Method, apparatus and computer-readable storage medium for rapid point cloud noise removal

Abstract

This invention discloses a method, apparatus, and computer-readable storage medium for rapid point cloud noise removal, which first calculates the neighborhood kernel N. m The i-th noise point P within i With the m-th pixel image I m Distance L i Then determine the neighboring kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i The size of the neighborhood radius R is used to determine the neighborhood kernel N. m The i-th noise point P within i Perform noise point counting to obtain information about the neighborhood kernel N. m Noise Counting Figure I mC Then, determine the noise count in Figure I. mC Each point within the image is related to the m-th pixel I. m The distance L and the preset threshold T are used to count noise points in image I. mC Mark each point within the neighborhood kernel N to determine if it is a point to be cleaned. m Points marked as to be cleaned are cleaned. This invention simplifies the noise removal process and reduces the difficulty of noise removal by introducing a neighborhood kernel to count and mark the noise of each pixel, thereby improving the accuracy of noise removal. It is suitable for deployment on imaging device terminals to improve the quality of point clouds and depth maps acquired through imaging.

Description

Technical Field

[0001] This invention relates to the field of machine vision 3D measurement technology, and in particular to a method, apparatus and computer-readable storage medium for rapid point cloud noise removal. Background Technology

[0002] Noise will be present in the point cloud generated after a 3D imaging device scans an object. Most of these noise points are discrete, isolated clusters of points. Therefore, noise removal is necessary to obtain better imaging results.

[0003] In existing methods, noise in point clouds is mainly removed using the following techniques:

[0004] 1. Using each point in the point cloud as the center, a relatively fast nearest neighbor search is achieved using data structures such as KDtree to remove noise from the point cloud;

[0005] 2. Using each point in the point cloud as the center, establish an initial radius to form a sphere and iteratively synthesize point clusters to remove noise from the point cloud;

[0006] 3. Triangulate the point cloud and divide it into multiple sub-images for processing to remove noise from the point cloud.

[0007] However, the above methods all have problems such as complex processes and long processing times. Furthermore, it is difficult to combine them with CUDA shared memory technology to clean up point cloud noise, and the cleaning effect is often different from that on the CPU side, which seriously restricts the development of point cloud noise cleaning technology. Summary of the Invention

[0008] The purpose of this invention is to provide a method, apparatus, and computer-readable storage medium for rapid point cloud noise removal. By introducing a neighborhood kernel to statistically identify and label the noise in each pixel, the method effectively simplifies the noise removal process, reduces the difficulty of removal, and greatly improves the accuracy of noise removal. This method is suitable for deployment on imaging device terminals to improve the quality of point clouds and depth maps acquired through imaging.

[0009] To achieve the above objectives, this invention discloses a method for rapid noise removal in point clouds, wherein the point cloud has an image of M pixels, and the method includes the following steps:

[0010] S1, using the m-th pixel image I m Centered on the m-th pixel image I, construct a new image based on the m-th pixel. m The neighborhood kernel N m , m = 1, ..., M, where m and M are natural numbers greater than or equal to 1;

[0011] S2, Calculate the neighborhood kernel N m The i-th noise point P within iWith the m-th pixel image I m Distance L i , where the neighboring kernel N m There are G noise points P in the memory, i = 1, ..., G, where i and G are natural numbers greater than or equal to 1;

[0012] S3, Determine the neighborhood kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i , and the size of the preset neighborhood radius R;

[0013] S4. Based on the judgment result of step S3, perform a judgment on the neighboring kernel N. m The i-th noise point P within i Perform noise counting;

[0014] S5. Repeat steps S2 to S4 until the neighborhood kernel N has been traversed. m For all noise points P within the region, obtain information about the neighborhood kernel N. m Noise Counting Figure I mC ;

[0015] S6. Determine the noise count in Figure I respectively. mC Each point within the image is related to the m-th pixel I. m The distance L and the magnitude of the preset threshold T;

[0016] S7. Based on the judgment result of step S6, perform the noise counting on image I respectively. mC Mark each point within the area to determine whether it is a point to be cleaned;

[0017] S8, for the neighboring kernel N m Clean up the points marked as to be cleaned.

[0018] S9. Repeat steps S1 to S8 until all pixel images of the point cloud have been traversed.

[0019] Preferably, the neighborhood radius R is introduced, and the expression for the neighborhood kernel N is N = (R*2-1)*(R*2-1).

[0020] Preferably, step S4 specifically includes:

[0021] If the neighboring kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i If the value is less than the neighborhood radius R, then for the neighborhood kernel N... m Noise Counting Figure I mC Add 1 to the total.

[0022] Preferably, the expression for the preset threshold T is T=(R*2-1)*(R*2-1) / 2.

[0023] Preferably, step S7 specifically includes:

[0024] If noise count diagram I mC Any point within the image and the m-th pixel I m If the distance L is less than the preset threshold T, then the noise count map I will be updated. mC The current point within is marked as a point to be cleaned.

[0025] Preferably, the method further includes the following steps before step S1:

[0026] Determine the m-th pixel image I m Is this a valid pixel image?

[0027] If so, proceed to step S1;

[0028] If not, then discard the m-th pixel image I. m .

[0029] Preferably, the method further includes the following steps before step S1:

[0030] Determine the type of point cloud;

[0031] If the point cloud is an ordered point cloud, proceed to step S1;

[0032] If the point cloud is an unordered point cloud, then the point cloud is rasterized to convert the unordered point cloud into an ordered point cloud before proceeding to step S1.

[0033] Preferably, the point cloud noise fast cleaning method is implemented using CUDA technology and shared memory technology.

[0034] Accordingly, the present invention also discloses a point cloud noise rapid cleaning device, wherein the point cloud has an M-pixel image, and the point cloud noise rapid cleaning device comprises:

[0035] Building unit, used to construct the image I at the m-th pixel m Centered on the m-th pixel image I, construct a new image based on the m-th pixel. m The neighborhood kernel N m , m = 1, ..., M, where m and M are natural numbers greater than or equal to 1;

[0036] Computational unit, used to compute neighborhood kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i , where the neighboring kernel N mThere are G noise points P in the memory, i = 1, ..., G, where i and G are natural numbers greater than or equal to 1;

[0037] The first judgment unit is used to judge the neighboring kernel N. m The i-th noise point P within i With the m-th pixel image I m Distance L i , and the size of the preset neighborhood radius R;

[0038] The counting unit is used to count the neighboring kernel N based on the judgment result of the first judgment unit. m The i-th noise point P within i Perform noise counting;

[0039] The first traversal unit is used to traverse the neighboring kernel N. m For all noise points P within the region, obtain information about the neighborhood kernel N. m Noise Counting Figure I mC ;

[0040] The second judgment unit is used to judge the noise point counting diagram I respectively. mC Each point within the image is related to the m-th pixel I. m The distance L and the magnitude of the preset threshold T;

[0041] The marking unit is used to mark the noise point counting image I according to the judgment result of the second judgment unit. mC Mark each point within the area to determine whether it is a point to be cleaned;

[0042] The cleanup unit is used to clean up neighboring kernel N. m Clean up the points marked as to be cleaned.

[0043] The second traversal unit is used to traverse all pixel images of the point cloud.

[0044] Accordingly, the present invention also discloses a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the point cloud noise fast cleaning method described above.

[0045] Compared with existing technologies, the present invention first calculates the neighborhood kernel N. m The i-th noise point P within i With the m-th pixel image I m Distance L i Then determine the neighboring kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i The size of the neighborhood radius R is used to determine the neighborhood kernel N. m The i-th noise point P withini Perform noise point counting to obtain information about the neighborhood kernel N. m Noise Counting Figure I mC Then, determine the noise count in Figure I. mC Each point within the image is related to the m-th pixel I. m The distance L and the preset threshold T are used to count noise points in image I. mC Mark each point within the neighborhood kernel N to determine if it is a point to be cleaned. m The method cleans up the noise of all pixels in the point cloud by marking the points to be cleaned. It introduces a neighborhood kernel to count and mark the noise of each pixel, which effectively simplifies the noise removal process and reduces the difficulty of removal, greatly improving the accuracy of noise removal. Moreover, this method is suitable for deployment on imaging device terminals to improve the quality of point clouds and depth maps acquired by imaging. Attached Figure Description

[0046] Figure 1 This is a flowchart of the point cloud noise rapid cleaning method of the present invention;

[0047] Figure 2 It is a localized marker map that traditionally uses the CPU to perform point cloud noise cleanup;

[0048] Figure 3 This is a local marker image of point cloud noise removal using the point cloud noise rapid removal method of the present invention;

[0049] Figure 4 This is a standard block depth map of the point cloud before noise removal;

[0050] Figure 5 The point cloud noise rapid cleaning method of the present invention is used to... Figure 4 Depth map of standard blocks after point cloud noise removal;

[0051] Figure 6 This is a time comparison chart of point cloud noise removal using CPU-based processing versus point cloud noise removal using CUDA shared memory technology.

[0052] Figure 7 This is a time comparison chart of point cloud noise removal using CUDA technology but without shared memory technology, and using CUDA technology and shared memory technology.

[0053] Figure 8 This is a structural block diagram of the point cloud noise rapid cleaning device of the present invention. Detailed Implementation

[0054] To illustrate the technical content, structural features, objectives, and effects of the present invention in detail, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0055] Please see Figures 1-7 As shown, the point cloud noise removal method of this embodiment is suitable for removing noise from point clouds obtained by visual acquisition devices such as imaging device terminals. The point cloud has M pixel images, and the point cloud noise removal method includes the following steps:

[0056] S1, using the m-th pixel image I m Centered on the m-th pixel image I, construct a new image based on the m-th pixel. m The neighborhood kernel N m Let m = 1, ..., M, where m and M are natural numbers greater than or equal to 1. Let I be the image of the m-th pixel. m The center coordinates are I m (u,v).

[0057] Preferably, the neighborhood radius R is introduced, and the expression for the neighborhood kernel N is N = (R*2-1)*(R*2-1).

[0058] Preferably, the method further includes the following steps before step S1:

[0059] Determine the type of point cloud, which is divided into ordered point clouds and unordered point clouds.

[0060] If the point cloud is an ordered point cloud, proceed to step S1.

[0061] If the point cloud is unordered, it is rasterized to convert it into an ordered point cloud before proceeding to step S1. That is, this step actually involves... m (u,v) equals V i At that time, I m (u,v) represents an unordered point cloud.

[0062] Preferably, the invalid value for different types of images is defined as V. i Before step S1, the method further includes:

[0063] Determine the m-th pixel image I m Is it a valid pixel image?

[0064] If so, proceed to step S1. Specifically, in I... m (u,v) is not equal to V i At that time, it can be considered that I m (u,v) represents the effective pixel image.

[0065] If not, then discard the m-th pixel image I. m Specifically, in I m (u,v) equals V i At that time, it can be considered that I m(u,v) represents an invalid pixel image. This invention does not process invalid pixel images; therefore, invalid pixel images are discarded.

[0066] S2, Calculate the neighborhood kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i , where the neighboring kernel N m There are G noisy points P in the memory, i = 1, ..., G, where i and G are natural numbers greater than or equal to 1. For a neighborhood kernel N... m The i-th noise point P within i , where the neighboring kernel N m The i-th noise point P within i The center coordinates are P m (u+i,v+j), i,j∈[0,2R-1].

[0067] S3, Determine the neighborhood kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i The size of the neighborhood radius R is related to the preset value.

[0068] Specifically, when the neighboring kernel N m The i-th noise point P within i When the image is a valid pixel image, according to formula L i =||I m (u,v)|-|P m (u+i,v+j)||, calculate the neighborhood kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i The size of the neighborhood radius R is related to the preset value.

[0069] S4. Based on the judgment result of step S3, perform a judgment on the neighboring kernel N. m The i-th noise point P within i Perform noise counting.

[0070] Preferably, step S4 specifically includes:

[0071] If the neighboring kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i If the value is less than the neighborhood radius R, then for the neighborhood kernel N... m Noise Counting Figure I mC The increment is 1. If the count is represented by C, then for the neighborhood kernel N...m Its initial count value is zero.

[0072] S5. Repeat steps S2 to S4 until the neighborhood kernel N has been traversed. m For all noise points P within the region, obtain information about the neighborhood kernel N. m Noise Counting Figure I mC That is, this step yields the neighborhood kernel N. m Total number of noise points within.

[0073] S6. Determine the noise count in Figure I respectively. mC Each point within the image is related to the m-th pixel I. m The distance L is related to the magnitude of the preset threshold T. Preferably, the expression for the preset threshold T is T = (R*2-1)*(R*2-1) / 2.

[0074] S7. Based on the judgment result of step S6, perform the noise counting on image I respectively. mC Each point within the area is marked as either a point to be cleaned.

[0075] Preferably, step S7 specifically includes:

[0076] If noise count diagram I mC Any point within the image and the m-th pixel I m If the distance L is less than the preset threshold T, then the noise count map I will be updated. mC The current point within is marked as a point to be cleaned.

[0077] S8, for the neighboring kernel N m Clean up the points marked as to be cleaned.

[0078] S9. Repeat steps S1 to S8 until all pixel images of the point cloud have been traversed.

[0079] Preferably, this embodiment employs CUDA and shared memory technologies to execute the fast point cloud noise removal method. Pseudocode for executing the fast point cloud noise removal method using CUDA and shared memory technologies is as follows:

[0080] "Knernel Label(Im,Ic,R)"

[0081] Ty = threadIdx.y

[0082] Tx = threadIdx.x

[0083] Ro = blockIDx.y * TILE_DIM + Ty

[0084] Co = blockIDx.x * TILE_DIM + Tx

[0085] Ri=Ro-(2R-1)

[0086] Ci=Co-(2R-1)

[0087] __shared__cache[BLOCK_W][BLOCK_H]

[0088] Do

[0089] Cache[Ty][Tx]=Im[Ri][Ci]

[0090] __syncthreads()

[0091] Count=0

[0092] If Ty<TILE_DIM&&Tx<TILE_DIM then

[0093] If Cache[Ty][Tx]!=Inval id then

[0094] For all(2R-1)do

[0095] If Cache[Ty+i][Tx+j]!=Inval id then

[0096] Len=abs(abs(Cache[Ty+i][Tx+j])-abs(Cache[Ty][Tx]))

[0097] Count=Count+1

[0098] End if

[0099] End for

[0100] End if

[0101] End if

[0102] __syncthreads()

[0103] If Ri<height&&Ci<width then

[0104] Ic[Ri][Ci]=Count

[0105] End if”。

[0106] It is understood that this embodiment combines CUDA technology and shared memory technology to execute the point cloud noise fast cleaning method, so as to achieve a cleaning technology similar to sliding window filtering, thereby achieving better processing efficiency and facilitating deployment to imaging device terminals, thus improving the quality of point clouds and depth maps acquired through imaging. Figure 6 The graph shows a time comparison between point cloud noise removal using CPU-based processing and point cloud noise removal using CUDA shared memory technology. Figure 7 The graph shows a time comparison for point cloud noise removal using CUDA technology without shared memory versus using both CUDA and shared memory technologies. Figure 6 and Figure 7 It is evident that combining CUDA technology and shared memory technology for point cloud noise removal can achieve higher removal efficiency.

[0107] Please see Figure 8 Accordingly, the present invention also discloses a point cloud noise rapid cleaning device, wherein the point cloud has an M-pixel image, and the point cloud noise rapid cleaning device includes:

[0108] Construction unit 10, used to construct the m-th pixel image I m Centered on the m-th pixel image I, construct a new image based on the m-th pixel. m The neighborhood kernel N m , m = 1, ..., M, where m and M are natural numbers greater than or equal to 1;

[0109] Computation unit 20 is used to compute the neighborhood kernel N. m The i-th noise point P within i With the m-th pixel image I m Distance L i , where the neighboring kernel N m There are G noise points P in the memory, i = 1, ..., G, where i and G are natural numbers greater than or equal to 1;

[0110] The first judgment unit 30 is used to judge the neighboring kernel N. m The i-th noise point P within i With the m-th pixel image I m Distance L i , and the size of the preset neighborhood radius R;

[0111] The counting unit 40 is used to count the neighboring kernel N based on the judgment result of the first judgment unit 30. m The i-th noise point P within i Perform noise counting;

[0112] The first traversal unit 50 is used to traverse the neighboring kernel N. m For all noise points P within the region, obtain information about the neighborhood kernel N.m Noise Counting Figure I mC ;

[0113] The second judgment unit 60 is used to judge the noise point counting map I respectively. mC Each point within the image is related to the m-th pixel I. m The distance L and the magnitude of the preset threshold T;

[0114] The marking unit 70 is used to mark the noise point counting map I according to the judgment result of the second judgment unit 60. mC Mark each point within the area to determine whether it is a point to be cleaned;

[0115] Cleanup unit 80 is used to clean up neighboring kernel N. m Clean up the points marked as to be cleaned.

[0116] The second traversal unit 90 is used to traverse all pixel images of the point cloud.

[0117] Accordingly, the present invention also discloses a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the point cloud noise fast cleaning method described above.

[0118] Combination Figures 1-8 This invention first calculates the neighborhood kernel N. m The i-th noise point P within i With the m-th pixel image I m Distance L i Then determine the neighboring kernel N m The i-th noise point P within i With the m-th pixel image I m Distance L i The size of the neighborhood radius R is used to determine the neighborhood kernel N. m The i-th noise point P within i Perform noise point counting to obtain information about the neighborhood kernel N. m Noise Counting Figure I mC Then, determine the noise count in Figure I. mC Each point within the image is related to the m-th pixel I. m The distance L and the preset threshold T are used to count noise points in image I. mC Mark each point within the neighborhood kernel N to determine if it is a point to be cleaned. m The method cleans up the noise of all pixels in the point cloud by marking the points to be cleaned. It introduces a neighborhood kernel to count and mark the noise of each pixel, which effectively simplifies the noise removal process and reduces the difficulty of removal, greatly improving the accuracy of noise removal. Moreover, this method is suitable for deployment on imaging device terminals to improve the quality of point clouds and depth maps acquired by imaging.

[0119] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for rapid point cloud noise removal, characterized in that, A point cloud has an image of M pixels. The method for rapid noise removal of the point cloud includes the following steps: S1, with the m-th pixel image Centered on the m-th pixel image, construct a model based on the m-th pixel. Neighborhood kernel , m=1, ...,M, where m and M are natural numbers greater than or equal to 1; S2, Calculate the neighborhood kernel The i-th noise point within With the m-th pixel image distance Among them, the neighboring kernel There are G noisy points P in memory, i=1,...,G, where i and G are natural numbers greater than or equal to 1; S3, Determine the neighboring kernel The i-th noise point within With the m-th pixel image distance , and the size of the preset neighborhood radius R; S4. Based on the judgment result of step S3, perform a judgment on the neighboring kernel. The i-th noise point within Perform noise counting; S5. Repeat steps S2 to S4 until the neighboring kernels have been traversed. For all noise points P within the kernel, obtain information about the neighborhood kernel. Noise count map ; S6. Determine the noise count map separately. Each point within the image is related to the m-th pixel. The distance L and the size of the preset threshold T; the expression for the preset threshold T is T=(R*2-1)*(R*2-1) / 2; S7. Based on the judgment result of step S6, perform a separate analysis of the noise point counting map. Mark each point within the area to determine whether it is a point to be cleaned; S8, for neighboring kernels Clean up the points marked as to be cleaned. S9. Repeat steps S1 to S8 until all pixel images of the point cloud are traversed. Step S7 specifically includes: If noise count map Any point within the image and the m-th pixel If the distance L is less than the preset threshold T, then the noise count map will be... The current point within is marked as a point to be cleaned.

2. The point cloud noise rapid cleaning method as described in claim 1, characterized in that, Introducing the neighborhood radius R, the expression for the neighborhood kernel N is N=(R*2-1)*(R*2-1).

3. The point cloud noise rapid cleaning method as described in claim 2, characterized in that, Step S4 specifically includes: If the neighboring kernel The i-th noise point within With the m-th pixel image distance If the value is less than the neighborhood radius R, then for the neighborhood kernel... Noise count map Add 1 to the total.

4. The point cloud noise rapid cleaning method as described in claim 1, characterized in that, The procedure preceding step S1 also includes: Determine the m-th pixel image Is this a valid pixel image? If so, proceed to step S1; If not, then discard the m-th pixel image. .

5. The point cloud noise rapid cleaning method as described in claim 1, characterized in that, The procedure preceding step S1 also includes: Determine the type of point cloud; If the point cloud is an ordered point cloud, proceed to step S1; If the point cloud is an unordered point cloud, then the point cloud is rasterized to convert the unordered point cloud into an ordered point cloud before proceeding to step S1.

6. The point cloud noise rapid cleaning method as described in claim 1, characterized in that, The point cloud noise fast cleaning method is implemented using CUDA technology and shared memory technology.

7. A point cloud noise rapid cleaning device, characterized in that, The point cloud has an image of M pixels, and the point cloud noise rapid cleaning device includes: Building unit, used for the m-th pixel image Centered on the m-th pixel image, construct a model based on the m-th pixel. Neighborhood kernel , m=1, ...,M, where m and M are natural numbers greater than or equal to 1; Computational unit, used to compute the neighborhood kernel The i-th noise point within With the m-th pixel image distance Among them, the neighboring kernel There are G noisy points P in memory, i=1,...,G, where i and G are natural numbers greater than or equal to 1; The first judgment unit is used to judge the neighboring kernel. The i-th noise point within With the m-th pixel image distance , and the size of the preset neighborhood radius R; The counting unit is used to count the neighboring kernel based on the judgment result of the first judgment unit. The i-th noise point within Perform noise counting; The first traversal unit is used to traverse the neighboring kernel. For all noise points P within the kernel, obtain information about the neighborhood kernel. Noise count map ; The second judgment unit is used to judge the noise count map respectively. Each point within the image is related to the m-th pixel. The distance L and the size of the preset threshold T; the expression for the preset threshold T is T=(R*2-1)*(R*2-1) / 2; The marking unit is used to mark the noise point count map according to the judgment result of the second judgment unit. Mark each point within the area to determine whether it is a point to be cleaned; The cleanup unit is used to clean up neighboring kernels. Clean up the points marked as to be cleaned. The second traversal unit is used to traverse all pixel images of the point cloud; The marking unit is specifically used for: If noise count map Any point within the image and the m-th pixel If the distance L is less than the preset threshold T, then the noise count map will be... The current point within is marked as a point to be cleaned.

8. A computer-readable storage medium for storing a computer program, characterized in that: When the program is executed by the processor, it implements the point cloud noise rapid cleaning method as described in any one of claims 1 to 6.

Images