A laser three-dimensional point cloud reconstruction method for a moving target scene

Abstract

The application provides a laser three-dimensional point cloud reconstruction method for a moving target scene, which comprises the following steps: obtaining under-sampled high-resolution high-density point cloud data, point cloud preprocessing, point cloud registration, scene reconstruction and optimization. The under-sampling is adopted, and a layered scene recovery algorithm is combined to obtain high-resolution high-density point cloud, so that the problem that traditional laser three-dimensional detection technology cannot realize high-speed long-distance target information acquisition can be solved. The point cloud preprocessing comprises point cloud filtering, point cloud simplification, point cloud smoothing and point normal vector estimation operation. The point cloud registration comprises three steps of rough registration, fine registration and global registration, the global scene information is obtained through the point cloud registration, and the scene reconstruction is realized by adopting an Alpha shapes algorithm. The method can accurately give high-resolution high-precision information of the moving target scene, and meets the identification requirement of the long-distance high-speed moving target.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional imaging technology for space targets, and in particular to a high-density laser three-dimensional point cloud reconstruction method for moving target scenes. Background Technology

[0002] 3D reconstruction of space targets enables the identification and confirmation of target details, serving as a crucial component of space situational awareness. Currently, 3D reconstruction of space targets primarily relies on two methods: stereo vision and lidar. LiDAR obtains target information by processing signals reflected from the target surface. It is widely used due to its advantages such as high accuracy, speed, efficiency, ability to measure a wide variety of targets, and high reliability. Furthermore, it is unaffected by lighting conditions and can operate around the clock, making it an important means of acquiring space target information.

[0003] Traditional laser 3D reconstruction primarily employs mechanical scanning and flash imaging. Mechanically scanned lidar is slow, bulky, and inefficient at high resolution. Lidar point cloud reconstruction methods based on this system, such as 3D point cloud reconstruction methods, systems, electronic devices, and storage media (patent application number: 2021100978067), cannot detect and identify high-speed targets. Flash lidar uses area array detectors; for long-range detection, to meet signal-to-noise ratio requirements, it requires ultra-large optical receiving apertures or excessive power supply. Lidar point cloud reconstruction methods based on this system, such as a lightweight point cloud reconstruction method and system based on 3D point cloud data features (patent application number: 2020103226397), cannot meet the needs for long-range, high-precision identification and perception of key spatial targets. For the observation of key spatial targets, which often operates at relatively high speeds and long distances, traditional laser 3D reconstruction technology fails to meet the urgent need for accurate long-range identification of key targets in situational awareness. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a laser 3D point cloud reconstruction method for moving target scenes, including the acquisition, preprocessing, scene reconstruction and optimization of high-resolution high-density point cloud data based on undersampling, so as to meet the recognition needs of long-distance high-speed moving targets.

[0005] The solution of this invention is: a laser 3D point cloud reconstruction method for moving target scenes, comprising the following steps:

[0006] S1. Acquire laser point cloud data to obtain a 3D point cloud image of the current position of the moving target scene; repeat the above operation to obtain 3D point cloud images of the moving target scene at different positions at multiple sampling times; preprocess the laser point cloud data of the 3D point cloud image to complete point cloud filtering, point cloud simplification, point cloud smoothing, and point normal vector estimation.

[0007] S2. Perform coarse and fine registration on the point cloud data in the three-dimensional point cloud images obtained from two adjacent samplings in step S1 to achieve positional alignment of the two sampled images;

[0008] S3. Extract the 3D point cloud image of the next sampling time in S1, repeat S2, and register it with the image of the previous sampling time until the 3D point cloud images of all sampling times are registered.

[0009] S4. Use the Alpha shapes algorithm to reconstruct the scene from the point cloud to the surface using the point cloud data registered in step S3.

[0010] S5. The laser point cloud reconstructed in step S4 is used to repair holes in the data of missing or sparse areas of the point cloud through a triangular mesh hole repair algorithm, thereby completing the scene reconstruction optimization.

[0011] Furthermore, step S1 uses undersampling to acquire laser point cloud data.

[0012] Furthermore, the laser point cloud data hardware platform includes: a pulsed laser, a beam expanding optical system, a receiving optical system, a spatial light modulator, a photodetector, and a data acquisition module; the spatial light modulator is placed at the primary focal plane of the optical path of the receiving optical system.

[0013] Based on the aforementioned hardware platform, the specific method for acquiring laser point cloud data of a moving target scene at the current location is as follows:

[0014] Using a pulsed laser as the laser source, the laser signal emitted by the laser source is collimated by a beam-expanding optical system and incident on the moving target scene. The incident laser is scattered by the moving target scene, resulting in a scattered echo laser signal. The scattered echo laser signal is converged by a receiving optical system. The echo laser signal converged at the primary focal plane of the receiving optical system is undersampled by a spatial light modulator according to a preset two-dimensional coding pattern. The undersampled laser signal is received by a photodetector and converted into an electrical signal, which is then output to a data acquisition module. The data acquisition module outputs a sequence of intensity quantized values ​​of the laser echo signal. The preset two-dimensional coding pattern is changed, and the above process is repeated for Q consecutive samplings. The resulting Q sequences of intensity quantized values ​​of the laser echo signal are input into a depth-layered scene reconstruction algorithm to calculate the three-dimensional coordinate data of the laser point cloud, thus obtaining a three-dimensional point cloud image of the moving target scene at its current position, where Q > 1.

[0015] Furthermore, the hardware platform also includes a time-regulating controller. The laser signal emitted by the laser source is time-coded in the time dimension by the time-regulating controller using "0" and "1", which is then applied to the laser source and received by the beam-expanding optical system.

[0016] Furthermore, the point cloud filtering described in step S1 is implemented using an isolated point filtering or statistical filtering algorithm to filter out non-numerical points, infinite value points, and outliers in the laser point cloud data.

[0017] Furthermore, in step S1, the point cloud simplification process uses a voxelization sampling method, and the set of sampling points serves as the input for point cloud smoothing.

[0018] Furthermore, the point cloud smoothing process described in step S1 employs the least squares method to filter out noise interference in the laser echo.

[0019] Furthermore, the point normal vector estimation in step S1 utilizes the depth gradient algorithm to obtain point normal vector information.

[0020] Furthermore, the coarse registration described in step S2 is implemented as follows: taking the moving target scene as a reference, the three-dimensional point cloud image A and the three-dimensional point cloud image B obtained from two adjacent samplings are projected onto the moving target scene coordinate system. The moving target scene coordinate system takes the centroid of the target in the scene as the origin, the direction of the target's movement as the X direction, the normal direction of the movement trajectory plane as the Z direction, and the Y direction is determined by the right-hand rule.

[0021] Extract *a* data points belonging to 3D point cloud image A and *b* data points belonging to 3D point cloud image B from the overlapping projection region. Obtain the point normal vectors of all extracted data points through point normal vector estimation. The steps are as follows:

[0022] p1. For each data point A extracted from image A... j Determine A j Does there exist a point in B in the neighborhood of A? j If no point in B exists in the neighborhood of A, then A is considered to be... j If no pairing point exists, increment j by 1 and re-execute step p1; otherwise, proceed to step p2.

[0023] p2. Set the initial value of the number of matching point pairs K to 0, and calculate the number of matching point pairs in the neighborhood image B. The point with the smallest value is taken as A. j matching point, K plus 1;

[0024] In the above formula, D1 is A j Distance to the moving target's location in the scene coordinate system. Let δ1 be the distance from each point in the neighborhood image B to the coordinate system of the moving target scene, and let A be the distance from each point in the neighborhood image B to the coordinate system of the moving target scene. j The orientation angle of the normal vector in the scene coordinate system of the moving target. Let be the orientation angle of the normal vector of each point in the neighborhood image B in the coordinate system of the moving target scene;

[0025] p3, j incremented by 1, and steps p1 to p3 were repeated until all j were traversed, j = 1, 2, ..., a;

[0026] p4. Obtain all K pairs of matching points after coarse registration of images A and B, where K ≤ a.

[0027] Furthermore, the implementation of fine registration in step S2 includes the following steps:

[0028] T1. From the K pairs of matching points obtained by coarse registration, randomly select two pairs and combine them to obtain... Various combinations;

[0029] T2. For each combination, perform a coordinate transformation on image B so that two matching points in image B coincide with the corresponding two matching points in image A; calculate the coordinates of the other matching points in image B after the coordinate transformation, and determine the distance d between K pairs of matching points. k The angle θ between the vector and the normal vector k Where: the distance between two overlapping pairs of matching points is 0;

[0030] T3. The distance d between matching point pairs calculated based on step T2. k The angle θ between the vector and the normal vector k Establish the objective function F1 and calculate the objective function value for each combination.

[0031]

[0032]

[0033] Where W1 is the weight representing distance in F1, and W2 is the weight representing the angle between the normal vectors in F1;

[0034] T4. Change to another combination method and repeat steps T2 to T4; continue until all are traversed. Various combinations;

[0035] T5. Select the combination with the smallest objective function value, retain the two pairs of matching points in the combination as the result of fine registration, and use these two pairs of matching points to realize the position alignment of the two sampled images.

[0036] The advantages of this invention compared to the prior art are:

[0037] (1) This invention provides a high-density laser three-dimensional point cloud reconstruction method for moving target scenes. It uses undersampling to obtain laser three-dimensional point clouds, which is lower than the number of samples required by the Nyquist sampling law. It also combines a layered scene recovery algorithm to obtain high-resolution, high-density point clouds, which can make up for the problem that traditional laser three-dimensional detection technology cannot achieve high-speed, long-distance target information acquisition.

[0038] (2) This invention uses the Alpha shapes algorithm to reconstruct the scene from the point cloud to the surface for high-resolution, high-density point cloud data, and can accurately provide high-resolution and high-precision information of the target scene. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the high-density laser three-dimensional point cloud reconstruction method according to an embodiment of the present invention.

[0040] Figure 2 This is a flowchart illustrating the point cloud processing of the high-density laser three-dimensional point cloud reconstruction method according to an embodiment of the present invention. Detailed Implementation

[0041] The invention will now be further described with reference to the accompanying drawings.

[0042] The method described in this embodiment of the invention includes the following steps:

[0043] S1. Acquire laser point cloud data to obtain a 3D point cloud image of the current position of the moving target scene; repeat the above operation to obtain 3D point cloud images of the moving target scene at different positions at multiple sampling times; preprocess the laser point cloud data of the 3D point cloud image to complete point cloud filtering, point cloud simplification, point cloud smoothing, and point normal vector estimation.

[0044] S2. Perform coarse and fine registration on the point cloud data in the three-dimensional point cloud images obtained from two adjacent samplings in step S1 to achieve positional alignment of the two sampled images;

[0045] S3. Extract the 3D point cloud image of the next sampling time in S1, repeat S2, and register it with the image of the previous sampling time until the 3D point cloud images of all sampling times are registered.

[0046] S4. Use the Alpha shapes algorithm to reconstruct the scene from the point cloud to the surface using the point cloud data registered in step S3.

[0047] S5. The laser point cloud reconstructed in step S4 is used to repair holes in the data of missing or sparse areas of the point cloud through a triangular mesh hole repair algorithm, thereby completing the scene reconstruction optimization.

[0048] like Figure 1 As shown, this embodiment of the invention uses undersampling to acquire laser point cloud data. The hardware platform on which it is based includes: a pulsed laser, a time-limited controller, a beam expander optical system, a receiving optical system, a spatial light modulator, a photodetector, and a data acquisition module.

[0049] Among them, when the time controller and spatial light modulator are in use, the working mode is selected by external trigger control. The external trigger control sends command signals to the time controller and spatial light modulator. Each command signal sent is recorded as command signal p, p = 1, ..., N, which controls the time controller and spatial light modulator to work under the p-th encoding. On the other hand, it sends command signal p to the data acquisition module to start acquisition, so as to realize the synchronization of the time controller, spatial light modulator and data acquisition module.

[0050] Based on the aforementioned hardware platform, the method for acquiring high-density laser 3D point clouds is as follows: A pulsed laser is selected as the laser source. The emitted laser light from the pulsed laser passes through a time-regulating modulator, where it is time-coded using "0" and "1" in the time dimension. This time coding is applied to the laser source and received by a beam-expanding optical system. The beam-expanding optical system then collimates and outputs the laser signal. The modulated laser light is transmitted to a moving target scene via a far-field input. The moving target scene generates a scattered echo signal from the incident laser light. The scattered echo signal is converged by a receiving optical system. A spatial light modulator is placed at the primary focal plane of the optical system's path. The spatial light modulator is randomly configured with "0" and "1". The two-dimensional coded pattern is formed, and the converged echo laser signal is sampled by a spatial light modulator according to the two-dimensional coded pattern. The sampled laser signal is received by a photodetector. The electrical signal output by the photodetector through photoelectric conversion is converted from analog signal to digital signal by a data acquisition module, and outputs a time-series intensity quantization value sequence of the laser echo signal. The preset two-dimensional coded pattern is changed, and the above process is repeated for 10 consecutive samplings. The intensity quantization value sequence of the 10 laser echo signals is input into the depth layering scene restoration algorithm to calculate the three-dimensional coordinate data of the laser point cloud and obtain the three-dimensional point cloud image of the current position of the moving target scene.

[0051] Furthermore, the deep layered scene recovery algorithm is as follows:

[0052] Ten intensity quantization sequences of laser echo signals form ten echo curves, serving as data points. The time at the peak of each echo curve is multiplied by the speed of light and divided by two to obtain multiple distance values ​​from the hardware platform to the moving target scene. A spectral estimation algorithm is used to layer the moving target scene according to these distance values, obtaining the intensity value and distance to the hardware platform for each layer. Then, the prior coding information of the spatial light modulator is used as a measurement matrix, and the intensity value of each layer is used as the measurement value. The Orthogonal Matching Pursuit (OMP) algorithm is used to obtain the two-dimensional coordinates of each layer's data points. Based on the two-dimensional coordinates of each layer's data points and the distance to the hardware platform for each layer, the three-dimensional coordinate data of the laser point cloud is obtained. The three-dimensional coordinate form of each layer is (X... i ,Y i Z i ).

[0053] Laser point cloud data (X) obtained by deep layered scene restoration algorithm i ,Y i Z i For a single measurement, the number of point clouds can reach i = M. 3 M represents the number of pixels in the X or Y dimension of the spatial light modulator, typically X = Y. The laser point cloud data obtained from the deep layered scene reconstruction algorithm serves as the input data for the preprocessing module.

[0054] like Figure 2 As shown, the acquired high-density 3D laser point cloud is first preprocessed, including point cloud filtering, point cloud simplification, point cloud smoothing, and point cloud normal vector estimation.

[0055] Point cloud filtering can remove non-numerical points, infinite points, and outliers in 3D laser point clouds. It can be achieved by using calculation methods such as isolated point filtering and statistical filtering.

[0056] In this embodiment, a statistical filtering method is employed. First, based on the input high-density point cloud data, the global density knn is calculated, and the global standard deviation or the standard deviation of the nearest neighbor distances of all points std is given as the input to the statistical filtering method. Second, for each laser point, the distances to its knn surrounding points are calculated, and their standard deviations are determined. Then, these distances are compared with the global standard deviation std. If the distance is greater than the global standard deviation, the point is filtered out; otherwise, it is retained.

[0057] Point cloud simplification removes redundant points with repetitive features, reducing the amount of data. In this embodiment, voxelization sampling is used in the point cloud simplification process. First, the 3D space is voxelized, and then the center point or the point closest to the center in each voxel is taken as the sampling point. The set of sampling points extracted from each voxel is used as the input for point cloud smoothing. The voxelized data can be efficiently used with spatial convolution, which is beneficial for extracting multi-scale and hierarchical local feature information.

[0058] The least squares method is used to smooth the point cloud and filter out noise interference in the laser echo.

[0059] After point cloud smoothing, the depth gradient algorithm is used to approximate the point normal vectors for subsequent point cloud reconstruction algorithms; the formula for calculating the point normal vectors using the depth gradient algorithm is as follows:

[0060]

[0061] In the above formula, P is the coordinate of the current three-dimensional laser point, I is the intensity of the current three-dimensional laser point P, and Q1, ..., Q n Let I be the coordinates of the laser point in the neighborhood of the current 3D laser point, with corresponding intensity values ​​I1, ..., I2. n ; Let v1, ..., v be the gradient at point P. n It is a unit direction vector.

[0062] Subsequently, the preprocessed point cloud is registered with images of moving targets at multiple sampling times through a three-step process: conventional coarse registration, fine registration, and global registration, in order to obtain full-field scene information. The principle of the three registration methods is to transform point cloud data obtained from different viewpoints at different times to achieve in-motion imaging.

[0063] (1) The coarse registration method is as follows:

[0064] Based on the imaging angle of the moving target scene, and taking the moving target scene as a reference, the images A and B obtained from two adjacent samplings are projected onto the moving target scene coordinate system. The moving target scene coordinate system takes the centroid of the target in the scene as the origin, the direction of the target's movement as the X direction, the normal direction of the movement trajectory plane as the Z direction, and the Y direction as determined by the right-hand rule.

[0065] Extract *a* data points belonging to image A and *b* data points belonging to image B from the overlapping projection region. Obtain the normal vectors of all extracted data points through point normal vector estimation. The steps are as follows:

[0066] p1. For each data point A extracted from image A... j Determine A j Does there exist a point in B in the neighborhood of A? j If no point in B exists in the neighborhood of A, then A is considered to be... j If no pairing point exists, increment j by 1 and re-execute step p1; otherwise, proceed to step p2.

[0067] p2. Set the initial value of the number of matching point pairs K to 0, and calculate the number of matching point pairs in the neighborhood image B. The point with the smallest value is taken as A. j matching point, K plus 1;

[0068] In the above formula, D1 is A j Distance to the moving target's location in the scene coordinate system. Let δ1 be the distance from each point in the neighborhood image B to the coordinate system of the moving target scene, and let A be the distance from each point in the neighborhood image B to the coordinate system of the moving target scene. j The orientation angle of the normal vector in the scene coordinate system of the moving target. Let be the orientation angle of the normal vector of each point in the neighborhood image B in the coordinate system of the moving target scene;

[0069] p3, j incremented by 1, and steps p1 to p3 were repeated until all j were traversed, j = 1, 2, ..., a;

[0070] p4. Obtain all K pairs of matching points after coarse registration of images A and B, where K ≤ a.

[0071] (2) The method for achieving fine registration includes the following steps:

[0072] T1. From the K pairs of matching points obtained by coarse registration, randomly select two pairs and combine them to obtain... Various combinations;

[0073] T2. For each combination, perform a coordinate transformation on image B so that two matching points in image B coincide with the corresponding two matching points in image A; calculate the coordinates of the other matching points in image B after the coordinate transformation, and redetermine the distance d between K pairs of matching points. k The angle θ between the vector and the normal vector k Where: the distance between two overlapping pairs of matching points is 0;

[0074] T3. The distance d between matching point pairs calculated based on step T2. k The angle θ between the vector and the normal vector k Establish the objective function F1 and calculate the objective function value for each combination.

[0075]

[0076]

[0077] Where W1 is the weight representing distance in F1, and W2 is the weight representing the angle between the normal vectors in F1;

[0078] T4. Change to another combination method and repeat steps T2 to T4; iterate through all... Various combinations;

[0079] T5. Select the combination with the smallest objective function value, and retain the two pairs of matching points in this combination as the result of fine registration.

[0080] (3) The global registration implementation method is as follows:

[0081] For 3D point cloud images of moving target scenes at different sampling times, extract new images and repeat coarse and fine registration operations with the images at the previous sampling time to complete global registration.

[0082] Then, the processed point cloud data is used to reconstruct the scene onto a 3D surface. The scene reconstruction can be achieved by using the Alpha Shapes algorithm or the Poisson algorithm to reconstruct the point cloud onto the surface.

[0083] In this embodiment, the Alpha shapes algorithm is used to reconstruct the scene of the three-dimensional surface from the point cloud data. The create_from_point_cloud_alpha_shape() function in the Open3d module of Python software is used. The point cloud data is substituted into the function as input parameters, and the output is the surface reconstruction variable.

[0084] For moving target scenes, when adjusting the alpha value, to ensure reconstruction results, the alpha value range used in sparse point cloud regions is 0-10; the alpha value range used in dense point cloud regions is 50-100. The point cloud density in dense regions is 10 times that in sparse regions. For example, regions with a point cloud density less than 5 are considered sparse regions, and regions with a point cloud density greater than 50 are considered dense regions.

[0085] Finally, the post-processing steps, such as hole repair, are completed using a triangular mesh hole repair algorithm to achieve scene reconstruction and optimization of the target 3D surface.

[0086] The described triangular mesh hole-filling algorithm primarily employs a continuously expanding approach based on neighborhood information to complete the hole-filling post-processing of data in point cloud that is missing or sparse. In this embodiment, the average point distance of all points is first calculated. Then, the uniformity of point distribution in the point cloud is used to determine boundary features. Finally, based on the above information, holes are uniformly filled, thereby completing the scene reconstruction optimization of the target 3D curved surface.

[0087] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A laser-based 3D point cloud reconstruction method for moving target scenes, characterized in that, Includes the following steps: S1. Acquire laser point cloud data to obtain a 3D point cloud image of the current position of the moving target scene; repeat the above operation to obtain 3D point cloud images of the moving target scene at different positions at multiple sampling times; preprocess the laser point cloud data of the 3D point cloud image to complete point cloud filtering, point cloud simplification, point cloud smoothing, and point normal vector estimation. S2. Perform coarse and fine registration on the point cloud data in the three-dimensional point cloud images obtained from two adjacent samplings in step S1 to achieve positional alignment of the two sampled images; S3. Extract the 3D point cloud image of the next sampling time in S1, repeat S2, and register it with the image of the previous sampling time until the 3D point cloud images of all sampling times are registered. S4. Use the Alpha shapes algorithm to reconstruct the scene from the point cloud to the surface using the point cloud data registered in step S3. S5. The laser point cloud reconstructed in step S4 is used to repair holes in the data of missing or sparse areas of the point cloud through the triangular mesh hole repair algorithm, thus completing the scene reconstruction optimization. Step S1 uses undersampling to acquire laser point cloud data; The laser point cloud data hardware platform includes: a pulsed laser, a beam expanding optical system, a receiving optical system, a spatial light modulator, a photodetector, and a data acquisition module; the spatial light modulator is placed at the primary focal plane of the optical path of the receiving optical system. Based on the aforementioned hardware platform, the specific method for acquiring laser point cloud data of a moving target scene at the current location is as follows: Using a pulsed laser as the laser source, the laser signal emitted by the laser source is collimated by a beam-expanding optical system and incident on the moving target scene. The incident laser is scattered by the moving target scene, resulting in a scattered echo laser signal. The scattered echo laser signal is converged by a receiving optical system. The echo laser signal converged at the primary focal plane of the receiving optical system is undersampled by a spatial light modulator according to a preset two-dimensional coding pattern. The undersampled laser signal is received by a photodetector and converted into an electrical signal, which is then output to a data acquisition module. The data acquisition module outputs a sequence of intensity quantized values ​​of the laser echo signal. The preset two-dimensional coding pattern is changed, and the above process is repeated for Q consecutive samplings. The resulting Q sequences of intensity quantized values ​​of the laser echo signal are input into a depth-layered scene reconstruction algorithm to calculate the three-dimensional coordinate data of the laser point cloud, thus obtaining a three-dimensional point cloud image of the moving target scene at its current position, where Q > 1.

2. The laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, The hardware platform also includes a time-controller. The laser signal emitted by the laser source is time-coded in the time dimension by the time-controller using "0" and "1", and then applied to the laser source and received by the beam-expanding optical system.

3. The laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, The point cloud filtering described in step S1 is implemented using an isolated point filtering or statistical filtering algorithm to filter out non-numerical points, infinite value points, and outliers in the laser point cloud data.

4. The laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, Step S1 describes a point cloud simplification process that uses a voxelization sampling method, where the set of sampling points serves as the input for point cloud smoothing.

5. The laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, The point cloud smoothing process described in step S1 uses the least squares method to filter out noise interference in the laser echo.

6. The laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, The point normal vector estimation in step S1 uses the depth gradient algorithm to obtain point normal vector information.

7. The laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, The implementation method of coarse registration in step S2 is as follows: taking the moving target scene as a reference, the three-dimensional point cloud image A and the three-dimensional point cloud image B obtained from two adjacent samplings are projected onto the moving target scene coordinate system. The moving target scene coordinate system takes the centroid of the target in the scene as the origin, the direction of the target's movement as the X direction, the normal direction of the movement trajectory plane as the Z direction, and the Y direction is determined by the right-hand rule. Extract *a* data points belonging to 3D point cloud image A and *b* data points belonging to 3D point cloud image B from the overlapping projection region. Obtain the point normal vectors of all extracted data points through point normal vector estimation. The steps are as follows: p1. For each data point A extracted from image A... j Determine A j Does there exist a point in B in the neighborhood of A? j If no point in B exists in the neighborhood of A, then A is considered to be... j If no pairing point exists, increment j by 1 and re-execute step p1; otherwise, proceed to step p2. p2. Set the initial value of the number of matching point pairs K to 0, and calculate the number of matching point pairs in the neighborhood image B. The point with the smallest value is taken as A. j matching point, K plus 1; In the above formula, For A j Distance to the moving target's location in the scene coordinate system. Let be the distance from each point in the neighborhood image B to the coordinate system of the moving target scene. For A j The orientation angle of the normal vector in the scene coordinate system of the moving target. Let be the orientation angle of the normal vector of each point in the neighborhood image B in the coordinate system of the moving target scene; p3, j incremented by 1, and steps p1~p3 were repeated until all j were traversed, j=1,2,…,a; p4. Obtain all K pairs of matching points after coarse registration of images A and B, where K ≤ a.

8. A laser 3D point cloud reconstruction method for a moving target scene according to claim 1, characterized in that, The fine registration described in step S2 includes the following steps: T1. From the K pairs of matching points obtained by coarse registration, randomly select two pairs and combine them to obtain... Various combinations; T2. For each combination, perform a coordinate transformation on image B so that two matching points in image B coincide with the corresponding two matching points in image A; calculate the coordinates of the other matching points in image B after the coordinate transformation, and determine the distance d between K pairs of matching points. k Angle with normal vector k Where: the distance between two overlapping pairs of matching points is 0; T3. The distance d between matching point pairs calculated based on step T2. k Angle with normal vector k Establish the objective function F1 and calculate the objective function value for each combination. Where W1 is the weight representing distance in F1, and W2 is the weight representing the angle between the normal vectors in F1; T4. Change to another combination method and repeat steps T2~T4; until all are traversed. Various combinations; T5. Select the combination with the smallest objective function value, retain the two pairs of matching points in the combination as the result of fine registration, and use these two pairs of matching points to realize the position alignment of the two sampled images.

Images