Three-dimensional visualization method of underground pipe wells based on RGB-D depth camera reconstruction
Through the three-dimensional visualization method based on RGB-D depth camera, the problems of missing three-dimensional information and inefficiency in underground well chamber measurement are solved, and the three-dimensional visualization and quantitative analysis of underground well chamber are realized, reducing costs and improving measurement accuracy and efficiency .
Patent Information
- Application Number
- CN202110371345.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-04-07
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2041-04-07
AI Technical Summary
The existing underground well chamber measurement methods have problems such as lack of three-dimensional information, inefficiency and high cost. Traditional two-dimensional images cannot meet the visualization requirements of underground well chambers and pipelines. Manual surveying and mapping are dangerous and costly, and 2D images lose depth information, 24 The underground robot has a complex structure, high cost and poor applicability.
Using a three-dimensional visualization method based on RGB-D depth camera, the image data frame is acquired, preprocessed and fusion is performed, and the initial position information and three-dimensional point cloud are calculated, point cloud registration is realized at different perspectives, and the reconstruction of underground pipe wells is obtained. Point clouds and surface reconstruction and texture mapping are carried out to improve measurement accuracy and efficiency.
The three-dimensional visualization of the underground well chamber is realized, and a three-dimensional measured point cloud model is provided, which supports quantitative analysis and disease detection of underground space, reduces measurement costs, and improves measurement efficiency and accuracy.
Smart Images

Figure CN113112588B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a three-dimensional visualization method for underground pipe wells based on RGB-D depth camera reconstruction, and belongs to the technical field of underground pipe well mapping. Background Art
[0002] With the increasing informatization of underground spaces, traditional two-dimensional image and video representation methods are no longer able to meet the visualization requirements of underground wells and pipelines. Traditional underground well measurements often rely on manual mapping, 2D imaging, and 24-mesh downhole robot measurement. However, manual mapping presents drawbacks such as high risk, high labor costs, and difficulty in ensuring accuracy. Ordinary 2D images lose depth information and cannot accurately measure three-dimensional scale information. Furthermore, the limited imaging area makes it impossible to obtain a complete view of the well. 24-mesh downhole robots are complex, bulky, and have poor applicability. Furthermore, the equipment requires calibration in a specific calibration room before use, which is cumbersome and costly. Overall, existing underground well measurement methods suffer from the lack of three-dimensional information, low efficiency, and high cost. Summary of the Invention
[0003] Based on the above, the present invention provides a three-dimensional visualization method for underground pipe wells based on RGB-D depth camera reconstruction to solve the technical problems of existing underground well chamber measurement methods such as lack of three-dimensional information, low efficiency and high cost.
[0004] The technical solution of the present invention is: a three-dimensional visualization method of underground pipe wells based on RGB-D depth camera reconstruction, wherein the method comprises:
[0005] Acquire image data frames of the underground pipe well from different viewing angles using an RGB-D depth camera, wherein the image data frames include target depth images and target RGB images;
[0006] Preprocessing the target depth image to obtain a corrected target depth image;
[0007] Fusing the corrected target depth image and the target RGB image under the same viewing angle to obtain a fused image data frame;
[0008] According to the fused image data frames, obtaining initial pose information between adjacent fused image data frames and fusing a three-dimensional point cloud;
[0009] Based on the initial pose information and the fused three-dimensional point cloud, registration of point clouds at different viewing angles is achieved to obtain a reconstructed three-dimensional point cloud of the underground pipe well;
[0010] The reconstructed three-dimensional point cloud is processed to obtain a true three-dimensional visual model of the underground pipe well.
[0011] Optionally, before the RGB-D depth camera takes pictures, the depth measurement error of the RGB-D depth camera is corrected based on a polynomial surface fitting method.
[0012] Optionally, the method for preprocessing the target depth image is: based on a joint bilateral filtering algorithm, using the target RGB image as a guide image to achieve fast filtering and denoising of the target depth image.
[0013] Optionally, the method for calculating the initial pose information is: matching adjacent fused image data frames according to a SIFT algorithm to obtain the initial pose information.
[0014] Optionally, an ICP algorithm based on neighborhood features is used to process the initial pose information and the fused three-dimensional point cloud to achieve registration of point clouds under different viewing angles.
[0015] Optionally, the method for processing the reconstructed three-dimensional point cloud includes:
[0016] Performing surface reconstruction on the reconstructed three-dimensional point cloud to obtain a surface reconstruction result;
[0017] According to the color image data of the target RGB image, texture mapping is performed on the surface reconstruction result.
[0018] The beneficial effects of the present invention are: the present invention performs visual three-dimensional reconstruction of the underground well chamber based on the RGB-D depth camera, realizes the three-dimensional visualization of its internal environmental conditions, and conducts quantitative analysis of the defect conditions of the underground space based on the three-dimensional measured point cloud model, which can provide technical support for the detection and maintenance of underground space diseases. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is the principle of Kinect triangulation;
[0020] Figure 2 Fitting the distance curve for Kinect ranging error;
[0021] Figure 3 is the depth error correction result;
[0022] Figure 4 is the original matching result;
[0023] Figure 5 This is the matching result after purification by the improved RANSAC algorithm;
[0024] Figure 6 It is a pinhole camera model;
[0025] Figure 7 Provides accurate registration process for ICP point clouds;
[0026] Figure 8 Accurately register ICP point clouds;
[0027] Figure 9 This is the point cloud result after global registration;
[0028] Figure 10 Poisson surface reconstruction results of underground well chamber
[0029] Figure 11 Texture model of an underground well chamber. DETAILED DESCRIPTION
[0030] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0032] See also Figures 1 to 11 , a method for three-dimensional visualization of underground pipe wells based on RGB-D depth camera reconstruction in an embodiment of the present invention, wherein the method comprises:
[0033] S1 acquires image data frames of the underground pipe well from different viewing angles using an RGB-D depth camera, wherein information of the image data frames includes a target depth image and a target RGB image;
[0034] Conventional laser scanners struggle to operate in underground wells, due to their confined space and monotonous texture. To address the needs of underground well applications, this embodiment uses a Kinect v2.0 depth camera based on the ToF ranging principle as the image acquisition device. This depth camera can acquire depth image data of the measured target under any lighting conditions, making it well suited for underground environments and crucial for 3D reconstruction of underground spaces.
[0035] To capture image data frames, the depth camera was first placed in the underground well chamber using a mechanical transmission mechanism. It was then rotated to acquire depth and color data from three different viewpoints: top, center, and bottom. The default rotation count was set to 16, with each rotation angle of 20 degrees. Data was collected from all three levels of the well chamber. During acquisition, the depth camera's measurement distance was maintained at least 1 meter to ensure optimal data quality. The acquisition range was set to 0.6-7.5 meters to ensure data validity. Because the Kinect v2.0 depth sensor has a field of view of 70° horizontally and 60° vertically, its vertical and horizontal coverage is limited. Therefore, during the actual measurement, the Kinect v2.0 depth camera was tilted upward and downward at specific angles, performing multiple scans to increase the camera's coverage. This ensured a high degree of overlap between the image data collected after one rotation.
[0036] In order to ensure the accuracy of the depth measurement of the Kinect depth camera, the depth measurement error of the Kinect depth camera is corrected based on the polynomial surface fitting method before the Kinect depth camera takes pictures.
[0037] Specifically, a depth measurement error correction method based on polynomial surface fitting is proposed for the Kinect depth camera in different directions. The triangulation principle of the Kinect depth camera can be Figure 1 The schematic diagram shown is used to represent it. According to the similarity of triangles, we can get:
[0038]
[0039] Among them, Z o is the reference plane depth value, f is the focal length of the infrared camera, b is the baseline length, and d is the parallax. The depth of the target to be measured is obtained from the above formula:
[0040]
[0041] If Δd is the parallax error, then the depth measurement error ΔZ k It can be expressed as follows:
[0042]
[0043] The parallax error is determined by the system error. The system error is fixed, so the parallax error is also constant. From formula (3), we can see that the depth measurement error is proportional to the measured distance. Through the error model of the above RGB-D depth camera triangulation, the true depth value F(u, v) of each pixel can be obtained with respect to the RGB-D depth camera measurement distance Z k The polynomial surface model is:
[0044]
[0045] Where a1(u, v),…,a n (u, v) represents the polynomial coefficients, and b0(u, υ) represents the constant term.
[0046] The experiment analyzed the ranging results of the Kinect depth camera and the laser ranging results, and obtained multiple sets of depth measurement errors and their changes within the Kinect ranging field of view. Using the acquired depth error data, a polynomial surface model of the true depth value with respect to the measured distance was constructed. The obtained measurement error data was fitted based on the principle of least squares, and a fitting formula for the measured depth with respect to the depth error was obtained. The fitting results are shown in the figure below. Figure 2 shown.
[0047] Figure 2 The depth error result in is fitted by the surface function model to obtain the following fitting formula:
[0048] y=4E-10x 3 -3E-6x 2 +0.01x (5)
[0049] Where y is the error correction value and x is the distance value. If the depth value of Kinect v2.0 is d, the measured and corrected depth error is:
[0050] d′=d-(4E-10x 3 -3E-6x 2 +0.01x) (6)
[0051] The depth error is corrected by the error correction equation of polynomial surface fitting, and the corrected error distribution result is obtained, such as Figure 3 As shown. Figure 3 Experimental results show that the original depth error increases rapidly with increasing measurement distance, while the depth error after error compensation changes more gradually with increasing distance. Furthermore, the Kinect depth camera's depth measurement error is less than 2cm within a range of 0.5m to 4.5m; within a range of 4.5m to 7m, the depth measurement error is maintained within 4.5cm. Experimental results demonstrate that the error correction method proposed in this paper can effectively correct the errors in Kinect depth measurement data and effectively improve its ranging accuracy.
[0052] S2 preprocesses the target depth image to obtain a corrected target depth image;
[0053] Preferably, based on the joint bilateral filtering algorithm, the target RGB image is used as a guide image to achieve fast filtering and denoising of the target depth image.
[0054] Specifically, the joint bilateral filter is an improved algorithm based on the bilateral filter that can be used to upsample low-resolution images, thereby filling in missing information in the image. In the bilateral filtering algorithm, all weights are calculated from a single image, while in the joint bilateral filtering algorithm, better weights can be obtained by introducing another image with richer information to calculate the weights. An RGB-D camera can simultaneously acquire a target depth image and an RGB image. The RGB color image contains complete information about the scene and can be used to fill in the missing parts of the depth image. Therefore, based on the joint bilateral filtering algorithm, this embodiment uses the RGB image as a guide image to achieve rapid filtering and denoising of the depth image.
[0055] The weight function of joint bilateral filtering is expressed as follows:
[0056]
[0057] Where (i, j) and (x, y) represent the coordinates of the two selected pixels, and the grayscale domain weight of the color image w r and the depth image spatial domain weight w g The calculation formula is as follows:
[0058]
[0059]
[0060] Where g(i, j) and g(x, y) represent the grayscale values obtained by converting the color image into a grayscale image at the pixel points (i, j) and (x, y), respectively.
[0061] S3 performs fusion processing on the corrected target depth image and the target RGB image under the same viewing angle to obtain a fused image data frame;
[0062] The fusion of depth images and RGB images is common knowledge to those skilled in the art and will not be described in detail here.
[0063] S4 obtains initial pose information between adjacent fused image data frames and fuses a three-dimensional point cloud based on the fused image data frames;
[0064] Preferably, adjacent fused image data frames are matched according to a SIFT algorithm to obtain initial pose information.
[0065] Depth cameras can quickly acquire depth and color information of scene objects, generating color point cloud data. However, due to the limited viewing angle of depth cameras, the large size of scene objects, and obstructions, scanning from a single angle can only capture partial data about the target. Therefore, multiple data acquisitions from different angles are necessary to obtain a global point cloud of the scene objects. The coordinate systems of point cloud data obtained from different perspectives are inconsistent, requiring point cloud registration to unify them into a single coordinate system.
[0066] The scale-invariant feature transform (SIFT) algorithm can detect and describe regional features in an image. Its principle is to detect key points in established spaces of different scales and calculate their directions. SIFT mainly detects key points with obvious information, and is also more effective in detecting occluded targets. SIFT features have a large amount of information and are suitable for fast and accurate matching in massive databases [i]. After obtaining the SIFT feature points of the image, feature matching is performed on two adjacent frames to obtain the posture transformation relationship between the two frames. The traditional RANSAC algorithm uses randomly selected feature points for matching, ignoring the differences in feature points, which can easily cause image mismatching. The improved RACSAC algorithm is based on the idea of iteration. It first screens the results of random sampling in the iterative process to eliminate matching point pairs with obvious errors. The principle is as follows:
[0067] 1) Assume that the set of ORB feature points extracted from two adjacent frames of images is:
[0068] P={p i |p i ∈P, i=1, 2, …, m)
[0069] Q={q i |q i ∈Q,i=1,2,…,m) (10)
[0070] Find the feature point p in P i The two feature points of the nearest neighbor and the next nearest neighbor in Q. If the ratio of the Euclidean distance to the nearest neighbor and the next nearest neighbor is less than the set threshold, then p is considered i The nearest neighbor ORB feature point is a pair of matching points.
[0071] 2) Using a bidirectional matching mechanism, find the feature points in P that correspond to the feature points in Q, and similarly find the points in Q that correspond to the feature points in P. If the two are a one-to-one match, the match is considered correct, and incorrect matches are discarded to achieve the purpose of screening.
[0072] 3) By setting a minimum matching distance threshold, the remaining matching results that are greater than a certain multiple of the minimum matching distance are eliminated. After the above steps, matching point pairs with high matching quality are obtained.
[0073] The experiment uses the SIFT feature point algorithm to match the acquired image data and obtain the initial pose estimation information. This provides a better calculation initial value for the subsequent ICP point cloud precise registration, further improving the registration accuracy of the point cloud image. Figure 4 The figure shows the original matching effect of SIFT feature points of underground well chamber images. There are many mismatches, which is not conducive to the initial pose estimation between image frames. In order to improve the accuracy of initial pose estimation, the improved RANSAC algorithm is used to process the feature matching results and eliminate the mismatch results. Figure 5 The experiment selected one set of data for statistics and obtained the matching results shown in Table 1. The experimental results show that the matching rate of feature points after processing by the improved RANSAC algorithm is significantly improved, and the correct matching rate can reach 95.24%, effectively eliminating the wrong matching point pairs.
[0074] Table 1 Statistics of SIFT feature point matching results
[0075]
[0076] In the present invention, the method for obtaining a fused 3D point cloud from a fused image data frame is based on the projection principle of a pinhole camera model. The pinhole camera model is a commonly used camera model, such as Figure 3-6 As shown. Where P(X C , Y C , Z C ) represents a coordinate point in the camera coordinate system, and its projection point on the image plane is P′(x, y).
[0077] Depend on Figure 6 The similarity relationship in can be obtained:
[0078]
[0079] Where f is the distance from O to O1. Expressing equation (11) in the form of homogeneous coordinates and matrices yields:
[0080]
[0081] Simplifying, we can get:
[0082]
[0083] Where (X w , Y w , Z w) represents the coordinates of a spatial point in the real scene in the world coordinate system, and (u, υ) represents its coordinates in the pixel coordinate system, with the unit being pixel. x =f / d x , a y =f / d y ; M is a 3×4 projection matrix, M1 is the intrinsic parameter a of the Kinect depth camera x , a y , u0, v0 determine, (u0, v0) is the coordinate of the main point, a x and a y They represent the scale factors on the u-axis and v-axis in the pixel coordinate system respectively. M2 is obtained from the external parameters of the Kinect depth camera.
[0084] From formula (13), we can see that if the internal and external parameters of the camera are known, the projection matrix M can also be determined. When the coordinates of the space point P and M are known, the vector The coordinates (u, v) of its projection point p can be obtained from formula (13).
[0085] S5 realizes registration of point clouds at different viewing angles based on the initial pose information and the fused three-dimensional point cloud, thereby obtaining a reconstructed three-dimensional point cloud of the underground pipe well;
[0086] Preferably, an ICP algorithm based on neighborhood features is used to process the initial pose information and the fused three-dimensional point cloud to achieve registration of point clouds under different viewing angles.
[0087] After matching the image data using the SIFT algorithm, the initial pose information between the image data frames is obtained. Based on the feature point matching, the iterative nearest neighbor point (ICP) algorithm based on the neighborhood feature is further used to achieve accurate pose estimation between the three-dimensional point clouds, thereby achieving accurate point cloud registration. The algorithm flow is as follows: Figure 6 shown.
[0088] First, establish the neighborhood features, the principle is as follows:
[0089] a. Select feature points: Assume r is the sampling point p i A neighborhood, p ij Indicates p i At all neighboring points in the domain, with curvature h ij , then the neighborhood curvature value of the current sampling point It can be expressed as follows:
[0090]
[0091] Where n is the number of neighboring points of the sampling point. The variance of the point cloud σ N and mean μ N The calculation is as follows:
[0092]
[0093]
[0094] In formula (10), N is the number of sampling points of the point cloud, and the neighborhood curvature of the sampling point is in μ N ±ασ N Points other than α are selected as feature points. α is an adjustment factor used to adjust the number of feature points.
[0095] b. Determine the corresponding points: Assume that P and Q are two adjacent point clouds. To find the feature point p in point cloud P i The corresponding points in the point cloud Q can be formed by selecting the three nearest points in Q to form a triangle.
[0096] c. Filter corresponding point pairs: There are some mismatched point pairs among the selected corresponding point pairs. You can filter them by the neighborhood features of the points to remove the wrong corresponding point pairs. The neighborhood features are described as Where n is the normal vector, It is the weighted average of the curvatures of neighboring points within the neighborhood of the sampling point.
[0097] After the domain features are established, the ICP algorithm is used to accurately estimate the pose between two adjacent frames of point clouds based on the established domain features. Assume that P and P' are already matched point pairs:
[0098] P={p1,…,p n},
[0099] P′=(p′1,…,p′ n} (16)
[0100] That is, find a Euclidean transformation R,t such that:
[0101]
[0102] Based on the initial pose estimation value provided by SIFT feature matching, the ICP algorithm combined with domain features can achieve accurate registration of local inter-frame point clouds and global point clouds. The inter-frame point cloud registration results are as follows: Figure 7 As shown in the figure, we can see that after the ICP algorithm accurately registers the two point clouds, a good splicing fusion is achieved, and the details of the local registration are also well performed. After the precise registration of the two point clouds, only the local point cloud registration is achieved, and the global point cloud is not fused. Further, the multi-viewpoint point cloud is fused by the ICP precise registration algorithm based on neighborhood features to obtain a globally consistent point cloud reconstruction registration result, as shown in the figure. Figure 8The experiment statistically analyzed the precise registration results of the two point clouds and the global point cloud, and compared them with the classic ICP algorithm. The registration results of the two algorithms are shown in Table 2. The experimental results show that the proposed method outperforms the classic ICP algorithm in the precise registration of both local and global point clouds, with a smaller point cloud registration error and significantly reduced registration time.
[0103] Table 2 Point cloud ICP registration results
[0104]
[0105] S6 processes the reconstructed three-dimensional point cloud to obtain a true three-dimensional visual model of the underground pipe well.
[0106] Preferably, the method for processing the reconstructed three-dimensional point cloud includes:
[0107] S61 performs surface reconstruction on the reconstructed three-dimensional point cloud to obtain a surface reconstruction result;
[0108] After point cloud reconstruction, sparse or dense point cloud reconstruction results of underground well chambers can be obtained, but it only stays at a point cloud map of a large number of point sets, and there is no obvious connection between 3D points, and the visibility is poor; further surface reconstruction is required to restore a more realistic original appearance of the physical object and enhance the visualization effect. The present invention uses the PCL point cloud data processing library based on VS2017 in the Windows 10 environment to perform surface reconstruction on the reconstructed underground well chamber three-dimensional point cloud. Figure 9 Figure 2 shows the Poisson surface reconstruction results based on the 3D reconstructed point cloud of the underground well chamber. The reconstructed surface model shows that after Poisson surface reconstruction, certain connections are established between the independent 3D point clouds. The basic contours and shape of the target surface are well expressed, and the model's visibility is significantly improved.
[0109] S62 performs texture mapping on the surface reconstruction result according to the color image data of the target RGB image.
[0110] The model obtained after surface reconstruction lacks color texture information, which is different from the visual effect of the actual scene. The visualization effect is not optimal, and the true three-dimensional real scene reconstruction is not achieved. To this end, the present invention combines the reconstructed surface model with the color image data obtained by the Kinect depth camera through texture mapping to reconstruct a model with color texture. Figure 10The following figure shows a texture view of an underground well scene model. The 3D reconstruction using real texture mapping achieves superior visualization compared to the previous point cloud reconstruction, showcasing more detailed local texture details in the underground well, essentially aligning with the actual scene. The original 3D point cloud reconstruction result is simply a collection of unconnected point clouds. Surface reconstruction methods can transform these points into surfaces, obtaining a 3D reconstructed model. Furthermore, real texture mapping restores the target scene's true appearance, achieving true 3D visualization.
[0111] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.
Claims
1. A 3D visualization method for underground pipe wells based on RGB-D depth camera reconstruction, where: The method comprises: Acquire image data frames of underground pipe wells from different viewing angles using an RGB-D depth camera, wherein information of the image data frames includes a target depth image and a target RGB image; Preprocessing the target depth image to obtain a corrected target depth image; Fusing the corrected target depth image and the target RGB image at the same viewing angle to obtain a fused image data frame; According to the fused image data frames, initial posture information between adjacent fused image data frames is obtained, and a fused three-dimensional point cloud is obtained, wherein the initial posture information is calculated by matching adjacent fused image data frames according to a SIFT algorithm to obtain initial posture information; According to the initial posture information and the fused three-dimensional point cloud, the point cloud registration at different viewing angles is realized to obtain the reconstructed three-dimensional point cloud of the underground pipe well; Processing the reconstructed three-dimensional point cloud to obtain a true three-dimensional visual model of the underground pipe well; Among them, before matching, the obviously wrong matching point pairs are eliminated based on the improved RACSAC algorithm, and the method is as follows: 1) Assume that the set of ORB feature points extracted from two adjacent frames of images is: P={p i |p i ∈P,i=1,2,...,m} Q={q i |q i ∈Q,i=1,2,...,m} Find the characteristic point p in P i If the ratio of the Euclidean distance between the nearest neighbor and the second nearest neighbor of the two feature points in Q is less than the set threshold, then p is considered i It is a pair of matching points with the nearest neighbor ORB feature point; 2) Using a two-way matching mechanism, that is, finding the feature points in P that correspond to the feature points in Q, and similarly finding the points in P that correspond to the feature points of Q; if the two are one-to-one matches, the match is considered correct, and the incorrect matching results are eliminated, thereby achieving the purpose of screening; 3) By setting a minimum matching distance threshold, the remaining matching results that are greater than a certain multiple of the minimum matching distance are eliminated; after the above steps, matching point pairs with high matching quality are obtained; Before the RGB-D depth camera takes pictures, the depth measurement error of the RGB-D depth camera is corrected based on a polynomial surface fitting method, wherein the correction method is: Firstly, the ranging results of Kinect depth camera and laser ranging results are analyzed, and multiple groups of depth measurement errors and their changes within the ranging field of view of Kinect are obtained. Secondly, the acquired depth error data is used to construct a polynomial surface model of the true depth value with respect to the measured distance. The obtained measurement error data is fitted based on the principle of least squares method to obtain a fitting formula for the measured depth with respect to the depth error. Finally, the depth error is corrected by the error correction equation of polynomial surface fitting to obtain the corrected error; Among them, the depth error result is fitted by the polynomial surface model to obtain the following fitting formula: y=4E-10x 3 -3E-6x 2 +0.01x In the formula, y is the error correction value, x is the distance value, if the Kinect depth value is d, the corrected depth error is: d′=d-(4E-10x 3 -3E-6x 2 +0.01x).
2. The underground pipe well three-dimensional visualization method according to claim 1, wherein: The method for preprocessing the target depth image is: based on the joint bilateral filtering algorithm, the target RGB image is used as a guide image to achieve fast filtering and denoising of the target depth image.
3. The underground pipe well three-dimensional visualization method according to claim 1, wherein: The initial pose information and the fused three-dimensional point cloud are processed using an ICP algorithm based on neighborhood features to achieve registration of point clouds under different viewing angles.
4. The underground pipe well three-dimensional visualization method according to claim 1, wherein: The method for processing the reconstructed three-dimensional point cloud includes: Performing surface reconstruction on the reconstructed three-dimensional point cloud to obtain a surface reconstruction result; According to the color image data of the target RGB image, texture mapping is performed on the surface reconstruction result.
