Coarse-to-fine point cloud registration method

Abstract

The invention discloses a coarse-to-fine point cloud registration method, and the method comprises the steps: carrying out the preprocessing of a to-be-registered source point cloud and a to-be-registered target point cloud, and obtaining a first point cloud and a second point cloud after preprocessing; performing coarse registration on the first point cloud and the second point to obtain an initial matching coefficient; performing iterative solution on the target function and the threshold function based on the initial matching coefficient, the source point cloud and the target point cloud until an optimal matching coefficient meeting a convergence threshold condition is obtained; and carrying out rotation and translation operation on the source point cloud based on the optimal matching coefficient, converting the source point cloud into a coordinate system where the target point cloud is located, enabling the source point cloud and the target point cloud to be aligned in the same coordinate system, and completing registration operation. According to the method, the point cloud is preprocessed and coarsely registered, and then the optimal matching coefficient is solved step by step through iteration, so that the defects of a traditional method can be effectively overcome, and the point cloud registration efficiency and precision are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud data processing, and in particular relates to a coarse-to-fine point cloud registration method. Background Art

[0002] In the field of water conservancy projects, the collection and processing of point cloud data is crucial for monitoring hydraulic structures. Advances in 3D laser scanning technology enable the rapid and precise acquisition of high-precision 3D point cloud data between hydraulic structure stations. This point cloud data comprehensively captures detailed changes in the dam structure and surrounding terrain, providing a key basis for safety assessments, maintenance decisions, and the long-term stable operation of hydraulic structures.

[0003] Traditional point cloud registration methods generally employ the iterative closest point (ICP) algorithm or a simplified variant thereof. Traditional ICP algorithms typically directly perform fine registration of the source and target point clouds during point cloud registration. However, this method has numerous drawbacks. For one thing, when the initial positions of the source and target point clouds differ significantly, the ICP algorithm can easily become trapped in a local optimum, resulting in inaccurate registration results. This is because the ICP algorithm relies on a good initial position estimate. If the initial positions deviate significantly, the iterative process may fail to converge to the global optimum, resulting in large rotational and translational errors in the registered point clouds. Furthermore, traditional ICP algorithms suffer from low computational efficiency when processing large point cloud data. Because the algorithm requires computation and optimization of all point cloud data in each iteration, computation time increases exponentially with increasing point cloud data volume, making it difficult to meet the requirements of fast registration in practical applications. Furthermore, traditional ICP algorithms are sensitive to noise and outliers in the point cloud, which can interfere with the algorithm's iterations and further impact registration accuracy. Therefore, a more efficient and accurate registration method is urgently needed to break through these technical bottlenecks. Summary of the Invention

[0004] The purpose of the present invention is to address the deficiencies of the above-mentioned background technology and to provide a coarse-to-fine point cloud registration method, which achieves efficient and accurate point cloud registration through multi-stage processing.

[0005] The technical solution adopted by the present invention is: a coarse-to-fine point cloud registration method, comprising the following steps:

[0006] Step 1: Preprocessing the source point cloud and the target point cloud to be registered respectively to obtain a preprocessed first point cloud and a second point cloud;

[0007] Step 2: Roughly align the first point cloud and the second point to obtain the initial matching coefficient; Step 3: Iteratively solve the objective function and threshold function based on the initial matching coefficient, the source point cloud and the target point cloud until the optimal matching coefficient that meets the convergence threshold condition is obtained; Step 4: Rotate and translate the source point cloud based on the optimal matching coefficient, convert the source point cloud to the coordinate system of the target point cloud, align the source point cloud and the target point cloud in the same coordinate system, and complete the registration operation.

[0008] Furthermore, the pre-processing includes sequentially performing memory mapping, voxel grid downsampling and noise filtering.

[0009] Furthermore, the coarse registration of the first point cloud and the second point to obtain the initial matching coefficient includes:

[0010] Constructing a first data structure using a KD tree on the first point cloud, and performing feature extraction on the first data structure to obtain a first feature point cloud;

[0011] Constructing the second point cloud using a KD tree to obtain a second data structure, and performing feature extraction on the second data structure to obtain a second feature point cloud;

[0012] The RANSAC algorithm is used to perform robust feature matching on the first feature point cloud and the second feature point cloud to obtain the initial matching coefficient.

[0013] Furthermore, the objective function is E(R, t) = ∑((ni T ·(R·pi+t-qi)) 2 )

[0014] Among them, E(R, t) is the sum of the vertical distances of all registered points to the corresponding plane, the registered point is the point in the source point cloud projected to the point in the target point cloud, the plane is the differential plane composed of several points near the registered point, R is the rotation matrix, t is the translation vector, p i and q i are the coordinates of the corresponding point i in the source point cloud and the target point cloud, n i is the normal vector at point i in the target point cloud, and ^T represents the vector transpose operation.

[0015] Furthermore, the iterative solution of the objective function must satisfy the following requirement: E(R, t) calculated in the current iterative calculation is less than E(R, t) calculated in the previous iterative calculation.

[0016] Furthermore, the threshold function is ε=max(ε min ,ε0·e -k·iter , where ε is the convergence threshold of the current iteration, ε minis the minimum error threshold allowed in the registration process, ε0 is the initial threshold at the beginning of registration, k is the attenuation coefficient that controls the rate of decrease of the threshold, and iter is the current number of iterations.

[0017] Furthermore, the convergence threshold condition is satisfied when ε<ε min , where ε is the convergence threshold of the current iteration, ε min The minimum error threshold allowed in the registration process.

[0018] Furthermore, if the number of points in the source point cloud and the target point cloud to be registered is greater than a set value, a multi-scale resolution registration method is used to obtain the optimal matching coefficient.

[0019] Furthermore, the method of obtaining the optimal matching coefficient by adopting multi-scale resolution registration includes the following steps:

[0020] Step a: Construct a resolution increasing sequence: L1, L2, ..., Lm, where Lj = (1 / bj) * Lm, j = 1, 2, ..., m, bj is a decreasing coefficient greater than or equal to 1, and b1>b2> ...>bm, bm = 1, and Lm is the original resolution;

[0021] Step b: Initialize j to 1, perform steps 1 to 3 on the source point cloud and target point cloud under L1, obtain the first matching coefficient, and perform step c;

[0022] Step c, determining the jth matching coefficient as the current matching coefficient;

[0023] Step d: Set j = j + 1. If j < m, proceed to step e; if j = m, proceed to step f.

[0024] Step e: Based on the current matching coefficient, perform step 3 on the source point cloud and target point cloud under Lj to obtain the jth matching coefficient, and return to step c;

[0025] Step f: Based on the current matching coefficient, perform step 3 on the source point cloud and target point cloud under Lm to obtain the optimal matching coefficient.

[0026] Furthermore, it also includes verifying the registration quality based on the converted point cloud. The process is: the root mean square error is calculated using the root mean square error formula. If the root mean square error is less than the set threshold, the registration quality meets the requirements, otherwise the registration quality does not meet the requirements.

[0027] The beneficial effects of the present invention are:

[0028] This paper proposes a comprehensive framework for a coarse-to-fine point cloud registration method, encompassing the key steps from preprocessing, coarse registration, iteratively solving for the optimal matching coefficient, to the final registration operation. This coarse-to-fine strategy effectively improves registration efficiency. Coarse registration quickly yields a rough match, which is then refined based on this result. This avoids the potential for falling into local optimal solutions when directly performing fine registration, thereby improving registration accuracy and reliability. This approach is particularly suitable for point clouds, which may have significant initial positional discrepancies.

[0029] The memory mapping in the preprocessing step of this invention can increase the reading speed of point cloud data, reduce memory usage, and make the processing of large-scale point cloud data more efficient. Voxel grid downsampling can reduce the density of point cloud data, reducing the amount of data and complexity of subsequent calculations, improving computational efficiency, while preserving the key features of the point cloud. Noise filtering can remove noise and outliers in the point cloud data, improving the quality of the point cloud data and thus enhancing the accuracy and efficiency of subsequent registration. These three steps are carried out in sequence, providing a high-quality, easy-to-process data foundation for subsequent registration operations.

[0030] The coarse registration of the present invention uses a KD tree to construct a data structure, enabling rapid feature extraction. The KD tree is an efficient spatial partitioning data structure that can accelerate operations such as nearest neighbor search, improving the efficiency of feature extraction. The RANSAC algorithm is a robust feature matching algorithm that can effectively remove false matches and improve the accuracy of the initial matching coefficient. By constructing KD trees for the first and second point clouds, extracting feature point clouds, and matching them using the RANSAC algorithm, reliable initial matching coefficients can be obtained, providing good initial values for subsequent fine registration, helping to improve the efficiency and accuracy of the entire registration process.

[0031] The objective function of the present invention uses the sum of the vertical distances from the projected points to the corresponding plane as the optimization target, which can more accurately reflect the degree of matching between the source point cloud and the target point cloud. During the registration process, by minimizing this objective function, the points in the source point cloud can be projected as accurately as possible onto the corresponding plane in the target point cloud, thereby improving the accuracy of the registration. The condition for iteratively solving the objective function is also clarified, namely that the E(R, t) calculated by the current iteration is less than the E(R, t) calculated by the previous iteration, ensuring that the objective function value can be reduced with each iteration, so that the registration result gradually approaches the optimal solution. Through this iterative method, the rotation matrix R and the translation vector t can be gradually optimized, improving the accuracy, efficiency and stability of the registration.

[0032] The threshold function of the present invention can dynamically adjust the convergence threshold according to the number of iterations. At the beginning of the registration, the initial threshold ε0 is large, allowing a larger error. As the number of iterations increases, the threshold gradually decreases, and the convergence condition becomes more stringent. This dynamic adjustment method can achieve rapid convergence in the early stage of registration and perform fine adjustments in the later stage. While ensuring the registration accuracy, it improves the registration efficiency and avoids the problem of premature convergence or late convergence.

[0033] The present invention clarifies the judgment criteria for meeting the convergence threshold condition. This condition provides a clear termination criterion for the iterative solution. When this condition is met, it means that the alignment result has reached the preset accuracy requirement and the iteration can be stopped, thereby avoiding unnecessary calculations and improving the alignment efficiency.

[0034] When the number of points in the source point cloud and the target point cloud to be registered is greater than a set value, the present invention adopts a multi-resolution registration method. The multi-resolution registration can start from low resolution and gradually transition to high resolution. Registration at low resolution can quickly obtain a rough matching result, provide an initial value for high-resolution registration, reduce the calculation amount of high-resolution registration, and improve registration efficiency. At the same time, multi-resolution registration can better process large-scale point cloud data and avoid calculation difficulties caused by excessive data volume.

[0035] This method, by constructing a sequence of increasing resolutions and gradually performing registration from low to high resolution, is more in line with the coarse-to-fine registration approach. It clarifies the operations on the preprocessed point cloud at different resolutions, making the multi-resolution registration process clearer and more rational, and improving the efficiency and accuracy of registration. At each resolution, the matching coefficient obtained at the previous resolution is used as the initial value, and an iterative solution is performed to obtain the matching coefficient at the current resolution, ultimately obtaining the optimal matching coefficient at high resolution. This step-by-step approach effectively improves the efficiency and accuracy of registration while ensuring the reliability of the registration results.

[0036] This method verifies registration quality based on the converted point cloud. Using the RMS error formula, the RMS error is calculated and compared with a set threshold, providing a visual indicator of the accuracy of the registration result. By comparing this with the set threshold, it is possible to determine whether the registration quality meets the requirements. This provides a quantitative assessment method for the reliability of the registration result, helping to ensure its accuracy and usability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of the present invention.

[0038] Figure 2 Schematic diagram of the initial point cloud (source point cloud) of the present invention.

[0039] Figure 3 Schematic diagram of the point cloud (source point cloud) after preprocessing of the present invention.

[0040] Figure 4 Schematic diagram of the coarse matching point cloud of the present invention.

[0041] Figure 5 Schematic diagram of the precise matching point cloud of the present invention.

[0042] Figure 6 Flowchart of the multi-resolution registration method of the present invention. DETAILED DESCRIPTION

[0043] The following is a further description of specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is intended to facilitate understanding of the present invention and does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0044] like Figure 1-5 As shown, the present invention provides a coarse-to-fine point cloud registration method, comprising the following steps:

[0045] Step 1: Preprocessing the source point cloud and the target point cloud to be registered respectively to obtain a preprocessed first point cloud and a second point cloud;

[0046] Step 2: Perform a rough registration on the first point cloud and the second point to obtain an initial matching coefficient;

[0047] Step 3: Iteratively solve the objective function and threshold function based on the initial matching coefficient, source point cloud, and target point cloud until the optimal matching coefficient that meets the convergence threshold condition is obtained;

[0048] Step 4: Rotate and translate the source point cloud based on the optimal matching coefficient, convert the source point cloud to the coordinate system of the target point cloud, align the source point cloud and the target point cloud in the same coordinate system, and complete the registration operation.

[0049] In some embodiments, the source and target point clouds to be registered are preprocessed separately to optimize data reading efficiency and reduce memory usage, laying the foundation for subsequent processing. By employing advanced data processing techniques, the processing capabilities of large-scale point cloud data are significantly improved. The specific preprocessing methods include sequential memory mapping, voxel grid downsampling, and noise filtering. Figure 2 This is a schematic diagram of the source point cloud before processing. Figure 3 This is a schematic diagram of the source point cloud after preprocessing, in which noise and outliers are removed and higher quality points are retained.

[0050] Efficient reading of large-scale point cloud data, achieved through memory mapping technology, can significantly reduce memory usage. The core principle of this technology is to map file contents directly into the virtual address space of the process, enabling direct memory access to the file. This approach not only reduces I / O operations but also improves data access speed, making it particularly suitable for processing large files. In actual implementation, mapping files to memory through system calls enables zero-copy data access, significantly improving reading efficiency. This technology is particularly effective when processing large-scale point cloud data, significantly reducing memory pressure and improving overall processing speed.

[0051] Voxel grid downsampling is an efficient data compression method that uses an octree structure to divide the point cloud into voxels, retaining only one representative point within each voxel. This method not only effectively reduces the amount of data but also preserves the key geometric features of the point cloud. The voxel size is calculated using a carefully designed formula:

[0052]

[0053] where v size is the voxel size, V bbox is the point cloud bounding box volume, N desired is the desired number of voxels. This calculation ensures that the geometric characteristics of the point cloud are preserved to the greatest extent possible while compressing the data. This method also uses an adaptive voxel size strategy, dynamically adjusting the voxel size based on the point cloud density, ensuring that more detail is preserved in critical areas while achieving greater compression in non-critical areas.

[0054] The noise filtering step uses the Statistical Outlier Removal (SOR) algorithm to effectively remove abnormal points and outliers. The core principle of the SOR algorithm is to calculate the average distance of each point to its k nearest neighbors. If this distance exceeds n times the standard deviation of the global average distance, the point is considered an outlier. The specific judgment conditions are:

[0055] d i >μ d +nσ d

[0056] in is the average distance from point i to its k nearest neighbors, μ d is the global average distance, σ d is the standard deviation, and n is the threshold coefficient. To improve filtering efficiency, this method utilizes a KD tree to accelerate the nearest neighbor search. Furthermore, local density estimation techniques are introduced to automatically adjust filtering parameters to accommodate complex terrain features. This improvement makes the filtering process more intelligent and efficient, effectively removing noise points while preserving important features.

[0057] In some embodiments, performing coarse registration on the first point cloud and the second point to obtain an initial matching coefficient includes:

[0058] Constructing a first data structure using a KD tree on the first point cloud, and performing feature extraction on the first data structure to obtain a first feature point cloud;

[0059] Constructing the second point cloud using a KD tree to obtain a second data structure, and performing feature extraction on the second data structure to obtain a second feature point cloud;

[0060] The RANSAC algorithm is used to perform robust feature matching on the first feature point cloud and the second feature point cloud to obtain the initial matching coefficient.

[0061] The above coarse registration obtains preliminary registration results through a fast algorithm, providing a good initial value for fine registration. This step significantly improves the registration efficiency and robustness. Figure 4 These are the effects of two angles of coarse registration. The red and yellow ones are the source point cloud and target point cloud after coarse registration, respectively.

[0062] KD tree construction is a key technology for spatial indexing. This paper employs an efficient algorithm with a time complexity of O(n log n). To further improve query efficiency, the SAH (Surface Area Heuristic) algorithm is introduced to optimize the tree's balance. Furthermore, a parallel KD tree construction process is implemented, utilizing multithreading to simultaneously construct different parts of the tree, significantly accelerating the construction process. This optimization is particularly important for processing large-scale point cloud data, significantly reducing initialization time.

[0063] Parallel computing optimization is key to improving processing efficiency. This method fully leverages the advantages of modern multi-core, multi-threaded processors and uses OpenMP parallel technology to achieve loop parallelization. That is, when calculating the objective function, each iteration distance is calculated separately. To maximize hardware utilization, the system dynamically adjusts the number of threads based on the number of CPU cores. At the same time, a dynamic scheduling strategy is adopted to ensure that the workload of each thread is balanced, avoiding situations where some threads are idle while others are overloaded. This parallel optimization strategy not only improves computing speed, but also enhances the scalability of the system, and can better adapt to different hardware environments.

[0064] Feature extraction and matching are the core steps of coarse registration. This paper uses FPFH (Fast Point Feature Histograms) descriptor to extract features. This descriptor can effectively capture the local geometric features of the point cloud. The calculation formula of FPFH descriptor is:

[0065]

[0066] Where SPFH is the simplified point feature histogram, k is the number of neighborhood points, and w k In the feature matching process, the RANSAC algorithm is used for robust matching, which can effectively handle the presence of noise and outliers.

[0067] In some embodiments, the objective function is E(R,t)=∑((ni T ·(R·pi+t-qi)) 2 )

[0068] Among them, E(R, t) is the sum of the vertical distances of all registered points to the corresponding plane, the registered point is the point in the source point cloud projected to the point in the target point cloud, the plane is the differential plane composed of several points near the registered point, R is the rotation matrix, t is the translation vector, p i and q i are the coordinates of the corresponding point i in the source point cloud and the target point cloud, n i is the normal vector at point i in the target point cloud, and ^T represents the vector transpose operation.

[0069] The threshold function is ε=max(ε min ,ε0·e -k·iter ), where ε is the convergence threshold of the current iteration, ε min is the minimum error threshold allowed in the registration process, ε0 is the initial threshold at the beginning of registration, k is the attenuation coefficient that controls the rate of decrease of the threshold, and iter is the current number of iterations.

[0070] The iterative solution of the objective function aims to minimize its value, that is, to obtain the smallest possible value of E(R, t). Therefore, the iterative solution of the objective function must satisfy the following requirement: the E(R, t) calculated in the current iteration is less than the E(R, t) calculated in the previous iteration. This ensures that the objective function value decreases with each iteration, allowing the registration result to gradually approach the optimal solution. This iterative approach can gradually optimize the rotation matrix R and the translation vector t, improving the accuracy, efficiency, and stability of the registration. To accelerate the registration process, a kd-tree is used to accelerate the nearest neighbor search, and the LM (Levenberg-Marquardt) algorithm is employed to optimize the solution. To improve the robustness of the algorithm, an M-estimator can also be introduced to effectively reduce the impact of outliers on the registration results. These optimization measures work together to ensure the efficiency and accuracy of the registration process.

[0071] During iterative solution, in addition to optimizing the objective function, it is also necessary to dynamically adjust the convergence threshold ε through an adaptive threshold mechanism. The principle of this mechanism is to use a larger threshold in the early stage of registration to achieve rapid convergence, and gradually reduce the threshold as the number of iterations increases to ensure that fine alignment can be achieved in the end. The threshold function is ε=max(εmin ,ε0·e -k·iter ), where ε is the convergence threshold of the current iteration, ε min is the minimum error threshold allowed in the registration process, ε0 is the initial threshold at the beginning of registration, k is the attenuation coefficient that controls the rate of decrease of the threshold, and iter is the current number of iterations.

[0072] The above convergence threshold condition is satisfied when ε < ε min , where ε is the convergence threshold of the current iteration, ε min The minimum error threshold allowed in the registration process. That is, when this condition is met, it means that the registration result has reached the preset accuracy requirement and the iteration can be stopped, thus avoiding unnecessary calculations and improving the registration efficiency.

[0073] In addition, the present invention also combines the registration error change rate and dynamically adjusts the attenuation coefficient k to achieve more intelligent convergence control. This adaptive mechanism greatly improves the efficiency and accuracy of registration.

[0074] k(iter)=k0+α·ΔE

[0075]

[0076] Among them, k(iter) is the attenuation coefficient at the iter-th iteration, k0 is the initial value of the attenuation coefficient, α is the control coefficient, ΔE is the error change rate, E iter is the error of the iter-th iteration (i.e., the root mean square of the vertical distance and the iter-th iteration), E iter－1 is the error of the iter-1th iteration.

[0077] In some embodiments, if the number of points in the source point cloud and the target point cloud to be registered is relatively small, the above-mentioned registration operation can be performed directly on the source point cloud and the target point cloud at the original resolution; and when the number of points in the source point cloud and the target point cloud to be registered is greater than the set value, the set value is set according to actual needs, and the multi-resolution registration method can be used to obtain the optimal matching coefficient, thereby realizing the registration operation.

[0078] like Figure 6 As shown, the multi-resolution registration method is used to obtain the optimal matching coefficient, including the following steps:

[0079] Step a, construct a sequence of increasing resolutions: L1, L2, ..., Lm, where Lj = (1 / bj) * Lm, j = 1, 2, ..., m, bj is a decreasing coefficient greater than or equal to 1, and b1>b2> ...>bm, the value of bi can be an exponential change or a multiple change, or other forms of change, bm = 1, Lm is the original resolution (corresponding to the source point cloud and target point cloud resolutions respectively in different preprocessing); in actual applications, only three or five smaller resolutions can be taken, that is, the value of m can be 4 or 6, so as to avoid computational redundancy caused by excessive subdivision.

[0080] Step b: Initialize j to 1, perform steps 1 to 3 on the source point cloud and target point cloud under L1, obtain the first matching coefficient, and then perform step c;

[0081] Step c: Determine the jth matching coefficient as the current matching coefficient and execute step d;

[0082] Step d: Set j = j + 1. If j < m, proceed to step e; if j = m (i.e., j is not less than m), proceed to step f.

[0083] Step e: Based on the current matching coefficient, perform step 3 on the source point cloud and target point cloud under Lj to obtain the jth matching coefficient, and return to step c;

[0084] Step f: Based on the current matching coefficient, perform step 3 on the source point cloud and target point cloud under Lm to obtain the optimal matching coefficient.

[0085] Through the above process, registration can be performed stepwise from low resolution to high resolution, which is more in line with the coarse-to-fine registration approach. At each resolution, the matching coefficient obtained at the previous resolution is used as the initial value, and the matching coefficient at the current resolution is iteratively solved. Finally, the optimal matching coefficient is obtained at the high resolution. This step-by-step approach can effectively improve the efficiency and accuracy of registration while ensuring the reliability of the registration results.

[0086] In some embodiments, the registration results need to be optimized and evaluated to ensure the registration quality, and the accuracy and reliability of the registration results can be fully verified through evaluation indicators.

[0087] The present invention uses the calculation of the root mean square error (RMSE) to evaluate the registration quality. If the root mean square error is less than a set threshold, the registration quality meets the requirements; otherwise, the registration quality does not meet the requirements.

[0088] The calculation formula of RMSE is:

[0089] Among them, p i ' is the coordinate of the corresponding point i in the transformed source point cloud, q iis the coordinate of the corresponding point i in the target point cloud, and n is the number of point pairs.

[0090] The present invention proposes an overall framework of a coarse-to-fine point cloud registration method, covering the key steps from preprocessing, coarse registration, iterative solution of the optimal matching coefficient to the final registration operation. This coarse-to-fine strategy can effectively improve the efficiency of registration. First, a rough matching result is quickly obtained through coarse registration, and then fine adjustments are made on this basis, avoiding the problem of falling into the local optimal solution when fine registration is performed directly, thereby improving the accuracy and reliability of registration. It is suitable for situations where there may be large initial position differences in point clouds between hydraulic structure stations. The solution of the present invention effectively solves the efficiency and accuracy problems faced by traditional methods when processing large-scale hydraulic structure point cloud data, and provides reliable technical support for hydraulic structure monitoring in water conservancy projects. In addition, the modular design of the method makes it have good scalability and adaptability, and it can be easily applied to other fields that require high-precision point cloud registration, such as topographic surveying and building monitoring.

[0091] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered by the scope of protection of the present invention. Matters not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A coarse-to-fine point cloud registration method, characterized by: The following steps are involved: Step 1: Preprocessing the source point cloud and the target point cloud to be registered respectively to obtain a preprocessed first point cloud and a second point cloud; Step 2: Perform a rough registration on the first point cloud and the second point cloud to obtain an initial matching coefficient; Step 3: Iteratively solve the objective function and threshold function based on the initial matching coefficient, source point cloud, and target point cloud until the optimal matching coefficient that meets the convergence threshold condition is obtained; Step 4: Rotate and translate the source point cloud based on the optimal matching coefficient, convert the source point cloud to the coordinate system of the target point cloud, align the source point cloud and the target point cloud in the same coordinate system, and complete the registration operation.

2. The coarse-to-fine point cloud registration method according to claim 1, characterized in that: The preprocessing includes sequentially performing memory mapping, voxel grid downsampling and noise filtering.

3. The coarse-to-fine point cloud registration method according to claim 1, characterized in that: The coarse registration of the first point cloud and the second point to obtain the initial matching coefficient includes: Constructing a first data structure using a KD tree on the first point cloud, and performing feature extraction on the first data structure to obtain a first feature point cloud; Constructing the second point cloud using a KD tree to obtain a second data structure, and performing feature extraction on the second data structure to obtain a second feature point cloud; The RANSAC algorithm is used to perform robust feature matching on the first feature point cloud and the second feature point cloud to obtain the initial matching coefficient.

4. The coarse-to-fine point cloud registration method according to claim 1, wherein: The objective function is E(R, t)=∑((ni T ·(R·pi+t-qi)) 2 ) Among them, E(R, t) is the sum of the vertical distances of all registered points to the corresponding plane, the registered point is the point in the source point cloud projected to the point in the target point cloud, the plane is the differential plane composed of several points near the registered point, R is the rotation matrix, t is the translation vector, p i and q i are the coordinates of the corresponding point i in the source point cloud and the target point cloud, n i is the normal vector at point i in the target point cloud, and ^T represents the vector transpose operation.

5. The coarse-to-fine point cloud registration method according to claim 4, characterized in that: The iterative solution of the objective function must satisfy the following requirement: E(R, t) calculated by the current iteration is less than E(R, t) calculated by the previous iteration.

6. The coarse-to-fine point cloud registration method according to claim 1, characterized in that: The threshold function is ε=max(ε min ,ε0·e -k·iter ), where ε is the convergence threshold of the current iteration, ε min is the minimum error threshold allowed in the registration process, ε0 is the initial threshold at the beginning of registration, k is the attenuation coefficient that controls the rate of decrease of the threshold, and iter is the current number of iterations.

7. The coarse-to-fine point cloud registration method according to claim 1, characterized in that: The convergence threshold condition is satisfied if ε<ε min , where ε is the convergence threshold of the current iteration, ε min The minimum error threshold allowed in the registration process.

8. The coarse-to-fine point cloud registration method according to claim 1, characterized in that: If the number of points in the source point cloud and the target point cloud to be registered is greater than the set value, a multi-scale resolution registration method is used to obtain the optimal matching coefficient.

9. The coarse-to-fine point cloud registration method according to claim 8, characterized in that: The method of obtaining the optimal matching coefficient by adopting the multi-scale resolution registration method includes the following steps: Step a: Construct a resolution increasing sequence: L1, L2, ..., Lm, where Lj = (1 / bj) * Lm, j = 1, 2, ..., m, bj is a decreasing coefficient greater than or equal to 1, and b1>b2> ...>bm, bm = 1, and Lm is the original resolution; Step b: Initialize j to 1, perform steps 1 to 3 on the source point cloud and target point cloud under L1, obtain the first matching coefficient, and perform step c; Step c: Determine the jth matching coefficient as the current matching coefficient and execute step d; Step d: Set j = j + 1. If j < m, proceed to step e; if j = m, proceed to step f. Step e: Based on the current matching coefficient, perform step 3 on the source point cloud and target point cloud under Lj to obtain the jth matching coefficient, and return to step c; Step f: Based on the current matching coefficient, perform step 3 on the source point cloud and target point cloud under Lm to obtain the optimal matching coefficient.

10. The coarse-to-fine point cloud registration method according to claim 1, characterized in that: It also includes verifying the registration quality based on the converted point cloud. The process is: the root mean square error is calculated using the root mean square error formula. If the root mean square error is less than the set threshold, the registration quality meets the requirements, otherwise the registration quality does not meet the requirements.

Images