Top falling detection method and system based on lateral projection

Abstract

The invention provides a plug falling detection method and system based on lateral projection, and the method comprises the steps: carrying out the preprocessing of original collection point cloud data, reducing the point cloud data size, and eliminating the noise influence; performing lateral projection on the top data obtained by segmentation, reducing the data dimension, and obtaining a top lateral projection drawing; performing data annotation on the projection image, and constructing a deep learning detection model; reasoning the projection drawing based on the model, and judging the plug state. The technical problems that an optical plug detection system is easily interfered by impurities and is low in reliability and convenience are solved.

Description

Technical Field

[0001] The invention relates to the technical field of industrial detection, and in particular to a method and system for detecting a head drop based on lateral projection. Background Art

[0002] Seamless steel pipes have high strength and good corrosion resistance and are widely used in industrial fields such as petroleum, natural gas, and chemicals. PQF limited mandrel continuous rolling tube technology has been widely used in the global seamless steel pipe production due to its advantages of high efficiency, high yield rate, and excellent steel pipe quality. It mainly includes heating, perforation, leveling, diameter reduction, cooling, and straightening. Among them, perforation is the key process. The steel billet is formed in the center through the rotation of the roller and the push of the perforating head in the perforator. A hole. In daily production, the head may fall off due to high temperature and mechanical impact, wear of the connection between the head and the ejector, failure of the locking or ejector device, improper operation, etc. If the head falls and is not discovered in time, the ejector directly contacts the steel billet, which is very likely to cause damage to the ejector and steel roller equipment or even scrapping, increasing equipment maintenance costs and production stoppage costs. In the current production process, manual observation is mainly used to determine whether the plug has fallen. This method requires workers to concentrate for a long time, which can easily cause visual fatigue. In addition, the plug runs at a high speed, and it is difficult to judge its status by naked eyes. During the hot rolling process, the borax and water vapor raised on site will interfere with the detection effect. With the development of optical systems and automation technology, non-contact plug automatic detection methods based on machine vision have become a development trend.

[0003] The existing invention patent application document "A Steel Plant Head Detection System" with publication number CN115400982A, the existing system adopts a head detection solution based on 2D camera and deep learning, the solution grayscales the image, removes small spots in the image and enhances the head boundary line features through a preprocessing algorithm, and then inputs the image into a deep learning neural network model for judgment. Due to the use of a 2D imaging system, one dimension of information is lost when collecting data. And the existing invention patent application document "A Steel Plant Head Detection System" with publication number CN115400982A, the existing method includes: a position detection trigger device, a high-speed camera device, an industrial control host and an alarm device; the position detection trigger device, the high-speed camera device, and the alarm device are all connected to the industrial control host; the position detection trigger device is arranged at the front end of the high-speed camera device, and the high-speed camera device continuously shoots the preset position with the preset shooting parameters according to the start signal sent by the industrial control host, and transmits the obtained n pictures to the industrial control host; the head feature recognition model trained in the industrial control host recognizes the received picture and gives a conclusion whether there is a head feature in the picture; if there is no head feature in the picture, the industrial control host sends a trigger signal to the alarm device; the alarm device sends an alarm message to the outside according to the trigger signal of the industrial control host. However, the above-mentioned existing scheme may not be able to separate the noise from the target when processing data containing piece noise, resulting in false detection or missed detection.

[0004] In summary, the optical plug detection system in the prior art has the technical problems of being easily interfered by impurities and having low reliability and convenience. Summary of the invention

[0005] The technical problem to be solved by the present invention is: how to solve the technical problem that the optical plug detection system in the prior art is easily interfered by impurities and has low reliability and convenience.

[0006] The present invention adopts the following technical solutions to solve the above technical problems: The head drop detection method based on lateral projection includes:

[0007] S1, preprocessing the original point cloud data to obtain preprocessed point cloud data, and segmenting it to obtain vertex data;

[0008] S2, performing lateral projection on the plug data to reduce the data dimension and obtain a lateral projection image of the plug;

[0009] S3, annotate the data of the lateral projection image of the head and build a deep learning detection model;

[0010] S4. Reasoning the lateral projection image of the top head based on the deep learning detection model, judging the state of the top head, and obtaining the drop determination information;

[0011] S5. Perform a drop alarm based on the drop determination information.

[0012] The present invention is based on point cloud data collected by a 3D camera, and realizes data downsampling and segmentation of top data and water vapor noise data through point cloud filtering and a density-based clustering algorithm, which fully utilizes the spatial three-dimensional information of the point cloud and improves the denoising effect.

[0013] In a more specific technical solution, S1 includes:

[0014] S11, collecting and obtaining original point cloud data;

[0015] S12, downsampling the original point cloud data through voxel filtering to obtain a downsampled point cloud;

[0016] S13, density-based point cloud segmentation, based on the density of points within the domain, points with accessible density are classified into the same cluster, points with unreachable density are divided into new clusters, and all points are traversed until all points have corresponding clusters, thus obtaining segmented point clouds and segmented clusters;

[0017] S14, introduce preset conditions for the segmented point cloud, screen the segmented clusters according to the preset conditions, and obtain the target cluster, wherein for each cluster C o , calculate the coordinates of the center point (x c ,y c ,Z c );

[0018] S15. Determine cluster C i Whether the preset point condition and the preset center condition are met; i The coordinates of the center point (x c ,y c ,z c ), compared with the preset coordinate range;

[0019] S16. If yes, then for the current cluster C i Execute the retention operation;

[0020] S17: If not, a discard operation is performed on the current cluster.

[0021] In a more specific technical solution, S12, the three-dimensional space of the original point cloud data is divided into no less than 2 cubic voxels; the side length of the cubic voxel is defined using the three-dimensional vector in the following logic:

[0022] size=[l x l y l z ] T

[0023] In the formula, l x , ly , l z are the projected lengths of the cubic voxel on the x-axis, y-axis, and z-axis, respectively;

[0024] Regarding the three-dimensional space as a regular three-dimensional grid, for each point p in the point cloud i = [x i y i z i T , using the following voxel index calculation logic, determine the non-empty voxel to which the point cloud belongs according to the coordinates of the point cloud:

[0025]

[0026] Using the following logic, calculate the centroid of the voxel for each non-empty voxel:

[0027]

[0028] Use all the voxel centroids to form the downsampled point cloud.

[0029] In the point cloud filtering and point cloud segmentation operations of the present invention, the main purpose of filtering is to downsample the original point cloud, remove redundant data while ensuring data semantics and detailed information, and reduce the computational complexity of the entire processing flow. Point cloud segmentation is to distinguish the head data from the noisy data brought by external factors such as water vapor borax, and improve the data quality.

[0030] In a more specific technical solution, in S13, according to the connectivity state between point clouds, the point cloud is segmented into different clusters, and the status of all points is set to unvisited. For each point p, when the point is not visited, mark the status of the point as visited;

[0031] Using the following logic, calculate the number of points N ε (p) within the preset domain range ε of the point:

[0032] N ε (p) = {q ∈ D|dist(p, q) ≤ ε}

[0033] where D is the point cloud data set, and dist(p, q) represents the distance between p and q;

[0034] For the preset point number threshold minPts, if it satisfies: N ε (p) < minPts, then mark the point as a noise point, otherwise mark the point as a core point, create a new cluster according to the core point, and add the core point p and all the points within its own domain to the new cluster;

[0035] Expand the new cluster. For each point q in the new cluster, calculate the ε-neighborhood N ε ​(q), if N ε (q)>minPts, then all unvisited points in the neighborhood of point q are added to the current new cluster, and each density-reachable point is traversed until no new core point can be found.

[0036] The density-based point cloud segmentation of the present invention can segment the point cloud into different clusters according to the connectivity between the point clouds, thereby improving the segmentation effect of block noise points and reducing false detection or missed detection caused by a large amount of noise data.

[0037] In a more specific technical solution, in S16, if cluster C i Satisfies: x min ≤x c ≤x max And y min ≤y c ≤y max And z min ≤z c ≤z max , then the center of the cluster meets the preset center condition.

[0038] In a more specific technical solution, in S17, if cluster C i The center point of is not within the preset coordinate range:

[0039] [x ,in ,x max ],[x min ,x max ],[x min ,x max ], then according to the following logic, the cluster C is discarded. i :

[0040]

[0041] In a more specific technical solution, in S17, a minimum point threshold N is set. min , if cluster C i The points satisfy:

[0042] |C i |≥N min

[0043] Then keep the cluster C i ;

[0044] Otherwise, discard the cluster C. i .

[0045] In a more specific technical solution, S2 includes:

[0046] S21, input point cloud data;

[0047] S22. Calculate the coordinate range of the projected point cloud according to the point cloud data, wherein the boundaries of the y coordinate and the z coordinate in the point cloud data are calculated, the physical boundaries of the two-dimensional image are defined, and the size of the two-dimensional projected image is determined:

[0048] y min =min(y i ),y max =max(y i ),z min =min(z i ),z max =max(z i );

[0049] S23, calculating the image resolution in the y direction and the z direction according to the boundary value of the physical boundary of the two-dimensional image and the size of the two-dimensional projection image:

[0050]

[0051] S24, calculating coordinate projection;

[0052] S25, acquiring a lateral projection image, and obtaining the image pixel position through a coordinate conversion operation.

[0053] The present invention projects the processed point cloud data into a 2D image by lateral projection, and detects the 2D projection image based on the deep learning pattern recognition method to determine the state of the plug. This method not only improves the reliability and convenience of plug detection, but also provides new ideas for subsequent tasks such as plug melt tip detection and plug meat loss detection, and has a reference role in research in related fields.

[0054] In a more specific technical solution, in S25, the y coordinate and z coordinate of each point in the point cloud data are converted to the image pixel position (i, j) using the following mapping logic:

[0055]

[0056]

[0057] Since the plug has obvious longitudinal (y direction) and height (z direction) features, the yOz plane can well show these features and clearly display the outline and size of the plug. However, the data in the x direction is relatively redundant and contributes less to the judgment of the plug state. Selecting the yOz plane not only retains the main shape features of the plug, but also makes the features after data projection more intuitive, providing a basis for subsequent morphological recognition.

[0058] In a more specific technical solution, the head drop detection system based on lateral projection includes:

[0059] The point cloud preprocessing module is used to preprocess the original point cloud data, obtain the preprocessed point cloud data, and segment the point cloud data to obtain the vertex data;

[0060] A lateral projection module is used to perform lateral projection on the top data, reduce the data dimension, and obtain the top lateral projection image. The dimensionality reduction module is connected with the point cloud preprocessing module;

[0061] A model building module is used to annotate data of the head lateral projection image and build a deep learning detection model. The model building module is connected to the lateral projection module;

[0062] A drop determination module is used to infer the lateral projection image of the top head based on the deep learning detection model, determine the state of the top head, and obtain drop determination information. The drop determination module is connected to the model construction module and the lateral projection module;

[0063] The alarm module is used to perform a drop alarm according to the drop determination information, and the alarm module is connected to the drop determination module.

[0064] Compared with the prior art, the present invention has the following advantages:

[0065] The present invention is based on point cloud data collected by a 3D camera, and realizes data downsampling and segmentation of top data and water vapor noise data through point cloud filtering and a density-based clustering algorithm, which fully utilizes the spatial three-dimensional information of the point cloud and improves the denoising effect.

[0066] In the point cloud filtering and point cloud segmentation operations of the present invention, the main purpose of filtering is to downsample the original point cloud, remove redundant data while ensuring data semantics and detail information, and reduce the amount of calculation of the entire processing flow. Point cloud segmentation is to distinguish the top data from the noise data caused by external factors such as water vapor and borax, so as to improve data quality.

[0067] The density-based point cloud segmentation of the present invention can segment the point cloud into different clusters according to the connectivity between the point clouds, thereby improving the segmentation effect of block noise points and reducing false detection or missed detection caused by a large amount of noise data.

[0068] The present invention projects the processed point cloud data into a 2D image by lateral projection, and detects the 2D projection image based on the deep learning pattern recognition method to determine the state of the plug. This method not only improves the reliability and convenience of plug detection, but also provides new ideas for subsequent tasks such as plug melt tip detection and plug meat loss detection, and has a reference role in research in related fields.

[0069] Since the plug has obvious longitudinal (y direction) and height (z direction) features, the yOz plane can well show these features and clearly display the outline and size of the plug. However, the data in the x direction is relatively redundant and contributes less to the judgment of the plug state. Selecting the yOz plane not only retains the main shape features of the plug, but also makes the features after data projection more intuitive, providing a basis for subsequent morphological recognition.

[0070] The invention solves the technical problems in the prior art that the optical plug detection system is easily interfered by impurities and has low reliability and convenience. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of basic steps of a plug drop detection method based on lateral projection according to Embodiment 1 of the present invention;

[0072] Figure 2 This is a schematic diagram of specific steps of preprocessing the original point cloud according to Example 1 of the present invention;

[0073] Figure 3 This is a schematic diagram of specific steps of performing lateral projection according to Embodiment 1 of the present invention;

[0074] Figure 4 This is a picture taken by the camera of Embodiment 2 of the present invention;

[0075] Figure 5 Graphs showing the detection results at different time points of Example 2 of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] Example 1

[0078] like Figure 1 As shown, the head drop detection method based on lateral projection provided by the present invention includes the following basic steps:

[0079] S1. Preprocess the original collected point cloud data to reduce the amount of point cloud data and eliminate the influence of noise, and segment to obtain the top data;

[0080] like Figure 2 As shown, in this embodiment, the step S1 of preprocessing the original point cloud also includes the following specific steps:

[0081] S11. Obtain the original point cloud data;

[0082] S12. Perform voxel filtering on the original point cloud data;

[0083] In this embodiment, voxel filtering is used to downsample the point cloud. First, the three-dimensional space of the point cloud data is divided into multiple cubic voxels. The side length of the voxel is defined by a three-dimensional vector size = [l x l y l z T where l x , l y , l z are the projected lengths of the cubic voxel on the x-axis, y-axis, and z-axis respectively. Then the entire space can be regarded as a regular three-dimensional grid composed of several voxels. For each point p i = [x i y i z i T in the point cloud, determine which voxel it is located in according to its coordinates. The calculation method of the voxel index is as follows:

[0084]

[0085] For each non-empty voxel, calculate the centroid of the points in the voxel as the representative point. The calculation formula of the centroid is:

[0086]

[0087] Finally, the centroids of all voxels form the downsampled point cloud. Compared with the original point cloud, the number of points in the downsampled point cloud is significantly reduced, while the geometric features of the original point cloud are retained;

[0088] S13. Perform segmentation operation based on the density point cloud;

[0089] In this embodiment, set the status of all points to unvisited. Then for each point p, if the point has not been visited, mark the point as visited. And calculate the number of points N ε (p) within the neighborhood range ε of the point, which can be expressed as:

[0090] N ε (p) = {q ∈ D|dist(p, q) ≤ ε}

[0091] where D is the point cloud data set, and dist(p, q) represents the distance between p and q. For the preset point number threshold minPts, if N ε (p) < minPts, mark the point as a noise point. If N ε ​​(p)≥minPts, then mark the point as a core point, create a new cluster, add p and all the points in its domain to the cluster. Then expand the cluster, and for each point q in the cluster, calculate q's ε-neighborhood N ε (q), if N ε (q)>minPts, then add all unvisited points in the neighborhood of q to the current cluster. In this way, gradually traverse each density-reachable point until no new core point can be found. Repeat the above steps until all points are visited;

[0092] S14, calculating the point cloud center of each cluster point set;

[0093] In this embodiment, for the segmented point cloud, there may be some small clusters that do not meet the target requirements or noise clusters that deviate from the specified area. In order to further extract the target clusters that meet the requirements, the following conditions are introduced to screen the segmented clusters to ensure that the retained clusters meet the specific spatial range and density requirements. For each cluster C i , calculate the coordinates of its center point (x c ,y c ,z c );

[0094] S15, judging whether the point number and center conditions are met;

[0095] In this embodiment, each cluster C i The coordinates of the center point (x c ,y c ,z c ) is compared with the set coordinate range;

[0096] S16: If yes, perform a reservation operation on the current cluster;

[0097] In this embodiment, if the cluster center is not within the predefined range, the cluster is discarded. The specific calculation is as follows:

[0098]

[0099] Set the xyz range of the cluster center to [x min ,x max ],[x min ,x max ],[x min ,x max ].

[0100] If x min ≤x c ≤x max And y min ≤y c ≤y max And z min≤z c ≤z max , then the cluster center meets the conditions and proceeds to the next step. Otherwise, the cluster is discarded.

[0101] S17: If not, a discard operation is performed on the current cluster.

[0102] In this embodiment, in order to exclude small clusters caused by noise or sparse point cloud areas, clusters with points greater than a specific value are retained. A minimum point threshold N is set min , if cluster C i The points satisfy:

[0103] |C i |≥N min

[0104] If the cluster is not found, the cluster is retained; otherwise, the cluster is discarded.

[0105] S2, performing lateral projection on the plug data segmented in the above step S1 to reduce the data dimension and obtain a lateral projection image of the plug;

[0106] like Figure 3 As shown, in this embodiment, the step S2 of performing the lateral projection further includes the following specific steps:

[0107] S21, input point cloud data;

[0108] In this embodiment, due to the large amount of 3D point cloud data and high dimensionality, direct processing often requires a lot of computing resources and time. Projection is an effective means to simplify calculations. Projecting point cloud data onto a two-dimensional plane can significantly reduce the amount of data, making it easier to complete subsequent analysis operations quickly and at low cost. Although some spatial information is discarded through projection, this processing method can significantly shorten the calculation time while ensuring key features.

[0109] S22, calculating the coordinate range of the projected point cloud;

[0110] In this embodiment, the boundaries of the y coordinates and z coordinates in the point cloud data are calculated. The boundary values ​​define the physical boundaries of the two-dimensional image so as to determine the size and resolution of the image. The specific calculation formula is as follows:

[0111] y min =min(y i ),y max =max(y i ),z mim =min(z i ),z max =max(z i );

[0112] S23, calculating image resolution;

[0113] In this embodiment, the resolutions in the y and z directions are calculated according to the boundary value and the sizes W and H of the two-dimensional projection image:

[0114]

[0115] S24, calculating coordinate projection;

[0116] S25, acquiring a lateral projection image;

[0117] In this embodiment, the y and z coordinates of each point in the point cloud are converted to the pixel position (i, j) in the image. The mapping formula is as follows:

[0118]

[0119] S3, annotate the projected image data and build a deep learning detection model;

[0120] S4, reasoning the projection image based on the deep learning detection model, judging the state of the head, and determining whether the head falls;

[0121] S5. Activate a drop alarm.

[0122] In this embodiment, the head feature recognition model of the present invention is implemented based on a convolutional neural network. The equipment continuously collects and enriches intact and dropped images of the head under different working conditions, and gradually expands the training data set, thereby steadily improving the system's recognition accuracy and detection accuracy of the head features. The system adopts an online detection method, which effectively combines deep learning with machine vision functions: specifically, the machine vision module is responsible for image capture and processing, and obtains image information in the production process in real time; the deep learning module uses the GPU of the same host to analyze and classify image features and distinguish the head status. While ensuring the detection effect, this design greatly improves the overall computing efficiency and real-time response speed of the system, further ensuring the rapidity and reliability of production detection.

[0123] Example 2

[0124] like Figure 4 and Figure 5 As shown, Figure 4 , Figure 5 For the test image, Figure 5 The upper left picture in the middle is the image taken by the 3D camera, and the lower left picture is the image after lateral projection. From the lower left picture, it can be seen that there is a top head, which means that the top head has not fallen off. Figure 5 The display shows the detection results at different time points. Since the 3D camera shoots the top from top to bottom, the data of the lower half of the top is missing, so the image of the top after lateral projection is in a downward state. It should be noted here that Figure 4 , Figure 5 The time in the figure is the testing time, not the technology disclosure time.

[0125] In summary, the present invention is based on the point cloud data collected by a 3D camera, and realizes data downsampling and segmentation of top data and water vapor noise data through point cloud filtering and a density-based clustering algorithm, which makes full use of the spatial three-dimensional information of the point cloud and improves the denoising effect.

[0126] In the point cloud filtering and point cloud segmentation operations of the present invention, the main purpose of filtering is to downsample the original point cloud, remove redundant data while ensuring data semantics and detail information, and reduce the amount of calculation of the entire processing flow. Point cloud segmentation is to distinguish the top data from the noise data caused by external factors such as water vapor and borax, so as to improve data quality.

[0127] The density-based point cloud segmentation of the present invention can segment the point cloud into different clusters according to the connectivity between the point clouds, thereby improving the segmentation effect of block noise points and reducing false detection or missed detection caused by a large amount of noise data.

[0128] The present invention projects the processed point cloud data into a 2D image by lateral projection, and detects the 2D projection image based on the deep learning pattern recognition method to determine the state of the plug. This method not only improves the reliability and convenience of plug detection, but also provides new ideas for subsequent tasks such as plug melt tip detection and plug meat loss detection, and has a reference role in research in related fields.

[0129] Since the plug has obvious longitudinal (y direction) and height (z direction) features, the yOz plane can well show these features and clearly display the outline and size of the plug. However, the data in the x direction is relatively redundant and contributes less to the judgment of the plug state. Selecting the yOz plane not only retains the main shape features of the plug, but also makes the features after data projection more intuitive, providing a basis for subsequent morphological recognition.

[0130] The invention solves the technical problems in the prior art that the optical plug detection system is easily interfered by impurities and has low reliability and convenience.

[0131] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A head drop detection method based on lateral projection, characterized in that: The method comprises: S1, preprocessing the original point cloud data to obtain preprocessed point cloud data, and segmenting it to obtain vertex data; S2, performing lateral projection on the plug data to reduce the data dimension and obtain a lateral projection image of the plug; S3, annotating the data of the lateral projection image of the top head and building a deep learning detection model; S4, inferring the lateral projection image of the plug based on the deep learning detection model, judging the plug status, and obtaining drop determination information; S5. Perform a drop alarm according to the drop determination information.

2. The head drop detection method based on lateral projection according to claim 1 is characterized in that: The S1 includes: S11, collecting and acquiring the original point cloud data; S12, downsampling the original point cloud data through voxel filtering to obtain a downsampled point cloud; S13, density-based point cloud segmentation, based on the density of points within the domain, points with accessible density are classified into the same cluster, points with unreachable density are divided into new clusters, and all points are traversed until all points have corresponding clusters, thus obtaining segmented point clouds and segmented clusters; S14, introducing preset conditions for the segmented point cloud, screening the segmented clusters according to the preset conditions, and obtaining target clusters, wherein for each cluster C i , calculate the coordinates of the center point (x c ,y c ,z c ); S15, judging the cluster C i Whether the preset point condition and the preset center condition are met; wherein each of the clusters C i The coordinates of the center point (x c ,y c ,z c ), compared with the preset coordinate range; S16, if yes, then for the current cluster C i Execute the retention operation; S17: If not, a discard operation is performed on the current cluster.

3. The method for detecting head drop based on lateral projection according to claim 2, characterized in that: In S12, the three-dimensional space of the original point cloud data is divided into at least two cubic voxels; and the side length of the cubic voxel is defined by using the three-dimensional vector in the following logic: size=[l x l y l z ] T In the formula, l x , l y , l z are the projection lengths of the cube voxel on the x-axis, y-axis, and z-axis respectively; The three-dimensional space is regarded as a regular three-dimensional grid, and for each point p in the point cloud i =[x i y i z i ] T , using the following voxel index calculation logic, determine the non-empty voxels to which the point cloud belongs according to the coordinates of the point cloud: The following logic is used to calculate the centroid of each non-empty voxel: The downsampled point cloud is constructed using all the body centroids.

4. The head drop detection method based on lateral projection according to claim 2 is characterized in that: In S13, the point cloud is divided into different clusters according to the connectivity status between the point clouds, the status of all points is set to unvisited, and for each point p, when the point is not visited, the status of the point is marked as visited; The number of points N within the preset domain range ε ​​is calculated using the following logic: ε (p): N ε (p)={q∈D|dist(p,q)≤ε} Where D is the point cloud dataset, dist(p,q) represents the distance between p and q; For a preset point threshold minPts, if the following condition is met: N ε (p) < minPts, then mark the point as a noise point; otherwise, mark the point as a core point. Create a new cluster based on the core point, and add the core point p and all the points within its own neighborhood to the new cluster; Expand the new cluster, and for each point q in the new cluster, calculate the ε neighborhood N of the point q ε (q), if N ε (q)>minPts, then all unvisited points in the neighborhood of the point q are added to the current new cluster, and each density-reachable point is traversed until no new core point can be found.

5. The method for detecting head drop based on lateral projection according to claim 2, characterized in that: In S16, if the cluster C i Satisfies: x min ≤x c ≤x max And y min ≤y c ≤y max And z min ≤z c ≤z max , then the center of the cluster meets the preset center condition.

6. The method for detecting head drop based on lateral projection according to claim 2, characterized in that: In S17, if the cluster C i The center point is not within the preset coordinate range: [x min ,x max ],[x min ,x max ],[x min ,x max ], then according to the following logic, the cluster C is discarded. i :

7. The method for detecting head drop based on lateral projection according to claim 1, characterized in that: In S17, a minimum point threshold N is set. min , if the cluster C i The points satisfy: |C i |≥N min Then keep the cluster C i ; Otherwise, discard the cluster C i .

8. The method for detecting head drop based on lateral projection according to claim 1, characterized in that: The S2 includes: S21, input point cloud data; S22. Calculate the coordinate range of the projected point cloud according to the point cloud data, wherein the boundaries of the y coordinate and the z coordinate in the point cloud data are calculated, the physical boundaries of the two-dimensional image are defined, and the size of the two-dimensional projected image is determined: y min =min(y i ),y max =max(y i ),z min =min(z i ),z max =max(z i ); S23, calculating the image resolution in the y direction and the z direction according to the boundary value of the physical boundary of the two-dimensional image and the size of the two-dimensional projection image: S24, calculating coordinate projection; S25, acquiring a lateral projection image, and obtaining the image pixel position through a coordinate conversion operation.

9. The method for detecting head drop based on lateral projection according to claim 8, characterized in that: In S25, the y coordinate and z coordinate of each point in the point cloud data are converted to the image pixel position (i, j) using the following mapping logic:

10. The head drop detection system based on lateral projection is characterized by: The system comprises: The point cloud preprocessing module is used to preprocess the original point cloud data, obtain the preprocessed point cloud data, and segment the point cloud data to obtain the vertex data; A lateral projection module, used for performing lateral projection on the top data, reducing the data dimension, and obtaining a lateral projection image of the top, wherein the dimensionality reduction module is connected to the point cloud preprocessing module; A model building module, used for data annotation of the head lateral projection image and building a deep learning detection model, wherein the model building module is connected to the lateral projection module; A drop determination module, used for inferring the lateral projection image of the plug based on the deep learning detection model, judging the plug state, and obtaining drop determination information, wherein the drop determination module is connected to the model building module and the lateral projection module; An alarm module is used to perform a drop alarm according to the drop determination information, and the alarm module is connected to the drop determination module.

Images