain a more representative descriptor f i (dimension is 256).
[0103]
[0104] Where W and b are weights and biases. Image B is similar to image A, and image A and image B are two images to be matched.
[0105] In the optimal matching layer, images A and B have M and N local feature points respectively. For the optimal matching layer, first construct the allocation matrix P∈[0,1] M×N And satisfy:
[0106]
[0107] For all possible matching points, the descriptor f obtained by the inner product calculation of the GNN part i A , The similarity of , then the matching pair score is expressed as:
[0108]
[0109] As with standard graph matching, the assignment matrix P can be calculated by the score matrix S∈R M×N Get and maximize the total score∑ i,j S i,j P i,j . Add a layer of dusbin channel in the last row or last column of the score matrix S to obtain the augmented matrix In order to filter out the wrong key points. The maximum total score ∑ i,j S i,j P i,j It is regarded as an optimal transmission problem. Entropy regularization is added to limit the complexity of the solution to the optimal transmission problem. In order to better play the role of entropy regularization, the Sinkhorn algorithm (iteration T = 100 times) is used to solve the optimal transmission problem with entropy regularization.
[0110] 3 Point cloud registration calculation
[0111] After using SuperPoint and SuperGlue to match the feature points of the point cloud grid, the corresponding feature points between the two point clouds, i.e., the conjugate points, can be inversely calculated. Further using the conjugate feature points, the present invention uses the SVD algorithm combined with the RANSAC method to complete the registration parameter solution. The specific solution steps are as follows:
[0112] (1) Sample n pairs of conjugate points from conjugate point clouds A and B respectively;
[0113] (2) Based on the sampling conjugate points, the transformation parameters are calculated using the SVD method. The specific calculation formulas are as follows (12)-(15):
[0114]
[0115] [U,S,V]=SVD(H) (13)
[0116] R=VU T (14)
[0117] T=-R*centroid A +centroid B (15)
[0118] in, Represents the coordinates of the sampled points in the conjugate point cloud A and B, centroid A and centroid B represents the mean center of the sampling points, H represents the covariance matrix of the sampling points, formula (13) represents the singular value decomposition of the covariance matrix to obtain the U, S, V matrices, R represents the rotation matrix between the sampling point clouds, and T represents the translation matrix between the sampling point clouds.
[0119] (3) Use the transformation parameters to determine whether other points are interior points. The specific determination method is to use formula (16) to calculate the non-sampling points in the conjugate point cloud A. After transformation Point, if The corresponding point in the conjugate point cloud B If the distance is less than a given threshold, the conjugate point j is considered to be an interior point, otherwise it is not an interior point.
[0120]
[0121] (4) If the proportion of interior points to all conjugate points reaches a specified value, the algorithm terminates and all interior points are used to optimize the transformation parameters;
[0122] (5) If the inlier point ratio does not reach the specified value, continue to resample at the conjugate point and execute steps (1) to (4) again until the inlier point ratio meets the requirement or the operation reaches the specified number of iterations. At this time, the transformation parameter corresponding to the highest inlier point ratio is used as the final transformation parameter.
[0123] 4 Experiment and analysis
[0124] In order to verify the effectiveness of the algorithm of the present invention, the 3D laser point cloud carried by unmanned aerial vehicle is used as the test data, and the evaluation index is established. The comparative test is carried out with several commonly used point cloud registration algorithms, including ICP, Colored_ICP, NDT [Biber P, Straber W. The normal distributions transform: a new approach to laser scan matching [C] / / Proceedings of 2003 IEEE / RSJ International Conference on Intelligent Robots and Systems, IEEE, 2003: 2743-2748.], PFH, FPFH, 3Dsc, LCD [Pham QH, Uy MA, Hua BS, et al. LCD: Learned Cross-Domain Descriptors for 2D-3D Matching. Proceedings of the AAAI Conference on ArtificialIntelligence.2020,34:11856-11864.] algorithm, among which the Colored_ICP algorithm is a variant of ICP, which also considers the color information of the point cloud in the registration; PFH, FPFH, 3Dsc, and LCD algorithms all extract local features through feature extraction algorithms, then find conjugate features, and perform registration calculations. The difference is that LCD uses deep learning methods for feature description. The hardware equipment used in the experiment is a laptop computer, the graphics card is NVIDIA GeForce RTX 3080Laptop, the video memory is 16G; the computer memory is 64G; the processor is Intel Core I7-11800H 8 cores and 16 threads; the hard disk is a 2T solid-state drive; the operating system is Window 10; the programming languages used in the experiment are Python and C++, and the deep learning network involved is implemented based on Pytorch.
[0125] 4.1 Test data
[0126] In order to verify the effectiveness of the algorithm, the embodiment of the present invention collects data from the same area at different heights and at different times to obtain test data. UAVs are used to obtain LiDAR data. The data acquisition platform is DJI M300, and the sensor is DJI Zenmuse L1. The device can obtain LiDAR point clouds containing color information. The test data acquisition area is the urban area of Zhengzhou City, Henan Province, and two acquisitions are performed at 60m and 80m altitudes respectively. Figure 5 and Figure 6Figure 1 is the effect diagram of the two collections, and Table 1 is the corresponding data information.
[0127] Table 1 Collection data information table
[0128]
[0129] 4.2 Evaluation Metrics
[0130] Since there will be non-overlapping areas between the data to be registered in the actual scene, in order to verify the effectiveness of the proposed algorithm, the registration effect diagram, registration coincidence degree (RCD), root mean square error (RMSE), mean absolute error (MAE), and registration time are used as evaluation criteria. RCD mainly refers to the ratio of the number of point clouds in the overlapping area to the number of point clouds to be matched after the registration of two point clouds. The specific formula of RCD is as follows:
[0131]
[0132] In formula (17), num C Indicates the number of point clouds in the overlapping area. You can use the point cloud in the registered data 1 as a reference to find the nearest point in the registered data 2. When the distance to the nearest point is less than a given threshold, the pair of points is considered to be overlapping points. C The number is increased by 1, and so on, to complete the number C Statistics, n is the number of point clouds of data 1.
[0133] The specific calculation formula of RMSE is as follows:
[0134]
[0135] In formula (18), n represents the number of point clouds in data 1, X i represents the coordinates of point i in data 1 after registration, Y i represents the coordinates of the nearest neighbor of point i in data 2 found with data 1 as reference after registration, |X i -Y i | represents the distance between the corresponding coincident points i. The smaller the RMSE, the better the registration result. The specific calculation formula of MAE is as follows:
[0136]
[0137] The meanings of the parameters in formula (19) are the same as those in formula (18).
[0138] 4.3 Results and Analysis
[0139] 4.3.1 Registration effect diagram
[0140] Figure 7 The process results of the registration, such as the raster image generation result, image matching result, corresponding 3D point cloud (conjugate point), and conjugate point cloud registration result, are shown. During the registration process, the data thinning interval is 10, that is, 1 point is retained for every 10 points. In the raster image generation result, the image size generated by data 1 is 583×2839, and the image size generated by data 2 is 499×2194. In the image matching result, the number of feature points, i.e., conjugate points, that are successfully matched is 2860 pairs, of which the number of feature points extracted from the raster image of data 1 is 7038, and the number of feature points extracted from the raster image of data 2 is 4678. In the corresponding 3D point cloud, they are the conjugate point clouds of data 1 and data 2 obtained by inverse calculation of the raster image, and the number of points in the point cloud and the number of matching points are the same, both 2080. The point cloud registration result shows the display effect of the conjugate points together before and after the registration, respectively. The left picture shows the effect of displaying together before the registration, and the right picture shows the effect of displaying together after the registration. In the final registration calculation, the number of internal points involved in the calculation is 708.
[0141] Figure 8 and Fig. 9 This is the display effect of the original point cloud before and after registration. Figure 8 This is the effect shown together before registration. Fig. 9 The result of displaying them together after registration. From the process results of point cloud registration and the final registration effect, it can be seen that the FCFRA algorithm of the present invention can effectively complete the registration between two point clouds.
[0142] Fig.10 The process result diagram of point cloud registration is that the color information of the point cloud is not used, but the layered coloring method is used to assign colors to the point cloud and then perform point cloud registration. From the results, it can be seen that the layered coloring method can still achieve effective point cloud registration, which proves that the FCFRA algorithm of the present invention is also applicable to laser point cloud data registration without color information.
[0143] 4.3.2 Comparison of evaluation index results
[0144] In order to further verify the effectiveness of the FCFRA algorithm of the present invention, the results of each evaluation index are analyzed in detail below.
[0145] (1) Experiment on the impact of point cloud downsampling
[0146] Before generating the point cloud raster map in the FCFRA algorithm, downsampling the point cloud is a critical step, which will affect the registration time and registration accuracy. In order to verify the effect of point cloud downsampling on registration, this paper compares the point cloud registration effects under different sampling step sizes. The specific results are as follows: Fig.11 As shown, Fig.11 (a) shows the number of point clouds after downsampling of two point cloud data at different sampling steps. Fig.11 (b) is the registration coincidence degree corresponding to different sampling conditions. Fig.11 (c) is the error distribution of RMSE and MAE under different downsampling conditions. Fig.11 (d) is the registration time under different downsampling conditions. Fig.11 Middle (b), Fig.11 Middle (c), Fig.11 The horizontal axis in (d) uses data 1 with a larger number of downsampled point clouds as a reference.
[0147] from Fig.11 From the experimental results, it can be seen that when the number of point clouds decreases, the registration coincidence will slowly decrease, and the RMSE and MAE error values will also slowly increase. The change trends of RMSE and MAE are consistent, with an average RMSE of 1.25m and an average MAE of 0.51m. In terms of registration time, when the point cloud data drops from more than 6 million points to about 1 million points, that is, when the corresponding adoption interval is 60, the registration time drops rapidly from about 6s to less than 1s. When the number of point clouds further drops from about 1 million to about 300,000, the registration time drops to about 0.3s, and the downward trend slows down. Combined with the principle of the FCFRA algorithm of the present invention, and further analysis of the time proportion of each step in the registration time, it is found that the generation of the raster map in the registration occupies more than half of the entire registration time. When the number of point clouds decreases, the raster map generation time also drops rapidly. Another major time consumption is mainly in the feature point registration, which is closely related to the size of the generated point cloud raster map. The smaller the image, the less time it takes to register. However, the smaller the image, the lower the registration accuracy. Based on the experimental results, when the number of point clouds is around 1 million, it can ensure both less time and better registration accuracy. Therefore, when the number of point clouds is large, the number of point clouds can be thinned to around 1 million by downsampling to ensure better registration efficiency and effect.
[0148] (2) Evaluation index results
[0149] Table 2 shows the registration results of experimental data using ICP, Colored_ICP, NDT, PFH, FPFH, 3Dsc, and LCD methods. The point cloud downsampling step size is uniformly 60. At this time, the number of point clouds after data 1 downsampling is 1055416, and the number of point clouds after data 2 downsampling is 704580.
[0150] Table 2 Evaluation index calculation results
[0151]
[0152]
[0153] The experimental results show that the FCFRA algorithm of the present invention has achieved the best results in the four indicators of RCD, RMSE, MAE, and registration time. Combined with the registration effect diagram and the registration coincidence, it is further found that the ICP, Colored_ICP, NDT, PFH, FPFH, 3Dsc, and LCD methods are not ideal in the registration of large-scale point clouds in large areas, and the registration coincidence is relatively low. Among them, the LCD method fails to register directly, and the Colored_ICP registration effect is relatively good, but the time used is much longer than the FCFRA algorithm of the present invention. The registration coincidence of the FCFRA algorithm of the present invention reaches 97.38%, which is close to the overlap rate of the actual coverage area of data 1 and data 2. Its registration time is only 0.85s, which is much faster than other methods. In terms of registration error, RMSE and MAE are 1.20m and 0.42m respectively. If only the RMSE and MAE of the repeated area are calculated, the values of the two are 0.31m and 0.27m respectively.
[0154] In addition, in theory, the algorithm that completes the registration by extracting point cloud feature points belongs to the coarse registration algorithm. The FCFRA algorithm of the present invention belongs to this category. If precise registration is required, it is necessary to further perform precise registration through algorithms such as ICP or NDT. However, for large-area outdoor point cloud data, the data collection range is different, and the collection height and collection path are different. This leads to the data to be registered. Even in the overlapping area, the corresponding points cannot completely overlap, and the registration error will have an extreme value. When it approaches this extreme value, even if the method is optimized, the registration error will not be reduced much. The present invention uses two sets of data acquired at different times during the experiment, so it is impossible to perform the registration test after translation and rotation transformation of one data, and the registration error can be very small. In order to verify the above analysis, on the basis of the registration results of the FCFRA algorithm, the ICP algorithm is further used to perform the registration test, which is recorded as ICP2. The specific results are shown in Table 3. In addition, for the convenience of comparison, Table 3 also lists the registration results of the ICP algorithm and the FCFRA algorithm.
[0155] Table 3 Calculation results of evaluation indexes of ICP precise registration algorithm
[0156]
[0157] It can be seen from the results in Table 3 that using the registration result of the FCFRA algorithm as the initial value can indeed improve the accuracy of the ICP algorithm. However, compared with the FCFRA algorithm of the present invention, the accuracy has not been improved but has decreased. The main reason for this is that the FCFRA algorithm mainly involves conjugate points in the transformation matrix calculation, while the ICP algorithm involves as many points as possible in the overlapping area of the two point clouds to be registered. The point clouds processed by the present invention are all point clouds collected outdoors. It is inevitable that there will be many noise points in these point clouds. There will also be many points in the overlapping area that are in data 1 but not in data 2, or in data 2 but not in data 1. This leads to the situation that the ICP algorithm not only fails to improve the accuracy, but reduces the accuracy.
[0158] Therefore, the FCFRA algorithm of the present invention has better performance in the registration of massive point clouds in large areas. In addition, since the FCFRA algorithm of the present invention is mainly based on conjugate point extraction based on raster images, it has a good tolerance for the number of point clouds. Combined with downsampling, it can be well applied to the registration of massive point cloud data in a large range.
[0159] 4.3.3 Algorithm adaptability verification
[0160] In order to further verify the adaptability of the FCFRA algorithm of the present invention to different data, the FCFRA algorithm was used to test another 4 groups of data. The first group is the LiDAR point cloud data obtained on the snow, and the acquisition device is also the DJI Zenmuse L1 laser radar; the second group is the point cloud data of winter snow and summer green land generated by image data using Agisoft Metashape software, where data 1 is obtained using a Sony DSC-RX1RM2 camera, and data 2 is obtained using a DJI EP800 camera; the third group uses the Vaihingen dataset published by the International Association for Photogrammetry and Remote Sensing, and the LiDAR point cloud data of two adjacent strips are selected in the experiment. The point cloud does not have color information, and the overlap rate between the strips is about 30%; the fourth group uses the Hessigheim 3D dataset published by the University of Stuttgart in Germany, and the LiDAR point cloud data of the same area obtained in two different periods in March 2016 and March 2019 are selected in the experiment. During the experiment, the point cloud coordinates of the above data were centroided to fully verify the feasibility of the algorithm. Fig.12 This is the relevant information of these 4 sets of data.
[0161] Figure 13 to Figure 16 The following is the registration effect diagram of these four sets of data, where the red one is data 1 and the green one is data 2. From the registration effect diagram, we can see that this algorithm has good adaptability to the airborne laser point cloud data obtained from snow and the point cloud data generated by images of winter snow and summer green land, and can complete the registration of these two types of data. It can also achieve good results for public data sets.
[0162] Table 4 Evaluation index calculation results
[0163]
[0164] Table 4 shows the evaluation index results of the registration of these four types of data. From the results, this method can quickly complete the registration of these two types of data, taking about 1 second. The RMSE and MAE values of all points in the first three groups of data in Table 4 are relatively large. The reason for this is that in the calculation of this evaluation index, the distance of the corresponding points is calculated by finding the nearest neighbor points in another group of data. However, in the experiment, the range of the data to be registered in the first three groups of data is quite different. In addition, the errors in the acquisition height, acquisition path, and image generation point cloud process, etc., make it impossible for the data to be registered to completely overlap. In this way, the nearest neighbor points found may not be overlapping points, and the two are far apart, which leads to relatively large RMSE and MAE values. The range of the data to be registered in the fourth group of data is basically the same, and the RMSE and MAE values are relatively small. For this reason, the present invention eliminates corresponding points that are far apart, that is, uses overlapping points to calculate RMSE and MAE. The results are shown in Table 4. The RMSE and MAE values of the overlapping points are basically maintained at around 0.5m. Considering that the points of the data to be registered themselves are not strictly overlapping, such registration accuracy is reasonable.
[0165] The registration of large-area outdoor laser point cloud data is the basis for subsequent applications. The FCFRA registration algorithm proposed in this paper makes full use of the color information of the point cloud, adopts the point cloud rasterization strategy, and then uses the excellent characteristics of SuperPoint and SuperGlue deep learning feature extraction and matching methods to complete the search for point cloud conjugate points and finally complete the registration of different point clouds. Experiments on large-area outdoor massive airborne point cloud data show that:
[0166] (1) The FCFRA point cloud registration algorithm proposed in the present invention has a clear principle and can be effectively used for the registration of massive airborne LiDAR point cloud data in large outdoor areas. It has high speed and accuracy, and the algorithm has good adaptability.
[0167] (2) The registration time of the FCFRA algorithm of the present invention is closely related to the size of the generated grid image. The experimental results show that when the number of point clouds is large, the number of point clouds can be thinned to about 1 million by downsampling. This can ensure a better registration effect and obtain better registration efficiency. The entire registration time can reach about 1 second.
[0168] (3) The comparison with ICP, Colored_ICP, NDT, PFH, FPFH, 3Dsc and LCD shows that the FCFRA algorithm of the present invention achieves the best results in terms of RCD, RMSE, MAE and registration time when processing large-area massive point cloud registration. In addition, the FCFRA algorithm is mainly based on the extraction of conjugate points based on the grid map, so it has a good tolerance for the number of point clouds. Combined with downsampling, it can be well applied to the registration of large-scale massive point cloud data.
[0169] (4) The test results of airborne laser point cloud data acquired from snow, point cloud data generated from images of winter snow and summer green land, and public data sets show that the FCFRA algorithm of the present invention can well complete point cloud registration even for data with large differences. It can also complete registration for point clouds without color information, and the algorithm has good adaptability.
[0170] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. A fast registration algorithm for massive airborne LiDAR point clouds integrating color features, It is characterized in that include: Step 1: Downsample the airborne 3D laser point cloud data to be registered, perform 2D rasterization on the downsampled 3D laser point cloud using central projection to obtain a raster image, perform median filtering on the raster image, and remove salt and pepper noise on the raster image; Step 2: Use the SuperPoint algorithm to extract feature points from the raster image after removing salt and pepper noise; Step 3: Use SuperGlue algorithm to match raster image feature points; Step 4: Use the matched 2D image feature points to inversely calculate the matched 3D point cloud feature points; Step 5: Using the matched 3D point cloud feature points, the singular value decomposition method combined with the random sampling consensus algorithm is used to calculate the optimal transformation parameters between the laser point cloud data to be registered; Step 6: Use the transformation parameters to perform transformation operations on the point cloud data to complete the registration of the 3D laser point cloud data.
2. According to claim 1, the rapid registration algorithm for massive airborne LiDAR point clouds integrating color features, It is characterized in that In step 1, the downsampled three-dimensional point cloud to be registered is subjected to two-dimensional rasterization processing in the following manner: Among them, u, v represent the pixel coordinates of the corresponding image point on the grid map, s is the scale factor, K represents the camera internal parameter matrix, R represents the rotation matrix between the camera coordinate system and the world coordinate system, T represents the translation matrix between the sampled point cloud, t represents the translation vector, [X w ,Y w ,Z w ] represents the three-dimensional coordinates of the point cloud, f represents the camera principal distance, d x and d y They represent the actual physical size of a single pixel in the horizontal u and vertical v directions on the camera sensor, respectively. 0 and v 0 is the coordinate of the center point of the image coordinate system in the pixel coordinate system; When generating a raster image, the corresponding pixel coordinates of each point in the three-dimensional laser point cloud are calculated according to the above formula, and the color value of the corresponding three-dimensional laser point cloud is assigned to the corresponding pixel point.
3. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 2, It is characterized in that The translation vector is determined according to the following formula: Among them, X max , X min Respectively represent the maximum and minimum coordinates of the point cloud in the X direction, and Y max , Y min They represent the maximum and minimum coordinates of the point cloud in the Y direction, and h is the height of the raster image.
4. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 1, It is characterized in that In step 1, the size of the raster image is determined according to the following formula: w=k×h, Among them, h is the height of the raster image, w is the width of the raster image in pixels, num is the number of points in the point cloud, s is the scale factor, and k is the aspect ratio of the raster image.
5. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 4, It is characterized in that The aspect ratio of the raster image is determined according to the following formula: Among them, X max , X min Respectively represent the maximum and minimum coordinates of the point cloud in the X direction, and Y max , Y min They represent the maximum and minimum coordinates in the Y direction of the point cloud respectively.
6. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 1, It is characterized in that In step 1, when the raster image is generated, the one-to-one correspondence between each image point and the three-dimensional point in the laser point cloud needs to be recorded to facilitate the subsequent inverse calculation of the corresponding three-dimensional point cloud feature points from the two-dimensional image feature points.
7. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 1, It is characterized in that In step 1, when the raster image is generated, Python is used for coding, the point cloud data operation is moved into the Numpy array and its vectorized processing is realized.
8. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 1, It is characterized in that In step 1, for laser point cloud data without color information, a layered coloring method is used to calculate the color value of the corresponding point cloud according to the elevation value of the point cloud: Where f = 255 × (Z-z_Min) / (z_Max-z_Min), L is a constant, R, G, B are the red, green, and blue components of the point cloud color, Z is the elevation value of the point cloud, and z_Max and z_Min are the maximum and minimum elevation values of all point clouds, respectively.
9. The rapid registration algorithm for massive airborne LiDAR point clouds integrating color features according to claim 1, It is characterized in that The step 5 comprises: Step 5.1: Sample n pairs of conjugate points from the feature points of the matched 3D point clouds A and B; Step 5.2: Based on the sampled conjugate points, the transformation parameters are calculated using the singular value decomposition method: [U,S,V]=SVD(H), R=VU T , T=-R*centroid A +ecntroid B , in, Respectively represent the coordinates of the sampling points in the conjugate point cloud A, B, ecntroid A and centroid B Respectively represent the average center of the sampling points in A and B, H represents the covariance matrix of the sampling points, U, S, V represent the matrix obtained by performing singular value decomposition on the covariance matrix, R represents the rotation matrix between the sampling point clouds, and T represents the translation matrix between the sampling point clouds; Step 5.3: Calculate the non-sampling points in A according to the following formula: After transformation Point, if The corresponding point in B If the distance is less than a given threshold, the conjugate point j is considered to be an interior point, otherwise it is not an interior point; Step 5.4: If the proportion of interior points to all conjugate points reaches the specified value, the algorithm terminates and all interior points are used to optimize the transformation parameters; Step 5.5: If the inlier ratio does not reach the specified value, continue to execute steps 5.1 to 5.4 until the inlier ratio meets the requirements or the operation reaches the specified number of iterations. At this time, the transformation parameters corresponding to the highest inlier ratio are used as the optimal transformation parameters.
Citation Information
Patent Citations
Three-dimensional point cloud registration method, device and system and storage medium
CN115841507A
Feature point detection method for two-dimensional and three-dimensional cross-domain matching
CN117409222A
Cited By
Training method of image generation algorithm, storage medium, lidar and vehicle
TWI929680B
