Bridge point cloud multi-view projection viewpoint screening method based on view angle consistency

Abstract

The invention discloses a bridge point cloud multi-view projection viewpoint screening method based on view angle consistency, and the method comprises the steps: 1), employing a rapid ground detection algorithm, and obtaining a visual prompt information proportion of an internal component of a visual point; 2) calculating a component visual rate based on the component visual prompt information proportion; 3) selecting the viewpoints with the component visibility greater than a preset threshold value, and calculating the viewing angle consistency rate of the selected viewpoints; and 4) screening out viewpoints with view angle consistency based on the view angle consistency rate. According to the method, visual information evaluation is carried out on the visual points which can be obtained by the viewpoints, and finally viewpoint screening is realized according to the consistency of the visual angles among the viewpoints, so that the viewpoints which can clearly present various components are obtained, and the method serves a downstream multi-view-based segmentation task.

Description

Technical Field

[0001] The invention relates to the technical field of information processing, and in particular to a bridge point cloud multi-view projection viewpoint screening method based on perspective consistency. Background Art

[0002] With the rapid development of bridge engineering and the increase in the number of aging bridges, bridge safety inspection and structural health monitoring have become increasingly important. Point cloud data, as a high-precision 3D measurement technology, has been widely used in bridge structure inspection and component identification. However, the complexity and massive amount of point cloud data make it challenging to directly extract effective bridge information from 3D bridge point clouds.

[0003] Therefore, using multi-view projection to obtain multiple views of point clouds and using a two-dimensional multi-view segmentation network or a large visual model to achieve rapid segmentation of bridge components has become a current research hotspot. However, existing point cloud multi-view projection methods only generate uniform annular or spherical viewpoints centered on the point cloud, without considering the screening of occluded viewpoints. These methods are difficult to deal with complex bridge structures, and are prone to problems such as view occlusion and unclear component identification, which affect the extraction of component information and the accuracy of subsequent segmentation. Therefore, a bridge point cloud multi-view projection viewpoint screening method that can ensure component visibility is needed. Summary of the invention

[0004] The purpose of the present invention is to provide a bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, comprising the following steps:

[0005] 1) Taking the center point of the bridge point cloud as the center, a fixed number of candidate viewpoints are uniformly generated;

[0006] 2) Voxelize the bridge point cloud and record the index and occupancy status of each voxel;

[0007] 3) Calculate the voxel center coordinates of the occupied voxels and record the correspondence between the occupied voxels and the points contained inside;

[0008] 4) Based on the ray traversal voxel algorithm, the visible voxels are confirmed and the visible points corresponding to each visible voxel are obtained;

[0009] 5) Use a fast ground detection algorithm to obtain the proportion of visual prompt information of components within the visible point;

[0010] 6) Calculate the component visibility rate based on the component visible prompt information ratio;

[0011] 7) Selecting a viewpoint whose component visibility is greater than a preset threshold, and calculating the viewing angle consistency rate of the selected viewpoint;

[0012] 8) Based on the perspective consistency rate, viewpoints with perspective consistency are screened out.

[0013] Furthermore, in step 1), the step of uniformly generating a fixed number of candidate viewpoints includes:

[0014] 1.1) Standardize the bridge point cloud;

[0015] 1.2) Taking the center point of the point cloud as the rotation center and the Z axis of the bridge point cloud as the central axis, at a distance of d units from the point cloud, set the elevation angle to 0° and the horizontal angle to 360°×(i-1) / N, and obtain N candidate viewpoints that evenly surround the point cloud. Calculate the three-dimensional coordinates u of each viewpoint based on the elevation angle and horizontal angle. i =(u ix ,u iy ,u iz ). i is the candidate viewpoint number.

[0016] Furthermore, in step 2), the steps of voxelizing the bridge point cloud and recording the index and occupancy status of each voxel include:

[0017] 2.1) Voxelize the standardized bridge point cloud and divide the point cloud into a voxelized grid G ​​with a voxel size of voxel_size;

[0018] 2.2) Record the index of each spatial voxel and occupancy status i ∈{0,1}; if o i =1, the voxel is considered to be occupied and there is a 3D point in the voxel; otherwise, i = 0, the voxel is considered to be an empty voxel and there is no 3D point in the voxel. ix,iy,iz are the coordinates of the index in the x, y, and z directions.

[0019] Furthermore, in step 3), the voxel center coordinates of the occupied voxel are v j =(v jx ,v jy ,v jz ), the correspondence between the occupied voxels and the points contained inside is vp j ={p k |p k ∈P and p k in v j}. P is the bridge point cloud.

[0020] Furthermore, in step 4), based on the ray traversal voxel algorithm, the steps of confirming the visible voxels and obtaining the visible points corresponding to each visible voxel include:

[0021] 4.1) Construct v with voxel as the starting point of the ray j =(v jx ,v jy ,v jz), with v j -u i is the direction of light, and r ji The current traversal voxel is initialized to

[0022] 4.2) Calculate the time interval Δt required for the light to advance one voxel along the three axes X =voxel_size / (v jx -u ix ), Δt Y =voxel_size / (v jy -u iy ), Δt Z =voxel_size / (v jz -u iz ), and initialize the time to enter the next voxel along the three-axis direction, that is, let T X ←Δt X 、T Y ←Δt Y 、T Z ←Δt Z ;

[0023] 4.3) According to the direction of the light, obtain the index increment of the voxel along the three axes Δx=sign(v jx -u ix ), Δy=sign(v jy -u iy )、Δz=sign(v jz -u iz );

[0024] If the voxel index of the current ray is within the voxel grid G, proceed to step 4.4), otherwise, obtain all voxels r that the ray passes through ji , and proceed to step 4.5);

[0025] 4.4) If min{T X ,T Y ,T Z =T X , then the X-axis index of the currently accessed voxel is updated to ix=ix+Δx, and the access time of a voxel under the X-axis is updated to T X =T X +Δt X , and add the current voxel information to the voxel set r that the light passes through ji ;;

[0026] If min{T X ,T Y ,T Z =T Y, then the Y-axis index of the currently accessed voxel is updated iy=iy+Δy, and the Y-axis voxel access time is updated T Y =T Y +Δt Y , and add the current voxel information to the voxel set r that the light passes through ji ;

[0027] If min{T X ,T Y ,T Z =T Z , then the Z-axis index of the currently accessed voxel is updated iz=iz+Δz, and the access time of a voxel under the Z axis is updated T Z =T Z +Δt Z , and add the current voxel information to the voxel set r that the light passes through ji , and repeat step 4.4);

[0028] 4.5) Count all voxels r that the light passes through ji The occupation status of Then the voxel is a visible voxel; if Then the voxel is an invisible voxel;

[0029] 4.6) Extract the points corresponding to all visible voxels as visible points P k .

[0030] Furthermore, in step 5), the step of using a fast ground detection algorithm to obtain the proportion of visual prompt information of components in the visible point includes:

[0031] 5.1) Divide the visible points into different sectors S in the XY plane s In Chinese, that is:

[0032]

[0033] Among them, n c =(x c -x k ,y c -y k ,z c -z k ), n i =(x i -x k ,y i -y k ,z i -z k );θ(n c ,n i ) is the angle between the two vectors, and its value range is [0,2π); Δα is the preset fan angle; (x c,y c , z c ) is the coordinate of the center point of the bridge point cloud; (x k ,y k , z k ) is the viewpoint coordinate; (x i ,y i , z i ) is the coordinate of any point in the bridge point cloud;

[0034] The points in a sector are as follows:

[0035] P ks ={p i ∈P k |s(p i )=s} (2)

[0036] 5.2) Divide the point cloud in the same sector into different blocks according to the distance between the point and the viewpoint In the division, if but

[0037] 5.3) Use the two-dimensional straight line fitting method to fit the two-dimensional representation of the lowest point of each block in the same sector to obtain the straight line line s ;

[0038] Among them, the two-dimensional representation of the lowest point of each block is as follows:

[0039]

[0040] 5.4) Calculate the distance from each point in each sector to the straight line s The distance is less than the threshold value, and the point is used as the component visual prompt information P kg .

[0041] Further, in step 6), the calculation formula of the component visibility is as follows:

[0042]

[0043] Where r is the visibility of the component, P kg is the number of points of component visual prompt information, P k is the number of visible points.

[0044] Further, in step 7), the viewpoint consistency rate calculation formula is as follows:

[0045]

[0046] In the formula, r threshold is the preset threshold. i 、ri-1 is the component visibility rate of viewpoints i and i-1; c is the viewing angle consistency rate.

[0047] Furthermore, in step 8), the step of selecting viewpoints with consistent viewing angles is as follows: based on the component consistency rate of each viewpoint, selecting the viewpoints with component consistency rate in the interval [c low ,c high ] as the final screening point U selected = {u j ,j=1,...,m}.

[0048] Furthermore, the screening points are used for multi-view projection of the bridge point cloud; and the multi-view projection of the bridge point cloud is used for semantic segmentation of the bridge point cloud, component identification and / or reconstruction of the completed bridge model.

[0049] The technical effect of the present invention is unquestionable. The present invention evaluates the visual information of the visible points that can be obtained by the viewpoints, and finally implements viewpoint screening based on the perspective consistency between viewpoints, obtains viewpoints that can clearly present various components, and serves the downstream multi-view based segmentation tasks.

[0050] The present invention reduces the computational cost of extracting visible points corresponding to viewpoints by voxelizing point clouds; extracting visible voxels and then extracting visible points effectively solves the problem of extracting visible points of viewpoints without semantic information; and obtains the proportion of component visual prompt information through fast ground detection algorithm detection, which can facilitate the calculation of viewpoint component visibility, and calculates the perspective consistency of the viewpoint based on the change of component visual prompt information of adjacent perspectives. By performing perspective consistency screening on viewpoints with certain component visibility, viewpoints that are not blocked by irrelevant information (such as retaining walls, etc.) and clearly provide component information can be screened out, which is beneficial to the improvement of subsequent segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of generating a fixed number of candidate viewpoints provided for an implementation example of this application;

[0052] Figure 2 A schematic diagram of bridge point cloud voxelization provided for the implementation example of this application;

[0053] Figure 3 A schematic diagram of the point cloud voxel occupancy state and visual state provided for the implementation example of this application;

[0054] Figure 4 A schematic diagram of viewpoints and certain voxel center rays provided for an implementation example of this application;

[0055] Figure 5 A schematic diagram of a certain viewpoint corresponding to a visible point provided for an implementation example of this application;

[0056] Figure 6 A schematic diagram of the detection result of the visual prompt information of the component in a certain viewpoint visual point provided for the implementation example of this application;

[0057] Figure 7 Schematic diagram of the bridge point cloud projection perspective screening results provided for the implementation of this application. DETAILED DESCRIPTION

[0058] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0059] Embodiment 1:

[0060] See also Figures 1 to 7 , a bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, comprising the following steps:

[0061] 1) Taking the center point of the bridge point cloud as the center, a fixed number of candidate viewpoints are uniformly generated;

[0062] 2) Voxelize the bridge point cloud and record the index and occupancy status of each voxel;

[0063] 3) Calculate the voxel center coordinates of the occupied voxels and record the correspondence between the occupied voxels and the points contained inside;

[0064] 4) Based on the ray traversal voxel algorithm, the visible voxels are confirmed and the visible points corresponding to each visible voxel are obtained;

[0065] 5) Use a fast ground detection algorithm to obtain the proportion of visual prompt information of components within the visible point;

[0066] 6) Calculate the component visibility rate based on the component visible prompt information ratio;

[0067] 7) Selecting a viewpoint whose component visibility is greater than a preset threshold, and calculating the viewing angle consistency rate of the selected viewpoint;

[0068] 8) Based on the perspective consistency rate, viewpoints with perspective consistency are screened out.

[0069] In step 1), the step of uniformly generating a fixed number of candidate viewpoints includes:

[0070] 1.1) Standardize the bridge point cloud;

[0071] 1.2) Taking the center point of the point cloud as the rotation center and the Z axis of the bridge point cloud as the central axis, at a distance of d units from the point cloud, set the elevation angle to 0° and the horizontal angle to 360°×(i-1) / N, and obtain N candidate viewpoints u that evenly surround the point cloud. i =(u ix ,u iy ,u iz ). i is the candidate viewpoint number. u ix ,u iy ,u iz are the coordinates in the x, y, and z directions;

[0072] In step 2), the steps of voxelizing the bridge point cloud and recording the index and occupancy status of each voxel include:

[0073] 2.1) Voxelize the standardized bridge point cloud and divide the point cloud into a voxelized grid G ​​with a voxel size of voxel_size;

[0074] 2.2) Record the index of each spatial voxel and occupancy status i ∈{0,1}; if o i =1, the voxel is considered to be occupied and there is a 3D point in the voxel; otherwise, i = 0, the voxel is considered to be an empty voxel and there is no 3D point in the voxel. ix,iy,iz are the coordinates of the index in the x, y, and z directions.

[0075] In step 3), the voxel center coordinates of the occupied voxel are v j=(v jx ,v jy ,v jz ), the correspondence between the occupied voxels and the points contained inside is vp j ={p k |p k ∈P and p k in v j}. P is the bridge point cloud.

[0076] In step 4), the steps of confirming the visible voxels and obtaining the visible points corresponding to each visible voxel include:

[0077] 4.1) Construct v with voxel as the starting point j =(v jx ,v jy ,v jz ), with v j -u i is the direction of the light, and the light r ji The current traversal voxel is initialized to

[0078] 4.2) Calculate the time interval Δt required for the light to advance one voxel along the three axes X =voxel_size / (v jx -u ix ), Δt Y =voxel_size / (v jy -u iy ), Δt Z =voxel_size / (v jz -u iz ), and initialize the time to enter the next voxel along the three-axis direction, that is, let T X ←Δt X 、T Y ←Δt Y 、T Z ←Δt Z ;

[0079] 4.3) According to the direction of the light, obtain the index increment of the voxel along the three axes Δx=sign(v jx -u ix ), Δy=sign(v jy -u iy )、Δz=sign(v jz -u iz );

[0080] If the voxel index of the current ray is within the voxel grid G, proceed to step 4.4), otherwise, obtain all voxels r that the ray passes through ji , and proceed to step 4.5);

[0081] 4.4) If min{T X ,T Y ,T Z =T X , then the X-axis index of the currently accessed voxel is updated to ix=ix+Δx, and the access time of a voxel under the X-axis is updated to T X =T X +Δt X , and add the current voxel information to the ray r ji ;;

[0082] If min{T X ,T Y ,T Z =T Y , then the Y-axis index of the currently accessed voxel is updated iy=iy+Δy, and the Y-axis voxel access time is updated T Y =T Y +Δt Y , and add the current voxel information to the ray r ji ;

[0083] If min{T X ,T Y ,T Z =T Z , then the Z-axis index of the currently accessed voxel is updated iz=iz+Δz, and the access time of a voxel under the Z axis is updated T Z =T Z +Δt Z , and add the current voxel information to the ray r ji , and return to step 4.2);

[0084] 4.5) Count all voxels r that the light passes through ji The occupation status of Then the voxel is a visible voxel; if Then the voxel is an invisible voxel;

[0085] 4.6) Extract the points corresponding to all visible voxels as visible points P k .

[0086] In step 5), the step of obtaining the visual prompt information ratio of the component in the visual point includes:

[0087] 5.1) Divide the visible points into different sectors S in the XY plane s In Chinese, that is:

[0088]

[0089] Among them, n c =(x c -x k ,y c -y k ,z c -z k ), n i =(x i -x k ,y i -y k ,z i -z k );θ(n c ,n i ) is the angle between the two vectors, and its value range is [0,2π); Δα is the preset fan angle; (x c ,y c , z c ) is the coordinate of the center point of the bridge point cloud; (x k ,y k , z k ) is the viewpoint coordinate; (x i ,y i , z i) is the coordinate of any point in the bridge point cloud;

[0090] The points in a sector are as follows:

[0091] P ks ={p i ∈P k |s(p i )=s} (2)

[0092] 5.2) Divide the point cloud in the same sector into different blocks according to the distance between the point and the viewpoint In the division, if but

[0093] 5.3) Use the two-dimensional straight line fitting method to fit the lowest point of each block in the same sector to obtain the straight line line s ;

[0094] Among them, the lowest points of each block are as follows:

[0095]

[0096] In the formula, take Calculate p′ at the lowest point of the z coordinate i .

[0097] 5.4) Calculate the distance from each point in each sector to the straight line s The distance is less than the threshold value, and the point is used as the component visual prompt information P kg .

[0098] In step 6), the component visibility is as follows:

[0099]

[0100] Where r is the visibility of the component.

[0101] In step 7), the view consistency rate of the viewpoint is as follows:

[0102]

[0103] In the formula, r threshold is the preset threshold. i 、r i-1 is the component visibility rate of viewpoints i and i-1; c is the viewing angle consistency rate.

[0104] In step 8), the step of selecting viewpoints with consistent viewing angles is as follows: based on the component consistency rate of each viewpoint, selecting the viewpoints with component consistency rate in the interval [c low ,c high] as the final screening point U selected = {u j ,j=1,...,m}.

[0105] The screening points are used for multi-view projection of bridge point cloud; and the multi-view projection of bridge point cloud is used for semantic segmentation of bridge point cloud, identification of bridge components and / or reconstruction of bridge completion model.

[0106] Embodiment 2:

[0107] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency includes the following steps:

[0108] 1) Taking the bridge point cloud as the center, a fixed number of candidate viewpoints are uniformly generated;

[0109] 2) Voxelize the bridge point cloud and record the index and occupancy status of each voxel;

[0110] 3) Calculate the voxel center coordinates of the occupied voxels and record the correspondence between the occupied voxels and the points contained inside;

[0111] 4) Based on the ray traversal voxel algorithm, the visible voxels are confirmed and the visible points corresponding to each visible voxel are obtained;

[0112] 5) Use a fast ground detection algorithm to obtain the proportion of visual prompt information of components within the visible point;

[0113] 6) Calculate the component visibility rate based on the component visible prompt information ratio;

[0114] 7) Selecting a viewpoint whose component visibility is greater than a preset threshold, and calculating the viewing angle consistency rate of the selected viewpoint;

[0115] 8) Based on the perspective consistency rate, viewpoints with perspective consistency are screened out.

[0116] Embodiment 3:

[0117] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as that of embodiment 2, further, in step 1), the step of uniformly generating a fixed number of candidate viewpoints includes:

[0118] 1.1) Standardize the bridge point cloud;

[0119] 1.2) Taking the center point of the point cloud as the rotation center and the Z axis of the bridge point cloud as the central axis, at a distance of d units from the point cloud, set the elevation angle to 0° and the horizontal angle to 360°×(i-1) / N, and obtain N candidate viewpoints u that evenly surround the point cloud. i =(u ix ,u iy ,u iz ).

[0120] Embodiment 4:

[0121] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as any one of embodiments 2-3, further, in step 2), the bridge point cloud is voxelized, and the step of recording the index and occupancy status of each voxel includes:

[0122] 2.1) Voxelize the standardized bridge point cloud and divide the point cloud into a voxelized grid G ​​with a voxel size of voxel_size;

[0123] 2.2) Record the index of each spatial voxel and occupancy status i ∈{0,1}; if o i =1, the voxel is considered to be occupied and there is a 3D point in the voxel; otherwise, i =0, the voxel is considered to be an empty voxel and there is no three-dimensional point in the voxel.

[0124] Embodiment 5:

[0125] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as any one of embodiments 2-4, further, in step 3), the voxel center coordinates of the occupied voxel are v j =(v jx ,v jy ,v jz ), the correspondence between the occupied voxels and the points contained inside is vp j ={p k |p k ∈P and p k in v j}.

[0126] Embodiment 6:

[0127] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as any one of embodiments 2-5, further, in step 4), the step of confirming the visible voxel and obtaining the visible point corresponding to each visible voxel includes:

[0128] 4.1) Construct v with voxel as the starting point j =(v jx ,v jy ,v jz ), with v j -u i is the direction of the light, and the light r ji The current traversal voxel is initialized to

[0129] 4.2) Calculate the time interval Δt required for the light to advance one voxel along the three axes X =voxel_size / (v jx -u ix ), Δt Y =voxel_size / (v jy -u iy ), Δt Z =voxel_size / (v jz -u iz ), and initialize the time to enter the next voxel along the three-axis direction, that is, let T X ←Δt X 、T Y ←Δt Y 、T Z ←Δt Z ;

[0130] 4.3) According to the direction of the light, obtain the index increment of the voxel along the three axes Δx=sign(v jx -u ix ), Δy=sign(v jy -u iy )、Δz=sign(v jz -u iz );

[0131] If the voxel index of the current ray is within the voxel grid G, proceed to step 4.4), otherwise, obtain all voxels r that the ray passes through ji , and proceed to step 4.5);

[0132] 4.4) If min{T X ,T Y ,T Z =T X , then the X-axis index of the currently accessed voxel is updated to ix=ix+Δx, and the access time of a voxel under the X-axis is updated to T X =T X +Δt X , and add the current voxel information to the ray r ji ;;

[0133] If min{T X ,T Y ,T Z =T Y , then the Y-axis index of the currently accessed voxel is updated iy=iy+Δy, and the Y-axis voxel access time is updated T Y =T Y +Δt Y , and add the current voxel information to the ray r ji ;

[0134] If min{T X ,T Y ,T Z =T Z , then the Z-axis index of the currently accessed voxel is updated iz=iz+Δz, and the access time of a voxel under the Z axis is updated T Z =T Z +Δt Z , and add the current voxel information to the ray r ji , and return to step 4.2);

[0135] 4.5) Count all voxels r that the light passes through ji The occupation status of Then the voxel is a visible voxel; if Then the voxel is an invisible voxel;

[0136] 4.6) Extract the points corresponding to all visible voxels as visible points P k .

[0137] Embodiment 7:

[0138] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as any one of embodiments 2-6, further, in step 5), the step of obtaining the visual prompt information ratio of the component in the visual point includes:

[0139] 5.1) Divide the visible points into different sectors S in the XY plane s In Chinese, that is:

[0140]

[0141] Among them, n c =(x c -x k ,y c -y k ,z c -z k ), n i =(x i -x k ,y i -y k ,z i -z k );θ(n c ,n i ) is the angle between two vectors, and its value range is [0,2π);

[0142] The points in a sector are as follows:

[0143] P ks ={p i ∈Pk |s(p i )=s} (2)

[0144] 5.2) Divide the point cloud in the same sector into different blocks according to the distance between the point and the viewpoint In the division, if but

[0145] 5.3) Use the two-dimensional straight line fitting method to fit the lowest point of each block in the same sector to obtain the straight line line s ;

[0146] Among them, the lowest points of each block are as follows:

[0147]

[0148] 5.4) Calculate the distance from each point in each sector to the straight line s The distance is less than the threshold value, and the point is used as the component visual prompt information P kg .

[0149] Embodiment 8:

[0150] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as any one of embodiments 2-7, further, in step 6), the component visibility is as follows:

[0151]

[0152] Where r is the visibility of the component.

[0153] Embodiment 9:

[0154] A bridge point cloud multi-view projection viewpoint screening method based on viewpoint consistency, the technical content is the same as any one of embodiments 2-8, further, in step 7), the viewpoint consistency rate of the viewpoint is as follows:

[0155]

[0156] In the formula, r threshold is the preset threshold.

[0157] Embodiment 10:

[0158] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the technical content is the same as any one of embodiments 2-9, further, in step 8), the step of screening out viewpoints with perspective consistency is: based on the component consistency rate of each viewpoint, screening out the component consistency rate in the interval [c low ,c high] as the final screening point U selected = {u j ,j=1,...,m}.

[0159] Embodiment 11:

[0160] A method for screening viewpoints of multi-view projection of bridge point clouds based on perspective consistency, the technical content of which is the same as any one of Examples 2-10. Furthermore, the screening points are used for multi-view projection of bridge point clouds, serving applications such as semantic segmentation of bridge point clouds, component identification, and reconstruction of bridge completion models.

[0161] Embodiment 12:

[0162] A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, the steps are as follows:

[0163] S101, taking the bridge point cloud as the center, uniformly generating a fixed number of candidate viewpoints;

[0164] S102, voxelizing the bridge point cloud, and recording the index and occupancy status of each voxel;

[0165] S103, calculating the voxel center coordinates of the occupied voxel, and recording the information of all points in the voxel;

[0166] S104, confirming visible voxels based on a ray traversal voxel algorithm, and obtaining visible points corresponding to each viewpoint;

[0167] S105, using a fast ground detection algorithm to obtain the proportion of visual prompt information of components in the visible point;

[0168] S106, calculating a component visibility rate based on the detected component visibility prompt information ratio;

[0169] S107, selecting viewpoints whose component visibility is greater than a certain threshold, and calculating the viewing angle consistency rate of each viewpoint;

[0170] S108. Screen out viewpoints with consistent perspectives according to the perspective consistency rate obtained in S107.

[0171] In this application, by voxelizing the point cloud, the computational cost of extracting the visible points corresponding to the viewpoints is reduced; by extracting the visible voxels, the visible points are extracted, which effectively solves the problem of extracting the visible points of the viewpoints without semantic information; by detecting with a fast ground detection algorithm, the proportion of the component visual prompt information is obtained, which can facilitate the calculation of the visibility of the viewpoint components, and at the same time, the perspective consistency of the viewpoint is calculated based on the changes in the component visual prompt information of the adjacent perspectives. By performing perspective consistency screening on viewpoints with certain component visibility, viewpoints that are not blocked by irrelevant information (such as retaining walls, etc.) and clearly provide component information can be screened out, which is beneficial to the improvement of subsequent segmentation accuracy.

[0172] S101, taking the bridge point cloud as the center, uniformly generating a fixed number of candidate viewpoints;

[0173] In the specific implementation, the bridge point cloud is first standardized, and the center point of the point cloud is used as the rotation center, the Z axis of the bridge point cloud is used as the central axis, and the elevation angle is set to 0° at a distance of d = 2 units from the point cloud. The horizontal angle is set to 360°×(i-1) / 20 in turn, and 20 candidate viewpoints u that evenly surround the point cloud are obtained. i =(u ix ,u iy ,u iz ),like Figure 1 shown.

[0174] S102, voxelizing the bridge point cloud, and recording the index and occupancy status of each voxel;

[0175] In the specific implementation, the standardized bridge point cloud is voxelized and the point cloud is divided into a voxelized grid G ​​with a voxel size of voxel_size = 0.01. The index of each spatial voxel is recorded and occupancy status i ∈{0,1}; if o i =1, the voxel is considered to be occupied and contains a three-dimensional point; otherwise, i = 0, the voxel is considered to be an empty voxel with no 3D point inside. Figure 2 Shows the implementation example bridge point cloud occupancy voxels.

[0176] S103, calculating the voxel center coordinates of the occupied voxel, and recording the information of all points in the voxel;

[0177] In specific implementation, for each occupied voxel, calculate its voxel center coordinate v j =(v jx ,v jy ,v jz ), and record the corresponding relationship vp between the occupied voxels and the points they contain j ={p k|p k ∈P and p k in v j}.

[0178] S104, confirming visible voxels based on a ray traversal voxel algorithm, and obtaining visible points corresponding to each viewpoint;

[0179] In specific implementation, for each viewpoint u i =(u ix ,u iy ,u iz ), and use the ray traversal voxel algorithm to extract the corresponding visible voxels, thereby realizing the visible point extraction.

[0180] Visual voxel extraction: Assume there is a viewpoint u i =(u ix ,u iy ,u iz ), construct its corresponding occupied voxel v j =(v jx ,v jy ,v jz ) between the two.

[0181] Starting from voxel v j =(v jx ,v jy ,v jz ), with v j -u i is the direction of the light, and the light r ji The current traversal voxel is initialized to Calculate the time interval Δt required for the light to advance one voxel along the three axes respectively X =0.01 / (v jx -u ix ), Δt Y =0.01 / (v jy -u iy ), Δt Y =0.01 / (v jz -u iz ), and initialize the time T for the current three directions to enter the next voxel X ←Δt X 、T Y ←Δt Y 、T Z ←Δt Z At the same time, according to the direction of the light, the index increment of the voxel along the three axes is obtained. jx -u ix ), Δy=sign(v jy -u iy)、Δz=sign(v jz -u iz ).

[0182] If the voxel index of the current ray is still within the voxel grid G, compare T X , T Y , T Z , get the coordinates of the next voxel to be visited, as follows:

[0183] If min{T X ,T Y ,T Z =T X , then the X-axis index of the currently accessed voxel is updated to ix=ix+Δx, and the access time of a voxel under the X-axis is updated to T X =T X +Δt X , and add the current voxel information to the ray r ji ;

[0184] If min{T X ,T Y ,T Z =T Y , then the Y-axis index of the currently accessed voxel is updated iy=iy+Δy, and the Y-axis voxel access time is updated T Y =T Y +Δt Y , and add the current voxel information to the ray r ji ;

[0185] If min{T X ,T Y ,T Z =T Z , then the Z-axis index of the currently accessed voxel is updated iz=iz+Δz, and the access time of a voxel under the Z axis is updated T Z =T Z +Δt Z , and add the current voxel information to the ray r ji

[0186] Repeat the above process of ray access voxel acquisition until the voxel where the current ray is located leaves the voxel grid G, and obtain all voxels r that the ray passes through ji .

[0187] Count all voxels r that the ray passes through ji The occupation status of Then the voxel is a visible voxel; if Then the voxel is an invisible voxel.

[0188] Visible point extraction: For each viewpoint u i =(uix ,u iy ,u iz ), use the above-mentioned visible voxel extraction method to judge each occupied speed, and extract the points corresponding to all visible voxels as the visible point P of the viewpoint k .

[0189] S105, using a fast ground detection algorithm to obtain the proportion of visual prompt information of components in the visible point;

[0190] In specific implementation, for each viewpoint’s visible point P k ,A rapid ground detection method is used to obtain the proportion of visual prompt information of bridge components within the visible points.

[0191] First, according to formula (1), the visible points are divided into different sectors S in the XY plane. s middle:

[0192]

[0193] Among them, n c =(x c -x k ,y c -y k ,z c -z k ), n i =(x i -x k ,y i -y k ,z i -z k );θ(n c ,n i ) is the angle between two vectors, and its value range is [0,2π). Therefore, a point in a sector can be recorded as:

[0194] P ks ={p i ∈P k |s(p i )=s} (2)

[0195] Divide the point cloud in the same sector into different blocks according to the distance between the point and the viewpoint In Then p i ∈P ks .

[0196] At the same time, the lowest point of each block is selected and expressed as follows:

[0197]

[0198] Use the two-dimensional straight line fitting method to fit the lowest point of each block in the same sector to obtain the straight line s . Calculate the distance from each point in each sector to the straight line s The point whose distance is less than the threshold value 0.01 is considered as the component visual prompt information P kg .

[0199] S106, calculating a component visibility rate based on the detected component visibility prompt information ratio;

[0200] In specific implementation, for each visible point of each viewpoint, based on the component visible prompt information obtained in S105, the component visibility rate is calculated using formula (4):

[0201]

[0202] S107, selecting viewpoints whose component visibility is greater than a certain threshold, and calculating the viewing angle consistency rate of each viewpoint;

[0203] In specific implementation, select the component visibility greater than r threshold = 0.2, the view consistency rate of the viewpoint is calculated using formula (5):

[0204]

[0205] S108. Screen out viewpoints with consistent perspectives according to the perspective consistency rate obtained in S107.

[0206] In the specific implementation, according to the component consistency rate of each viewpoint calculated in S107, the viewpoints whose component consistency rate is within the interval [0.2, 5] are selected as the final selection points U selected = {u j ,j=1,...,m}.

Claims

1. A bridge point cloud multi-view projection viewpoint screening method based on perspective consistency, characterized in that: The following steps are involved: 1) Taking the center point of the bridge point cloud as the center, a fixed number of candidate viewpoints are uniformly generated; 2) Voxelize the bridge point cloud and record the index and occupancy status of each voxel; 3) Calculate the voxel center coordinates of the occupied voxels and record the correspondence between the occupied voxels and the points contained inside; 4) Based on the ray traversal voxel algorithm, the visible voxels are confirmed and the visible points corresponding to each visible voxel are obtained; 5) Use a fast ground detection algorithm to obtain the proportion of visual prompt information of components within the visible point; 6) Calculate the component visibility rate based on the component visible prompt information ratio; 7) Selecting a viewpoint whose component visibility is greater than a preset threshold, and calculating the viewing angle consistency rate of the selected viewpoint; 8) Based on the perspective consistency rate, viewpoints with perspective consistency are screened out.

2. According to the method for selecting viewpoints of bridge point cloud multi-view projection based on perspective consistency according to claim 1, it is characterized in that: In step 1), the step of uniformly generating a fixed number of candidate viewpoints includes: 1.1) Standardize the bridge point cloud; 1.2) Taking the center point of the point cloud as the center and the Z axis of the bridge point cloud as the center axis, at a distance of d units from the point cloud, set the elevation angle to 0° and the horizontal angle to 360°×(i-1) / N, and obtain N candidate viewpoints that evenly surround the point cloud. Calculate the three-dimensional coordinates u of each viewpoint based on the elevation angle and horizontal angle. i =(u ix ,u iy ,u iz ), i is the candidate viewpoint number, u ix ,u iy ,u iz are the coordinates in the x, y, and z directions.

3. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 2), the steps of voxelizing the bridge point cloud and recording the index and occupancy status of each voxel include: 2.1) Voxelize the standardized bridge point cloud and divide the point cloud into a voxelized grid G ​​with a voxel size of voxel_size; 2.2) Record the index of each spatial voxel and occupancy status i ∈{0,1}; if o i =1, the voxel is considered to be occupied and there is a 3D point in the voxel; otherwise, i =0, the voxel is considered to be an empty voxel and there is no three-dimensional point in the voxel; ix,iy,iz are the coordinates of the index in the x, y, and z directions.

4. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 3), the voxel center coordinates of the occupied voxel are v j =(v jx ,v jy ,v jz ), the correspondence between the occupied voxels and the points contained inside is vp j ={p k |p k ∈P and p k in v j }; P is the bridge point cloud.

5. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 4), based on the ray traversal voxel algorithm, the steps of confirming the visible voxels and obtaining the visible points corresponding to each visible voxel include: 4.1) Construct a voxel center v j =(v jx ,v jy ,v jz ) is the starting point of the light, with v j -u i is the direction of light, and r ji The current traversal voxel is initialized to Candidate point u i =(u ix ,u iy ,u iz ); 4.2) Calculate the time interval Δt required for the light to advance one voxel along the three axes X =voxel_size / (v jx -u ix ), Δt Y =voxel_size / (v jy -u iy ), Δt Z =voxel_size / (v jz -u iz ), and initialize the time to enter the next voxel along the three-axis direction, that is, let T X ←Δt X 、T Y ←Δt Y 、T Z ←Δt Z ; 4.3) According to the direction of light, obtain the index increment of the voxel along the three axes Δx=sign(v jx -u ix ), Δy=sign(v jy -u iy )、Δz=sign(v jz -u iz );u ix ,u iy ,u iz are the coordinates of the candidate point in the x, y, and z directions; If the voxel index of the current ray is within the voxel grid G, proceed to step 4.4), otherwise, obtain all voxels r that the ray passes through ji , and proceed to step 4.5); 4.4) If min{T X ,T Y ,T Z =T X , then the X-axis index of the currently accessed voxel is updated to ix=ix+Δx, and the access time of a voxel under the X-axis is updated to T X =T X +Δt X , and add the current voxel information to the voxel set r that the light passes through ji ;; If min{T X ,T Y ,T Z =T Y , then the Y-axis index of the currently accessed voxel is updated iy=iy+Δy, and the Y-axis voxel access time is updated T Y =T Y +Δt Y , and add the current voxel information to the voxel set r that the light passes through ji ; If min{T X ,T Y ,T Z =T Z , then the Z-axis index of the currently accessed voxel is updated iz=iz+Δz, and the access time of a voxel under the Z axis is updated T Z =T Z +Δt Z , and add the current voxel information to the voxel set r that the light passes through ji , and repeat step 4.4); 4.5) Count all voxels r that the light passes through ji The occupation status of Then the voxel is a visible voxel; if Then the voxel is an invisible voxel; 4.6) Extract the points corresponding to all visible voxels as visible points P k .

6. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 5), the steps of using a fast ground detection algorithm to obtain the proportion of visual prompt information of components in the visible point include: 5.1) Divide the visible points into different sectors S in the XY plane s In Chinese, that is: Among them, n c =(x c -x k ,y c -y k ,z c -z k ), n i =(x i -x k ,y i -y k ,z i -z k );θ(n c ,n i ) is the angle between two vectors, and its value range is [0,2π); s(p i ) is the visible point p i The sector to which it belongs; Δα is the preset sector angle; (x c ,y c , z c ) is the coordinate of the center point of the bridge point cloud; (x k ,y k , z k ) is the viewpoint coordinate; (x i ,y i , z i ) is the coordinate of any point in the bridge point cloud; The points in a sector are as follows: P ks ={p i ∈P k |s(p i )=s} (2) Where P ks is a set of points in a sector; 5.2) Divide the point cloud in the same sector into different blocks according to the distance between the point and the viewpoint In the division, if but 5.3) Use the two-dimensional straight line fitting method to fit the two-dimensional representation of the lowest point of each block in the same sector to obtain the straight line line s ; Among them, the two-dimensional representation of the lowest point of each block is as follows: 5.4) Calculate the distance from each point in each sector to the straight line s The distance is less than the threshold value, and the point is used as the component visual prompt information P kg .

7. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 6), the calculation formula of component visibility is as follows: Where r is the visibility of the component, P kg is the number of points of component visual prompt information, P k is the number of visible points.

8. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 7), the viewpoint consistency rate calculation formula is as follows: In the formula, r threshold is the preset threshold; r i 、r i-1 is the component visibility rate of viewpoints i and i-1; c is the viewing angle consistency rate.

9. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1 is characterized in that: In step 8), the step of selecting viewpoints with consistent viewing angles is as follows: based on the component consistency rate of each viewpoint, selecting the viewpoints with component consistency rate in the interval [c low ,c high ] as the final screening point U selected = {u j ,j=1,...,m}.

10. The bridge point cloud multi-view projection viewpoint screening method based on perspective consistency according to claim 1, characterized in that: The screening points are used for multi-view projection of the bridge point cloud; the multi-view projection of the bridge point cloud is used for semantic segmentation of the bridge point cloud, component identification and / or reconstruction of the completed bridge model.

Images