Point cloud differential defect detection system and detection method based on multi-feature fusion

Abstract

The invention discloses a point cloud differential defect detection system and detection method based on multi-feature fusion, which are mainly used for high-precision detection of surface defects of a side wallboard of a railway passenger car, and the method comprises the following steps: obtaining datum point cloud data through a designed standard side wallboard model of the railway passenger car, and collecting real-time point cloud data through a laser scanner; preprocessing the obtained point cloud data; calculating the difference between the real-time point cloud and the reference point cloud by adopting a difference method for fusing geometric features such as Euclidean distance, curvature and normal vector, namely calculating the Euclidean distance difference, curvature difference and normal vector difference between the reference point cloud and the real-time point cloud; and a comprehensive abnormal score is generated by further combining a dynamic weight fusion mechanism, and different region characteristics (such as a plane, a curved surface and an edge region) are adapted through dynamic threshold adjustment. According to the method, the defect region is segmented through the clustering algorithm, and the defect area and depth are quantified. The method has the advantages of high detection precision, strong adaptability and good real-time performance, and is suitable for complex surface defect detection in industrial production.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional object detection, and specifically to a point cloud differential defect detection system and method based on multi-feature fusion. Background Art

[0002] With the development of laser scanning technology and three-dimensional point cloud data processing technology, point cloud data has been widely used in various defect detections. Especially in the field of defect detection of complex surfaces such as the side wall panels of rail passenger cars, point cloud data can provide accurate geometric shape information and help achieve efficient defect recognition. However, the existing technologies still face some significant challenges when processing these point cloud data.

[0003] Most traditional point cloud defect detection methods rely on single geometric features to identify defects. Although these methods can solve the detection problems of simple shape defects to a certain extent, they cannot effectively cope with the challenges of complex surface defects, especially in the case of variable geometric features such as minor deformations or local curved surfaces on the surface. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a point cloud differential defect detection system and method based on multi-feature fusion for the deficiencies of the above-mentioned existing technologies. The point cloud differential defect detection system and method based on multi-feature fusion can comprehensively consider defects of different geometric shapes, especially defects in minor deformations, curved surfaces, and edge regions, by fusing geometric features such as Euclidean distance difference, curvature difference, and normal vector difference and combining a dynamic weight adjustment mechanism of regional characteristics (types), so as to improve the recognition accuracy of different types of defects and can effectively cope with the challenges of complex surface defects.

[0005] To achieve the above technical purpose, the technical solution adopted by the present invention is as follows:

[0006] A point cloud differential defect detection method based on multi-feature fusion, comprising:

[0007] Step 1: Obtain reference point cloud data through a designed standard side wall panel model of a rail passenger car, and collect real-time point cloud data of the side wall panel of the rail passenger car through a laser scanner;

[0008] Step 2: Preprocess the obtained reference point cloud data and real-time point cloud data respectively. The preprocessing includes denoising, downsampling, and point cloud registration;

[0009] Step 3: After completing the registration of the point cloud, calculate the Euclidean distance difference, curvature difference, and normal vector difference between the reference point cloud and the real-time point cloud;

[0010] Step 4: Dynamically adjust the weights of the Euclidean distance difference, curvature difference, and normal vector difference in the comprehensive anomaly score according to the regional characteristics respectively. Based on the weights of the Euclidean distance difference, curvature difference, and normal vector difference, fuse the Euclidean distance difference, curvature difference, and normal vector difference to generate a comprehensive anomaly score, and obtain all the defective points in the real-time point cloud according to the comprehensive anomaly score;

[0011] Step 5: Based on the defective points, use the clustering algorithm to group the points with spatial continuity into the same defective area, thereby realizing the segmentation of the defective area; perform geometric quantization analysis on each segmented defective area;

[0012] Step 6: Visualize the analysis results obtained in Step 5.

[0013] As a further improved technical solution of the present invention, Step 1 specifically includes:

[0014] Step 1.1: Use the designed standard side wall panel model of the railway passenger car for data conversion, and convert it into point cloud data as the reference point cloud data;

[0015] Step 1.2: Select a three-dimensional laser scanner to collect the point cloud data on the surface of the side wall panel of the railway passenger car. The spatial resolution of the three-dimensional laser scanner is not less than 0.5 mm, the three-dimensional laser scanner has a measurement range of at least 5 m, and the scanning accuracy requirement reaches 0.1 mm;

[0016] Step 1.3: Before collecting the point cloud data, first set the scanning area and calibrate the three-dimensional laser scanner; install the three-dimensional laser scanner on a scanning platform to ensure that the optical axis of the three-dimensional laser scanner is parallel to the surface to be measured or aligned at a predetermined angle; use a standard calibration block to calibrate the three-dimensional laser scanner to ensure the consistency of the scanned data with the actual physical space;

[0017] Step 1.4: Collect the reference point cloud data to represent the geometric state of the side wall panel of the railway passenger car without defects; collect the real-time point cloud data to reflect the state of the side wall panel of the railway passenger car during actual production or use;

[0018] Both the collected reference point cloud and real-time point cloud data contain the three-dimensional coordinate information of the object surface and also contain the reflection intensity; store all the collected point cloud data and convert the point cloud data into a unified PCD standard format.

[0019] As a further improved technical solution of the present invention, Step 2 specifically includes:

[0020] Step 2.1: Use the statistical outlier removal algorithm to identify and remove outliers:

[0021] By calculating the Euclidean distance between each point and other points in its neighborhood, determine whether the point is a noise point. The specific formula is as follows:

[0022]

[0023] Among them, D i represents the average distance between the i-th point and n i points in its neighborhood, P i and P j represent the coordinates of the i-th point and the j-th point respectively;

[0024] Set a distance threshold θ. If D i is greater than θ, then the i-th point is determined to be a noise point and removed;

[0025] Step 2.2: Use the voxel grid filtering algorithm to downsample the point cloud data:

[0026] Divide the point cloud data into voxel grids, set the voxel side length v, and for the points in each voxel, select its centroid as the representative point for calculation:

[0027]

[0028] Among them, P v is the representative point of voxel v, n v is the number of points in this voxel, and P i is the coordinate of the i-th point in the voxel;

[0029] After performing the denoising process in Step 2.1 and the downsampling process in Step 2.2 on the collected reference point cloud and real-time point cloud data respectively, use the ICP algorithm to register the reference point cloud and the real-time point cloud, and align the reference point cloud with the real-time point cloud to ensure that both are in the same coordinate system; specifically as follows:

[0030] Step 2.3.1: Initially align the reference point cloud and the real-time point cloud through a coarse registration algorithm to obtain a transformation matrix T0;

[0031] Step 2.3.2: Use the ICP algorithm to continuously optimize the transformation matrix by iteratively calculating the minimum distance between each pair of points between the source point cloud and the target point cloud, and gradually reduce the difference between the two groups of point clouds; the formula is:

[0032]

[0033] Among them, P i is the coordinate of the i-th point in the source point cloud, and T(Pt i ) is the coordinate of the point after the coordinate of the i-th point Pt i in the target point cloud is transformed by the transformation matrix T, and T is the transformation matrix, represents the sum of squared errors of all corresponding point pairs;

[0034] At the initial iteration, T = T0. At each iteration, the T from the previous iteration * is used as the T for the current iteration, and the T for the current iteration is obtained according to formula (3) * ;

[0035] Through iterative optimization, until the iteration stop condition is met, the optimal transformation matrix T is obtained * ;

[0036] Step 2.3.3, after registration is completed, calculate the root mean square error; if the root mean square error is greater than the preset threshold, return to the initial registration step of the ICP algorithm and execute it again;

[0037] The calculation formula for the root mean square error is:

[0038]

[0039] where n is the number of point pairs, and T * (Pt i ) is the coordinate of the point after Pt i is transformed by the optimal transformation matrix T * , that is, the coordinate of the registered point.

[0040] As a further improved technical solution of the present invention, the specific steps of step 3 are as follows:

[0041] Step 3.1, for each point between the reference point cloud P base and the real-time point cloud P real , calculate its Euclidean distance difference:

[0042]

[0043] where P base (x i , y i , z i ) represents the coordinate of the i-th point in the reference point cloud, specifically (x basw , y basw , z base ); P real (x i , y i , z i ) represents the coordinate of the i-th point in the real-time point cloud, specifically (x real , y real , z real ); d i is the Euclidean distance difference between the i-th point in the reference point cloud and the i-th point in the real-time point cloud;

[0044] Step 3.2: For each point in the reference point cloud and the real-time point cloud, use the local plane fitting method to calculate the curvature based on the neighborhood fitting of the point cloud;

[0045] Local curvature formula:

[0046]

[0047] where k i is the curvature of the i-th point, and R i is the distance from the least-squares plane fitted in its neighborhood to the point P i ;

[0048] Calculate the curvature of the i-th point in the reference point cloud and the curvature of the i-th point in the real-time point cloud according to formula (6) respectively, and take the difference between the curvature of the i-th point in the reference point cloud and the curvature of the i-th point in the real-time point cloud to calculate the curvature difference Δk i :

[0049] Δk i =|k base,i -k real,i |(7);

[0050] where k base,i represents the curvature of the i-th point in the reference point cloud, and k real,i represents the curvature of the i-th point in the real-time point cloud;

[0051] Step 3.3: For each point in the reference point cloud and the real-time point cloud, calculate its normal vector, and the normal vector n i can be expressed as:

[0052]

[0053] where P i is the coordinate of the i-th point in the point cloud, P j is the coordinate of its neighborhood point, and P avg is the average position of the neighborhood points;

[0054] Calculate the normal vector n i of the i-th point in the reference point cloud and the normal vector n i of the i-th point in the real-time point cloud according to formula (8) respectively, and take the difference between the normal vector of the reference point cloud and the normal vector of the real-time point cloud to calculate the normal vector difference Δn i :

[0055] Δn i =||n base,i -n real,i || (9);

[0056] where n base,iDenote the normal vector of the $i$-th point in the reference point cloud as $\mathbf{n}$ base,i Denote the normal vector of the $i$-th point in the real-time point cloud.

[0057] As a further improved technical solution of the present invention, step 4 specifically includes:[[]]

[0058] Step 4.1: Based on the weight $w_{\text{ed}}$ of the Euclidean distance difference euc , the weight $w_{\text{cur}}$ of the curvature difference curv and the weight $w_{\text{nor}}$ of the normal vector difference norm fuse the Euclidean distance difference $d$ i , the curvature difference $\Delta k$ i and the normal vector difference $\Delta\mathbf{n}$ i to generate the comprehensive anomaly score of the $i$-th point in the real-time point cloud, and the formula is:[[]]

[0059] $S$ i $=$ $w_{\text{ed}}$ euc $d$ i $+$ $w_{\text{cur}}$ curv $\Delta k$ i $+$ $w_{\text{nor}}$ norm $\Delta\mathbf{n}$ i (10);

[0060] Step 4.2: According to the geometric characteristics of the point cloud data, identify the type of the region where the $i$-th point in the current real-time point cloud is located. The region types can be divided into flat regions, curved surface regions, and edge regions;

[0061] According to the region type, dynamically adjust each weight:[[]]

[0062] When the region where the $i$-th point in the real-time point cloud is located is a flat region, increase the weight $w_{\text{ed}}$ of the Euclidean distance difference euc , decrease the weight $w_{\text{cur}}$ of the curvature difference curv and decrease the weight $w_{\text{nor}}$ of the normal vector difference norm ;

[0063] When the region where the $i$-th point in the real-time point cloud is located is a curved surface region, increase the weight $w_{\text{cur}}$ of the curvature difference curv , decrease the weight $w_{\text{nor}}$ of the normal vector difference norm and decrease the weight $w_{\text{ed}}$ of the Euclidean distance difference euc ;

[0064] When the region where the $i$-th point in the real-time point cloud is located is an edge region, increase the weight $w_{\text{nor}}$ of the normal vector difference norm , decrease the weight $w_{\text{cur}}$ of the curvature difference curv and decrease the weight $w_{\text{ed}}$ of the Euclidean distance difference euc ;

[0065] Step 4.3: Perform a threshold determination on the comprehensive anomaly score $S$ of the $i$-th point in the real-time point cloud i . If $S$i Greater than the set threshold θ fusion , then the i-th point in the real-time point cloud is considered a defective point:

[0066] S i > θ fusion (11).

[0067] As a further improved technical solution of the present invention, step 5 specifically includes:

[0068] Step 5.1: Set the neighborhood radius ∈ and the minimum number of neighborhood points MinPts;

[0069] For each defective point P i , first determine its neighborhood range, that is, all points whose distance from this defective point P i is less than or equal to ∈. The number of points in the neighborhood is denoted as N ∈ (P i );

[0070] If N ∈ (P i ) ≥ MinPts, then this defective point P i is regarded as a core point; starting from the core point, the points in its neighborhood are successively grouped into the same cluster until no further expansion is possible;

[0071] If N ∈ (P i ) < MinPts and it is not in the neighborhood of any core point, then this defective point is regarded as a noise point and will not be classified into any cluster;

[0072] The result after clustering is multiple clusters containing defective points, and each cluster represents an independent defective area;

[0073] Step 5.2: After completing the segmentation of the defective area, perform geometric quantization analysis on each segmented defective area, including calculating the area of the defective area and the depth of the defective area:

[0074] Calculate the area A of each segmented defective area through the convex hull algorithm defect ;

[0075] The depth of the defective area refers to the depth difference between the deepest point in the defective area and the reference point cloud; assume that the deepest point in the defective area is P deepest , and its coordinates are The depth of the corresponding area in the reference point cloud is z base , then the depth D defect of the defective area is:

[0076] D defect = z base - z deepest(12);

[0077] Step 5.3: After Steps 5.1 and 5.2, all the segmented defect regions will be given corresponding geometric features, including the area A of the defect region defect and the depth D of the defect region defect .

[0078] As a further improved technical solution of the present invention, Step 6 specifically includes:

[0079] Obtain the area A defect , depth D defect and position of each defect region through Steps 1, 2, 3, 4, and 5;

[0080] Visualize the real-time point cloud data through 3D rendering technology. According to the calculated area A defect , depth D defect and position of each defect region, mark the defect regions on the 3D point cloud with different colors.

[0081] To achieve the above technical purpose, another technical solution adopted by the present invention is:

[0082] A point cloud differential defect detection system based on multi-feature fusion, including:

[0083] An image acquisition module, which is used to obtain reference point cloud data for the designed standard model of the side wall panel of the rail vehicle, and is used to collect the reference point cloud and real-time point cloud data of the side wall panel of the rail vehicle through a laser scanner;

[0084] An image preprocessing module, which is used to preprocess the collected reference point cloud data and real-time point cloud data respectively. The preprocessing includes denoising processing, downsampling processing, and point cloud registration;

[0085] A feature extraction module, which is used to calculate the Euclidean distance difference, curvature difference, and normal vector difference between the reference point cloud and the real-time point cloud;

[0086] A dynamic threshold adjustment module, which is used to dynamically adjust the weights of the Euclidean distance difference, curvature difference, and normal vector difference in the comprehensive anomaly score respectively according to the regional characteristics, and is used to fuse the Euclidean distance difference, curvature difference, and normal vector difference based on the weights of the Euclidean distance difference, curvature difference, and normal vector difference to generate a comprehensive anomaly score, and is used to obtain all the defect points in the real-time point cloud according to the comprehensive anomaly score;

[0087] A defect segmentation module, which is used to group the points with spatial continuity into the same defect region based on the defect points through a clustering algorithm, thereby realizing the segmentation of the defect region;

[0088] A geometric quantization analysis module for performing geometric quantization analysis on each segmented defect region, including the area and depth of the defect region;

[0089] A result output and visualization module for visually displaying the analysis results obtained by the geometric quantization analysis module through three-dimensional rendering technology.

[0090] The beneficial effects of the present invention are as follows:

[0091] (1) By combining geometric features such as Euclidean distance, curvature, and normal vector, the present invention overcomes the limitation of traditional point cloud difference methods that rely on a single geometric feature. Through multi-feature fusion technology, it can comprehensively reflect the geometric features of different defects, especially having higher recognition accuracy for defects with complex shapes. At the same time, the dynamic weight fusion mechanism automatically adjusts the weights of each feature according to the characteristics of different regions, thereby enhancing the robustness of detection and ensuring that defects in different regions can be effectively identified.

[0092] (2) Traditional point cloud difference methods usually adopt fixed thresholds, which are difficult to meet the detection requirements of complex surface regions. The present invention designs a dynamic weight adjustment mechanism to dynamically adjust the weights of each feature according to regional characteristics during the detection process, which can flexibly adapt to different defect characteristics in planar, curved, and edge regions. This mechanism not only optimizes the detection accuracy of defects but also improves the response ability to various geometric deformations.

[0093] (3) The present invention uses the DBSCAN clustering algorithm to process the differential point cloud data. By classifying points with spatial continuity into the same cluster, it can effectively segment the defect regions. Compared with traditional methods, the DBSCAN algorithm can better identify continuous defects in high-density defect regions and avoid segmentation errors caused by sparse point cloud data or local noise effects, thereby improving the accuracy of defect segmentation.

[0094] (4) Through fine geometric quantization analysis, the present invention calculates geometric characteristics such as the area and depth of the defect region. The surface area of the defect region is calculated by the convex hull algorithm, and the depth of the defect is calculated by the difference between the reference point cloud and the real-time point cloud, which provides a quantitative basis for subsequent quality assessment and repair decision-making. Compared with traditional qualitative assessment methods, the present invention can provide more accurate quantitative data for defect detection, which helps to scientifically and reasonably evaluate the impact of defects on the structure.

[0095] (5) The present invention visualizes the defect region using three-dimensional rendering technology, which can intuitively display information such as the type, location, area, and depth of the defect. By annotating the defect region with different colors on the three-dimensional point cloud model, this visualization method improves the readability of the detection results and provides convenient defect detection feedback.

[0096] (6) The present invention greatly improves the detection efficiency through an automated process of point cloud acquisition, preprocessing, differential analysis, feature fusion, defect segmentation, and quantitative analysis. By dynamically adjusting the feature weights and thresholds, the method has higher real-time performance and adaptability in complex surface defect detection, meeting the requirements of industrial production lines for efficient, accurate, and real-time detection.

[0097] (7) The present invention combines geometric features such as Euclidean distance, curvature difference, and normal vector difference through a dynamic weight adjustment mechanism that combines regional characteristics. Through multi-feature fusion, it can comprehensively consider defects of different geometric forms, especially defects in micro-deformations, curved surfaces, and edge regions. Dynamically adjust the weights of each feature according to the characteristics of different regions, thereby improving the recognition accuracy of different types of defects. The present invention realizes accurate segmentation and geometric quantification of defect regions through an optimized clustering algorithm and weighted fusion strategy, improving the overall reliability of defect detection.

[0098] (8) Through the dynamic weight fusion and dynamic threshold adjustment mechanism, the accuracy and robustness of the surface defect detection of the side wall panels of rail passenger cars are improved, meeting the real-time detection requirements of industrial production lines. Description of the Drawings

[0099] Figure 1 It is a flowchart of a point cloud differential defect detection method based on multi-feature fusion.

[0100] Figure 2 It is a process diagram of realizing defect detection of the side wall panels of rail passenger cars through multi-feature fusion.

[0101] Figure 3 In (a), it is a real-time point cloud map of the side wall panels of rail passenger cars.

[0102] Figure 3 In (b), it is a defect recognition map of the side wall panels of rail passenger cars.

[0103] Figure 3 In (c), it is a defect recognition map of the side wall panels of rail passenger cars.

[0104] Figure 3 In (d), it is a defect recognition map of the side wall panels of rail passenger cars.

[0105] Figure 3 In (e), it is a defect recognition map of the side wall panels of rail passenger cars.

[0106] Figure 3 In (f), it is a defect recognition map of the side wall panels of rail passenger cars.

[0107] Figure 3 In (g), it is a defect recognition map of the side wall panels of rail passenger cars. Detailed Embodiment

[0108] The following further describes the specific embodiments of the present invention with reference to the accompanying drawings:

[0109] A point cloud differential defect detection method based on multi-feature fusion has a process as Figure 1 shown, including:

[0110] Step 1: Obtain reference point cloud data through a designed standard track passenger car side wall panel model, and collect real-time point cloud data of the track passenger car side wall panel through a laser scanner;

[0111] The collected point cloud data contains attributes such as three-dimensional coordinate information and reflection intensity on the object surface, so as to comprehensively reflect the geometric shape and material characteristics of the measured surface.

[0112] Step 2: For the collected original point cloud data, to ensure the quality of the point cloud and provide reliable input for subsequent processing, preprocess the original data, including: denoising processing, downsampling processing, and point cloud registration. After the above steps, the preprocessing work of the point cloud is completed, and high-quality point cloud data is generated.

[0113] Step 3: After completing the registration of the point cloud, calculate the Euclidean distance difference, curvature difference, and normal vector difference between the reference point cloud and the real-time point cloud.

[0114] After completing the registration of the point cloud, the present invention uses the differences between the reference point cloud and the real-time point cloud to identify possible defect areas. A multi-feature fusion detection method based on Euclidean distance difference, curvature difference, and normal vector difference is proposed. The Euclidean distance is suitable for global deformation detection, but is insensitive to local minute deformations. The curvature difference is suitable for local geometric deformation detection, but may be more sensitive to noise. The normal vector difference is suitable for detecting mutation regions, but has limited effects on smooth regions. These features are combined to adapt to the characteristics of different detection regions. In this processing process, as Figure 2 shown, it includes a differential calculation module.

[0115] Step 4: Dynamically adjust the weights of the Euclidean distance difference, curvature difference, and normal vector difference in the comprehensive anomaly score respectively according to the regional characteristics, fuse the Euclidean distance difference, curvature difference, and normal vector difference based on the weights of the Euclidean distance difference, curvature difference, and normal vector difference to generate a comprehensive anomaly score, and obtain all the defect points in the real-time point cloud according to the comprehensive anomaly score.

[0116] Regarding the geometric features of differential point clouds, the present invention generates a comprehensive feature score by weighted fusion of Euclidean distance, curvature features, and normal vector features. To meet the detection requirements of different regional characteristics (such as flat surfaces, curved surfaces, and edge regions), a dynamic weight adjustment mechanism is designed to dynamically adjust the feature weights according to the regional characteristics. When a flat area is detected, the weight of the Euclidean distance is increased; when a curved surface or edge region is detected, the weights of the curvature and / or normal vector features are increased. According to the fusion result, a comprehensive anomaly score of the differential point cloud is generated to reflect the significance of the point cloud difference. During this processing, as Figure 2 shown, it includes a regional characteristic judgment module, a dynamic weight adjustment module, a weighted fusion module, and a defect identification module.

[0117] Step 5: Based on the defect points, use a clustering algorithm to group points with spatial continuity into the same defect region, thereby achieving the segmentation of the defect region; perform geometric quantization analysis on each segmented defect region.

[0118] After obtaining the comprehensive anomaly score, the present invention segments the defect region through a clustering algorithm and quantifies the geometric characteristics of the defect. The DBSCAN clustering algorithm is used to analyze the comprehensive anomaly score to segment the defect region with spatial continuity. For each segmented defect region, geometric quantization analysis is performed, including defect area and defect depth.

[0119] Step 6: Visualize the analysis results obtained in Step 5.

[0120] The present invention outputs and visualizes the defect detection results, including: outputting the type and location of the defect, and using three-dimensional rendering technology to label the defect region with different colors. The type of the defect can be identified manually.

[0121] The specific content of Step 1 includes:

[0122] Step 1.1: Select a high-precision three-dimensional laser scanner to collect real-time point cloud data, ensuring that the three-dimensional surface information of the side wall panel of the rail passenger car can be accurately obtained. The spatial resolution of the laser scanner should be no less than 0.5 mm to ensure that small surface defects can be captured. The scanner should have a measurement range of at least 5 meters to ensure that the entire side wall panel surface can be covered; the scanning accuracy requirement is 0.1 mm to ensure that the obtained real-time point cloud data accurately reflects the geometric shape of the side wall panel surface.

[0123] Step 1.2: Before collecting point cloud data, first set the scanning area and calibrate the 3D laser scanner to ensure the accuracy and consistency of the collected data. Install the 3D laser scanner on a stable scanning platform to ensure that the optical axis of the 3D laser scanner is parallel to the surface to be measured or aligned at a predetermined angle; use a standard calibration block to calibrate the 3D laser scanner to ensure the consistency between the scanned data and the actual physical space.

[0124] Step 1.3: Collect reference point cloud data to represent the geometric state of the side wall panel of the rail vehicle when there are no defects, as a reference for subsequent comparison with real-time point cloud data; collect real-time point cloud data to reflect the state of the side wall panel of the rail vehicle during actual production or use.

[0125] Convert the standard design model into point cloud data to obtain the reference point cloud data, and the reference point cloud is used to provide the geometric shape of the side wall panel of the rail vehicle. The point cloud data collected during production or use may contain defect information. The real-time point cloud is used to compare with the reference point cloud to identify potential defects. The collected point cloud data should include the three-dimensional coordinate information of the object surface, as well as physical properties such as reflection intensity. All collected point cloud data is stored, and the point cloud data is converted into a unified PCD standard format.

[0126] The specific steps of Step 2 are as follows:

[0127] Step 2.1: Use the Statistical Outlier Removal (SOR) algorithm to identify and remove outliers.

[0128] By calculating the Euclidean distance between each point and other points in its neighborhood, determine whether the point is a noise point. The specific formula is as follows:

[0129]

[0130] where D i represents the average distance between the i-th point and n i points in its neighborhood, P i and P j represent the coordinates of the i-th point and the j-th point respectively; set a distance threshold θ. If D i is greater than θ, then the i-th point is determined to be a noise point and removed.

[0131] Step 2.2: In order to reduce the computational amount of the point cloud data and improve the efficiency of subsequent processing, use the Voxel Grid Filter algorithm to downsample the point cloud data. Divide the point cloud data into voxel grids, set the voxel side length v, and for the points in each voxel, select its centroid as the representative point for calculation:

[0132]

[0133] Among them, P v is the representative point of voxel v, and n v is the number of points in the voxel, and P i is the coordinate of the i-th point in the voxel.

[0134] Step 2.3: After performing the denoising process in Step 2.1 and the downsampling process in Step 2.2 on the collected reference point cloud and real-time point cloud data respectively, use the Iterative Closest Point (ICP) algorithm to register the reference point cloud and the real-time point cloud, align the reference point cloud with the real-time point cloud, and ensure that both are in the same coordinate system; specifically as follows:

[0135] Step 2.3.1: Initially align the reference point cloud and the real-time point cloud through a rough registration algorithm to obtain a rough transformation matrix T0;

[0136] Step 2.3.2: Use the ICP algorithm to iteratively calculate the minimum distance between each pair of points between the source point cloud and the target point cloud, continuously optimize the transformation matrix, and gradually reduce the difference between the two sets of point clouds; the formula is:

[0137]

[0138] Among them, P i is the coordinate of the i-th point in the source point cloud, and T(Pt i ) is the coordinate of the point after the coordinate of the i-th point Pt i in the target point cloud is transformed by the transformation matrix T, T is the transformation matrix, represents the sum of the squared errors of all corresponding point pairs;

[0139] At the initial iteration, T = T0. At each iteration, use the T * from the previous iteration as the T for the current iteration, and calculate the T * for the current iteration according to formula (3);

[0140] Through iterative optimization, until the stopping iteration condition is met (such as the iteration number reaches the upper limit, or the change amount of the solution is less than a certain threshold), obtain the optimal transformation matrix T * ;

[0141] Step 2.3.3: After the registration is completed, use the Root Mean Square Error (RMSE) to evaluate the accuracy of the registration result. Calculate the root mean square error; if the root mean square error is greater than the preset threshold, return to the initial registration step of the ICP algorithm to re-execute to ensure the accuracy of the registration result;

[0142] The calculation formula of the root mean square error is:

[0143]

[0144] Among them, n is the number of point pairs, and T * (Pt i ) is the coordinate of the point after Pt i is transformed by the optimal transformation matrix T * , that is, the coordinate of the registered point.

[0145] The specific steps of step 3 include:

[0146] Step 3.1: For defects in a large range, Euclidean distance difference is used to detect global deformation. For each point between the reference point cloud P base and the real-time point cloud P real , calculate its Euclidean distance difference:

[0147]

[0148] Among them, P nase (x i , y i , z i ) represents the coordinate of the i-th point in the reference point cloud, specifically (x base , y base , z base ) ; P real (x i , y i , z i ) represents the coordinate of the i-th point in the real-time point cloud, specifically (x real , y real , z real ) ; d i is the Euclidean distance difference between the i-th point in the reference point cloud and the i-th point in the real-time point cloud;

[0149] For each point in the detection area, if its Euclidean distance difference d i exceeds the set threshold, it is considered that this point may belong to the defect area: d i > θ euc ; Among them, θ euc is the threshold of the Euclidean distance, which determines whether this point is marked as a possible defect.

[0150] Step 3.2: For the detection of small-range defects, curvature difference is used to detect local geometric deformation. For each point in the reference point cloud and the real-time point cloud, the local plane fitting method is used to calculate the curvature based on the neighborhood fitting of the point cloud;

[0151] Local curvature formula:

[0152]

[0153] Among them, k iis the curvature of the ith point, R i Point P i The distance from the point to the least squares plane fitted to its neighborhood; greater curvature indicates more dramatic surface changes.

[0154] According to formula (6), the curvature of the i-th point in the reference point cloud and the curvature of the i-th point in the real-time point cloud are calculated respectively, and the curvature of the i-th point in the reference point cloud and the curvature of the i-th point in the real-time point cloud are differentiated to calculate the curvature difference Δk i :

[0155] Δk i =|k base,i -k real,i |(7);

[0156] Among them, k base,i represents the curvature of the i-th point in the reference point cloud, k real,i Represents the curvature of the i-th point in the real-time point cloud.

[0157] If the curvature difference Δk i If it exceeds the set threshold, the point is judged as a possible defect area: Δk i >θ curv ; where θ curv is the threshold of curvature difference, which is used to determine whether there is abnormal deformation.

[0158] Step 3.3: For the detection of surface mutation areas, normal vector difference is used for detection. For each point in the reference point cloud and the real-time point cloud, its normal vector is calculated. The normal vector is usually obtained by fitting a plane with neighboring points. i It can be expressed as:

[0159]

[0160] Among them, P i is the coordinate of the i-th point in the point cloud, P j is the coordinate of its neighborhood point, P avg is the average position of the neighborhood points;

[0161] According to formula (8), the normal vector n of the i-th point in the reference point cloud is calculated respectively: i and the normal vector n of the i-th point in the real-time point cloud i , the normal vector of the reference point cloud and the normal vector of the real-time point cloud are differentiated to calculate the normal vector difference Δn i :

[0162] Δn i =||n base,i -n real,i || (9);

[0163] Where nbase,i represents the normal vector of the i-th point in the reference point cloud, n base,i represents the normal vector of the i-th point in the real-time point cloud.

[0164] If the normal vector difference Δn i exceeds the set threshold, then it is determined that this point is a possible defect area: Δn i > θ norm ; where θ norm is the threshold of the normal vector difference, used to determine whether it is a mutation area.

[0165] The specific steps of step 4 include:

[0166] In order to integrate the advantages of each feature, the present invention fuses the results of Euclidean distance difference, curvature difference, and normal vector difference to generate a comprehensive anomaly score. According to the characteristics of each region, the weights of Euclidean distance difference, curvature difference, and normal vector difference (i.e., the weighting coefficients of each feature) are dynamically adjusted.

[0167] Step 4.1, based on the weight w of the Euclidean distance difference euc , the weight w of the curvature difference curv and the weight w of the normal vector difference norm fuse the Euclidean distance difference d i , the curvature difference Δk i and the normal vector difference Δn i to generate the comprehensive anomaly score of the i-th point in the real-time point cloud. The formula is:

[0168] S i = w euc d i + w curv Δk i + w norm Δn i (10);

[0169] where: d i is the Euclidean distance difference; Δk i is the curvature difference; Δn i is the normal vector difference; w euc + w curv + w norm = 1.

[0170] Without considering the regional characteristics, the initial weight values can be set evenly. For example: w euc = 0.33, w curv = 0.33, w norm = 0.34; This initial weight assumes that all features contribute equally to defect detection and is applicable to general situations.

[0171] Step 4.2: The present invention designs a dynamic weight adjustment mechanism to adjust the weights of features according to the geometric characteristics of the detection area.

[0172] According to the geometric characteristics of the point cloud data, identify the type of the area where the i-th point in the current real-time point cloud is located. The area types can be divided into flat areas, curved surface areas, and edge areas.

[0173] When the curvature value is small (close to 0) and the normal vector change of the point cloud is small, it is a flat area. When the curvature value is large and the normal vector has no obvious change, it is a curved surface area. When the curvature value is large and the normal vector changes violently, it is an edge area. For the division of areas, specifically: when the curvature value is less than 0.01 and the normal vector change of the point cloud is less than 0.05, this area is a flat area. When the curvature value is greater than 0.1 and the normal vector change of the point cloud is less than 0.1, this area is a curved surface area. Edge area: when the curvature value is greater than 0.1 and the normal vector change is greater than 0.2, this area is an edge area. The above values can be changed according to the actual situation.

[0174] According to the area type, dynamically adjust the weights of each feature:

[0175] When the area where the i-th point in the real-time point cloud is located is a plane area, increase the weight w of the Euclidean distance difference euc , reduce the weight w of the curvature difference curv , reduce the weight w of the normal vector difference norm ; for example: w euc = 0.6, w curv = 0.2, w norm = 0.2.

[0176] When the area where the i-th point in the real-time point cloud is located is a curved surface area, increase the weight w of the curvature difference curv , reduce the weight w of the Euclidean distance difference euc , reduce the weight w of the normal vector difference norm ; for example: w euc = 0.2, w curv = 0.6, w norm = 0.2.

[0177] When the area where the i-th point in the real-time point cloud is located is an edge area, increase the weight w of the normal vector difference norm , reduce the weight w of the Euclidean distance difference euc , reduce the weight w of the curvature difference curv ; for example: w euc = 0.2, w curv = 0.2, w norm = 0.6.

[0178] These dynamically adjusted weights can enable the most important geometric features to obtain stronger responses in different regions, thereby improving the accuracy of defect detection.

[0179] Step 4.3, the comprehensive anomaly score S of the i-th point in the real-time point cloud i is subjected to a threshold determination. If S i is greater than the set threshold θ fusion , then the i-th point in the real-time point cloud is considered a defect point:

[0180] S i > θ fusion (11).

[0181] The specific steps of step 5 include:

[0182] The present invention uses the DBSCAN clustering algorithm to group points with spatial continuity into the same defect region according to the comprehensive anomaly score, thereby realizing the segmentation of the defect region. For each point P i , its comprehensive anomaly score S i has been calculated according to the weighted fusion in step 4. The comprehensive anomaly scores of all points form a data set D = {S1, S2,..., S n}}.

[0183] Step 5.1, the DBSCAN algorithm used in the present invention performs point clustering by setting the neighborhood radius ∈ and the minimum number of neighborhood points MinPts.

[0184] For each defect point P i , first determine its neighborhood range, that is, all points whose distance from this defect point P i is less than or equal to ∈. The number of points in the neighborhood is denoted as N ∈ (P i );

[0185] If N ∈ (P i ) ≥ MinPts, then this defect point P i is regarded as a core point; starting from the core point, the points in its neighborhood are successively grouped into the same cluster until no further expansion is possible;

[0186] If N ∈ (P i ) < MinPts and it is not in the neighborhood of any core point, then this defect point is regarded as a noise point and will not be grouped into any cluster;

[0187] The result after clustering is multiple clusters containing defect points, and each cluster represents an independent defect region.

[0188] Step 5.2. After the segmentation of the defect areas is completed, geometric quantitative analysis is performed on each segmented defect area, including the calculation of the area of the defect area and the depth of the defect area:

[0189] Calculate the area A of each segmented defect area through the convex hull algorithm defect .

[0190] For each segmented defect area, calculate the area of the defect area. Assume that the defect area contains n points, denoted as P1, P2, …, P n , and the coordinates of each point are (x i , y i , z i ). Its area A defect can be obtained by calculating the two-dimensional projection of the point cloud through the convex hull algorithm and calculating the polygon area:

[0191] A defect =Area of Convex Hull(P1, P2, …, P n );

[0192] Area of Convex Hull(P1, P2, …, P n ) is the result of applying the convex hull algorithm to the point set (P1, P2, …, P n ), and (P1, P2, …, P n ) represents the point cloud within the defect area. By calculating the outer bounding area of the defect area, the surface area of the defect area is obtained, which is also the area of the defect area.

[0193] The depth of the defect area refers to the depth difference between the deepest point in the defect area and the reference point cloud; assume that the deepest point in the defect area is P deepest , and its coordinates are (x deepest , y deepest , z deepest ), and the depth of the corresponding area in the reference point cloud is z base , then the depth D defect of the defect area is:

[0194] D defect =z base -z deepest (12);

[0195] The depth of the defect area is obtained, that is, the vertical distance from the surface of the reference point cloud to the deepest point of the defect area.

[0196] Step 5.3. After Steps 5.1 and 5.2, all segmented defect areas will be assigned corresponding geometric properties, including the area A defect of the defect area and the depth Ddefect These geometric features will serve as quantitative information about the defects. Finally, the system will output information for each defect area, including: the identification and location of the defect area; the area of the defect; the depth of the defect.

[0197] Step 6 specifically includes:

[0198] Obtain the area A of each defect area through Steps 1, 2, 3, 4, and 5 defect and depth D defect and location.

[0199] Visualize the real-time point cloud data through 3D rendering technology. According to the calculated area A of each defect area defect and depth D defect and location, mark the defect areas on the 3D point cloud with different colors. Among them, the defect types (the main defect types identified for the side wall panel are dents, protrusions, cracks, etc., and the defect types can be identified manually) can also be visualized through 3D rendering technology.

[0200] Visualize the point cloud data through 3D rendering technology to ensure that the spatial distribution of the point cloud and the defect areas are clearly visible. Use the open3d visualization library to convert the point cloud data into a 3D model. According to the calculated defect area positions, as well as the size and depth of the defects, mark the defect areas on the 3D point cloud with different colors.

[0201] Figure 3 In (a) is the real-time point cloud map of the side wall panel of the rail passenger car. Figure 3 In (b)-(g) are the defect recognition result maps of the real-time point cloud data of the side wall panel of the rail passenger car in (a) by the method of the present invention. Among them, the red within the rectangular frame is the marked position of the recognized defect area. Figure 3

[0202] This embodiment also provides a point cloud differential defect detection system based on multi-feature fusion, including:

[0203] An image acquisition module, which is used to obtain the reference point cloud data for the designed standard model of the side wall panel of the rail passenger car, to collect the real-time point cloud data of the side wall panel of the rail passenger car through a laser scanner, and to perform preliminary processing on the collected point cloud to ensure that the image quality meets the requirements of subsequent processing;

[0204] An image preprocessing module, which is used to perform preprocessing on the collected reference point cloud data and real-time point cloud data respectively. The preprocessing includes denoising processing, downsampling processing, and point cloud registration; to ensure the quality of the point cloud data and provide reliable input for subsequent defect detection;

[0205] A feature extraction module, that is Figure 2 ​The difference calculation module is used to calculate the Euclidean distance difference, curvature difference, and normal vector difference between the reference point cloud and the real-time point cloud; it is used to extract geometric features such as Euclidean distance, curvature, and normal vector from the preprocessed point cloud, and generate differential data by combining different feature information through a multi-feature fusion method for defect area identification;

[0206] The dynamic threshold adjustment module includes Figure 2 the area feature judgment module, dynamic weight adjustment module, weighted fusion module, and defect identification module in it. They are respectively used to judge the area characteristics; used to dynamically adjust the weights of the Euclidean distance difference, curvature difference, and normal vector difference in the comprehensive anomaly score according to the area characteristics, used to fuse the Euclidean distance difference, curvature difference, and normal vector difference based on the weights of the Euclidean distance difference, curvature difference, and normal vector difference to generate a comprehensive anomaly score, used to obtain all defect points in the real-time point cloud according to the comprehensive anomaly score; used to dynamically adjust the weights of each feature according to different area characteristics to ensure the defect detection accuracy in different areas;

[0207] The defect segmentation module is used to group points with spatial continuity into the same defect area based on the defect points, and thus realize the segmentation of the defect area;

[0208] The geometric quantization analysis module is used to perform geometric quantization analysis on each segmented defect area, including the area of the defect area and the depth of the defect area, to provide accurate quantitative results to assist subsequent processing and decision-making;

[0209] The result output and visualization module is used to visually display the analysis results (the area of the defect area, the depth of the defect area, the position of the defect area, etc.) obtained by the geometric quantization analysis module through 3D rendering technology.

[0210] The protection scope of the present invention includes but is not limited to the above embodiments. The protection scope of the present invention is subject to the claims. Any replacement, deformation, and improvement that are easily conceivable by those skilled in the art to this technology fall within the protection scope of the present invention.

Claims

1. A point cloud differential defect detection method based on multi-feature fusion, characterized in that: include: Step 1: Obtain reference point cloud data through the designed standard rail passenger car side wall panel model, and collect real-time point cloud data of the rail passenger car side wall panel through a laser scanner; Step 2: Preprocess the acquired reference point cloud data and real-time point cloud data respectively, including denoising, downsampling and point cloud registration; Step 3: After completing the registration of the point cloud, calculate the Euclidean distance difference, curvature difference, and normal vector difference between the reference point cloud and the real-time point cloud; Step 4: Dynamically adjust the weights of the Euclidean distance difference, the curvature difference, and the normal vector difference in the comprehensive anomaly score according to the regional characteristics. Based on the weights of the Euclidean distance difference, the curvature difference, and the normal vector difference, the Euclidean distance difference, the curvature difference, and the normal vector difference are fused to generate a comprehensive anomaly score. All defect points in the real-time point cloud are obtained according to the comprehensive anomaly score. Step 5: Based on the defect points, points with spatial continuity are classified into the same defect area through a clustering algorithm, thereby segmenting the defect area; and geometric quantitative analysis is performed on each segmented defect area; Step 6: Visualize the analysis results obtained in step 5.

2. The point cloud difference defect detection method based on multi-feature fusion according to claim 1 is characterized in that: The step 1 specifically includes: Step 1.1, use the designed standard rail passenger car side wall panel model to convert data into point cloud data as reference point cloud data; Step 1.2: Use a 3D laser scanner to collect real-time point cloud data on the surface of the side wall of the railway passenger car. The spatial resolution of the 3D laser scanner shall not be less than 0.5 mm. The 3D laser scanner shall have a measurement range of at least 5 meters and a scanning accuracy of 0.1 mm. Step 1.3: Before collecting point cloud data, first set the scanning area and calibrate the 3D laser scanner; install the 3D laser scanner on a scanning platform, ensure that the optical axis of the 3D laser scanner is parallel to the surface to be measured or aligned at a predetermined angle; calibrate the 3D laser scanner using a standard calibration block to ensure the consistency of the scanned data with the actual physical space; Step 1.4, collecting reference point cloud data to represent the geometric state of the rail passenger car side wall panel when there are no defects on the surface; collecting real-time point cloud data to reflect the state of the rail passenger car side wall panel during actual production or use; The collected reference point cloud and real-time point cloud data both contain the three-dimensional coordinate information of the object surface and the reflection intensity; all collected point cloud data are stored and converted into a unified PCD standard format.

3. The point cloud difference defect detection method based on multi-feature fusion according to claim 1 is characterized in that: The step 2 specifically includes: Step 2.1: Use statistical outlier removal algorithm to identify and remove outliers: By calculating the Euclidean distance between each point and other points in its neighborhood, we can determine whether the point is a noise point. The specific formula is as follows: Among them, D i Represents the relationship between the i-th point and its neighborhood n i The average distance of points, P i and P j Represent the coordinates of the i-th point and the j-th point respectively; Set a distance threshold θ, if D i If it is greater than θ, the i-th point is judged as a noise point and removed; Step 2.2: Use voxel grid filtering algorithm to downsample the point cloud data: Divide the point cloud data into voxel grids, set the voxel side length v, and for each point in the voxel, select its centroid as the representative point for calculation: Among them, P v is the representative point of voxel v, n v is the number of points in the voxel, P i is the coordinate of the i-th point in the voxel; Step 2.3: After the collected reference point cloud and real-time point cloud data are subjected to denoising processing in step 2.1 and downsampling processing in step 2.2, the reference point cloud and the real-time point cloud are registered using the ICP algorithm to align the reference point cloud with the real-time point cloud to ensure that they are in the same coordinate system; the details are as follows: Step 2.3.1: Perform preliminary alignment between the reference point cloud and the real-time point cloud through a coarse registration algorithm to obtain a transformation matrix T0; Step 2.3.2: Use the ICP algorithm to iteratively calculate the minimum distance between each pair of points between the source point cloud and the target point cloud, continuously optimize the transformation matrix, and gradually reduce the difference between the two groups of point clouds; the formula is: Among them, P i is the coordinate of the i-th point in the source point cloud, T(Pt i ) is the i-th point Pt in the target point cloud i The coordinates of the point after the coordinates are transformed by the transformation matrix T, where T is the transformation matrix, represents the sum of squared errors of all corresponding point pairs; At the initial iteration, T = T0. At each iteration, T * As T at the current iteration, calculate T at the current iteration according to formula (3) * ; Through iterative optimization, until the conditions for stopping iteration are met, the optimal transformation matrix T is obtained. * ; Step 2.3.3: After the registration is completed, calculate the root mean square error; if the root mean square error is greater than the preset threshold, return to the initial registration step of the ICP algorithm and execute it again; The formula for calculating the root mean square error is: Where n is the number of point pairs, T * (Pt i ) is Pt i After the optimal transformation matrix T * The coordinates of the transformed points are the coordinates of the registered points.

4. The point cloud difference defect detection method based on multi-feature fusion according to claim 1 is characterized in that: The step 3 specifically includes: Step 3.1: For the reference point cloud P base and real-time point cloud P real For each point between , calculate the difference in Euclidean distance: Among them, P base (x i ,y i ,z i ) represents the coordinates of the i-th point in the reference point cloud, specifically (x base ,y base ,z base );P real (x i ,y i ,z i ) represents the coordinates of the i-th point in the real-time point cloud, specifically (x real ,y real ,z real );d i is the difference in Euclidean distance between the i-th point in the reference point cloud and the i-th point in the real-time point cloud; Step 3.2, for each point in the reference point cloud and the real-time point cloud, a local plane fitting method is used to calculate the curvature based on the neighborhood fitting of the point cloud; Local curvature formula: Among them, k i is the curvature of the ith point, R i For point P i The distance from the point to the least squares plane fitted in its neighborhood; According to formula (6), the curvature of the i-th point in the reference point cloud and the curvature of the i-th point in the real-time point cloud are calculated respectively, and the curvature of the i-th point in the reference point cloud and the curvature of the i-th point in the real-time point cloud are differentiated to calculate the curvature difference Δk i : Δk i =|k base,i -k real,i |(7); Among them, k base,i represents the curvature of the i-th point in the reference point cloud, k real,i Represents the curvature of the i-th point in the real-time point cloud; Step 3.3: For each point in the reference point cloud and the real-time point cloud, calculate its normal vector, the normal vector n i It can be expressed as: Among them, P i is the coordinate of the i-th point in the point cloud, P j is the coordinate of its neighborhood point, P avg is the average position of the neighborhood points; According to formula (8), the normal vector n of the i-th point in the reference point cloud is calculated respectively: i and the normal vector n of the i-th point in the real-time point cloud i , the normal vector of the reference point cloud and the normal vector of the real-time point cloud are differentiated to calculate the normal vector difference Δn i : Δn i =||n base,i -n real,i || (9); Where n base,i Represents the normal vector of the i-th point in the reference point cloud, n base,i Represents the normal vector of the i-th point in the real-time point cloud.

5. The point cloud difference defect detection method based on multi-feature fusion according to claim 4 is characterized in that: The step 4 specifically includes: Step 4.

1. Weight w based on Euclidean distance difference euc , the weight of the curvature difference w curv and the weight w of the normal vector difference norm The Euclidean distance difference d i , curvature difference Δk i and the normal vector difference Δn i Fusion is performed to generate a comprehensive anomaly score for the i-th point in the real-time point cloud. The formula is: S i =w euc d i +w curv Δk i +w norm Δn i (10); Step 4.2, according to the geometric features of the point cloud data, identify the type of the region where the i-th point in the current real-time point cloud is located. The region type can be divided into flat region, curved region and edge region; According to the area type, dynamically adjust the weights: When the area where the i-th point in the real-time point cloud is located is a plane area, increase the weight w of the Euclidean distance difference euc , weight w to reduce the curvature difference curv and the weight w to reduce the difference in normal vectors norm ; When the area where the i-th point in the real-time point cloud is located is a curved area, increase the weight of the curvature difference w curv , weight w to reduce the difference in normal vectors norm and weight w to reduce the difference in Euclidean distance euc ; When the area where the i-th point in the real-time point cloud is located is the edge area, increase the weight w of the normal vector difference norm , weight w to reduce the curvature difference curv and weight w to reduce the difference in Euclidean distance euc ; Step 4.3: S the comprehensive anomaly score of the i-th point in the real-time point cloud i Perform threshold determination. If S i is greater than the set threshold θ fusion , then the i-th point in the real-time point cloud is considered to be a defect point: S i >θ fusion (11)。 6. The point cloud difference defect detection method based on multi-feature fusion according to claim 1 is characterized in that: The step 5 specifically includes: Step 5.1, set the neighborhood radius ∈ and the minimum number of neighborhood points MinPts; For each defect point P i First, determine its neighborhood range, that is, all the neighbors related to the defect point P i The distance between the points is less than or equal to ∈, and the number of points in the neighborhood is recorded as N ∈ (P i ); If N ∈ (P i )≥MinPts, then the defect point P i is regarded as a core point; starting from the core point, the points in its neighborhood are grouped into the same cluster in turn until it cannot be expanded further; If N ∈ (P i ) < MinPts and is not within the neighborhood of any core point, then this defective point is regarded as a noise point and will not be classified into any cluster; The result of clustering is multiple clusters containing defect points, each cluster represents an independent defect area; Step 5.2: After completing the segmentation of the defect area, perform geometric quantitative analysis on each segmented defect area, including calculation of the area and depth of the defect area: The area A of each segmented defect area is calculated by the convex hull algorithm. defect ; The depth of the defect area refers to the depth difference between the deepest point in the defect area and the reference point cloud. Assume that the deepest point in the defect area is P deepest , whose coordinates are (x deepest ,y deepest ,z deepest ), the depth of the corresponding area in the reference point cloud is z base , then the depth of the defect area D defect for: D defect =z base -z deepest (12); Step 5.3: After steps 5.1 and 5.2, all segmented defect areas will be assigned corresponding geometric characteristics, including the area A of the defect area. defect and the depth D of the defect area defect .

7. The point cloud difference defect detection method based on multi-feature fusion according to claim 6 is characterized in that: The step 6 specifically includes: The area A of each defect area is obtained through steps 1, 2, 3, 4, and 5. defect , the depth of the defect area D defect and the location of the defective area; The real-time point cloud data is visualized through 3D rendering technology, and the area A of each defect area is calculated. defect , the depth of the defect area D defect And the position of the defect area, the defect area is marked on the 3D point cloud with different colors.

8. A point cloud differential defect detection system based on multi-feature fusion, characterized in that: include: An image acquisition module is used to obtain reference point cloud data of the designed standard model of the rail passenger car side wall panel, and is used to collect reference point cloud and real-time point cloud data of the rail passenger car side wall panel through a laser scanner; The image preprocessing module is used to preprocess the collected reference point cloud data and real-time point cloud data respectively, and the preprocessing includes denoising, downsampling and point cloud registration; A feature extraction module for calculating the Euclidean distance difference, curvature difference, and normal vector difference between the reference point cloud and the real-time point cloud; A dynamic threshold adjustment module is used to dynamically adjust the weight of the Euclidean distance difference, the weight of the curvature difference and the weight of the normal vector difference in the comprehensive anomaly score according to the regional characteristics, and to fuse the Euclidean distance difference, the curvature difference and the normal vector difference based on the weight of the Euclidean distance difference, the weight of the curvature difference and the weight of the normal vector difference to generate a comprehensive anomaly score, and to obtain all defect points in the real-time point cloud according to the comprehensive anomaly score; The defect segmentation module is used to classify points with spatial continuity into the same defect area based on defect points and through a clustering algorithm, thereby achieving segmentation of the defect area; A geometric quantitative analysis module is used to perform geometric quantitative analysis on each segmented defect area, including the area of ​​the defect area and the depth of the defect area; The result output and visualization module is used to visualize the analysis results obtained by the geometric quantitative analysis module through three-dimensional rendering technology.

Images