Three-dimensional data acquisition method for underground pipe gallery equipment based on SLAM and AI

Abstract

The present invention relates to the technical field of three-dimensional data acquisition methods, and in particular to a three-dimensional data acquisition method for underground pipe gallery equipment based on SLAM and AI, comprising the following steps: selecting a fire partition, scanning and acquiring high-precision point cloud data using a laser scanning instrument; manually marking the equipment type and location in the high-precision point cloud data, and exporting it as a labeled data set; training an AI model using the labeled data set, and forming a trained AI model; for the remaining fire partitions, scanning and acquiring low-precision point cloud data using a handheld SLAM instrument; matching feature points in the low-precision point cloud data with feature points in the high-precision point cloud data and correcting errors, and forming corrected point cloud data; using a trained AI model to repair the corrected point cloud data, and forming repaired point cloud data; inputting the repaired point cloud data into the trained AI model, and automatically marking the equipment type and location. The above provides more comprehensive and accurate spatial information.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional data acquisition methods, and in particular to a three-dimensional data acquisition method for underground pipe gallery equipment based on SLAM and AI. Background Art

[0002] Underground pipe corridors are long and wide, and the equipment in the corridors is mainly monitoring alarm, electrical, and fire-fighting equipment, including a large number of small devices such as sensors, cameras, control boxes, switches, etc. Taking a district-level pipe corridor with a single compartment of nearly 100 kilometers as an example, the number of equipment is about 50,000; if it is completely dependent on personnel entering the corridor for on-site inventory, the workload is huge, it is easy to miss, and it is impossible to obtain the exact location.

[0003] The utility corridor is generally designed to have multiple compartments horizontally, and each compartment is divided into multiple fire zones (about 200 meters) vertically. The type, quantity and layout of equipment in each zone are basically the same (which can be understood as several repeated linear spaces).

[0004] The application document with publication number CN115657069A discloses a three-dimensional map creation method, device and surveying equipment, the method comprising: based on SLAM technology, using laser radar to perform dynamic scanning to obtain a global map of the target area; determining the target local area in the global map; planning multiple static mapping control points in the target local area, wherein the control points are static scanning points; controlling the laser radar to move to each control point, and performing static scanning at each control point to obtain a target local map of the target local area; splicing the target local map on the global map to obtain a target three-dimensional map.

[0005] The existing technology combines static and dynamic scanning methods to obtain three-dimensional data. However, insufficient environmental constraints in degraded spaces such as pipe corridors can lead to positioning failures and reduce the accuracy of data acquisition. Summary of the invention

[0006] The purpose of the present invention is to improve the accuracy of data. In view of the above-mentioned shortcomings, a three-dimensional data acquisition method for underground pipe gallery equipment based on SLAM and AI is proposed.

[0007] The present invention adopts the following technical solution:

[0008] The method for collecting three-dimensional data of underground pipe gallery equipment based on SLAM and AI includes the following steps:

[0009] S1: Select a fire partition, scan it with a laser scanner and obtain high-precision point cloud data;

[0010] S2: Manually mark the equipment type and location in the high-precision point cloud data and export it as a labeled dataset;

[0011] S3: Use the labeled data set to train the AI ​​model and form a trained AI model;

[0012] S4: For the remaining fire partitions, use a handheld SLAM instrument to scan and obtain low-precision point cloud data;

[0013] S5: matching feature points in the low-precision point cloud data with feature points in the high-precision point cloud data and correcting errors to form corrected point cloud data;

[0014] S6: Use the trained AI model to repair the corrected point cloud data and form the repaired point cloud data;

[0015] S7: Input the repaired point cloud data into the trained AI model to automatically mark the device type and location.

[0016] Optionally, in step S3, the labeled data set includes geometric features, color features and texture features, and the AI ​​model is trained by a learning algorithm and using the geometric features, color features and texture features, so that the trained AI model can learn the characteristics of the device and mark the information of the device.

[0017] Optionally, in step S5,

[0018] Extract corner points, edge points, and surface points from low-precision point cloud data and high-precision point cloud data and match them through matching algorithms.

[0019] According to the matching results, the ICP algorithm and RANSAC algorithm are used to estimate the transformation matrix from low-precision point cloud data to high-precision point cloud data. According to the matching results, the ICP algorithm and RANSAC algorithm are used to align the low-precision point cloud data to the high-precision point cloud data. The corresponding transformation results are obtained according to the alignment and transformation matrix operations.

[0020] The error is corrected according to the transformation result, and the correction error includes position correction and attitude correction.

[0021] Optionally, in step S6, the trained AI model repair content includes:

[0022] Fill in missing details in the corrected point cloud data and improve edge resolution and clarity;

[0023] Identify and remove noisy points in the rectified point cloud data.

[0024] Optionally, in step S1, the laser scanning instrument is a Faro S350 laser scanning instrument.

[0025] Optionally, in step S2, manual marking is performed using point cloud processing software.

[0026] The beneficial effects achieved by the present invention are:

[0027] Based on the characteristics of the tunnel design with basically the same layout of multiple partitions, a laser scanner is used to scan and obtain high-precision point cloud data for one of the partitions, and a handheld SLAM instrument is used to scan and obtain low-precision point cloud data for the remaining partitions. The high-precision point cloud data is combined with AI technology to repair the obtained low-precision point cloud data, so as to obtain the data of the entire corridor, thereby reducing labor costs and improving the efficiency of tunnel data acquisition. Using a laser scanner to scan and obtain high-precision point cloud data can improve data accuracy and integrity, and using a handheld SLAM instrument to scan and obtain low-precision point cloud data can obtain data in real time, quickly, flexibly and comprehensively, thereby improving work efficiency. Combined with AI technology, the effect of repairing, analyzing and identifying data can be achieved. The fusion of the three data can form a relatively complete unit model, providing more comprehensive and accurate spatial information.

[0028] To further understand the features and technical contents of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are only for reference and description and are not intended to limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of the method of the present invention;

[0030] Figure 2 It is a schematic diagram of the structure of the laser scanning device in the present invention when it is working;

[0031] Figure 3 This is a schematic diagram of the structure obtained by scanning with a handheld SLAM instrument in the present invention;

[0032] Figure 4 It is a schematic diagram of the structure obtained by scanning with a laser scanning device in the present invention;

[0033] Figure 5 This is a flow chart of the method of Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0034] The following is an explanation of the embodiments of the present invention through specific embodiments. Those skilled in the art can understand the advantages and effects of the present invention from the contents disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depicted according to actual dimensions. It is stated in advance. The following embodiments will further explain the relevant technical contents of the present invention in detail, but the disclosed contents are not intended to limit the scope of protection of the present invention.

[0035] Embodiment 1: This embodiment provides a method for collecting three-dimensional data of underground pipe gallery equipment based on SLAM and AI, combined with Figures 1 to 4 shown.

[0036] The method for collecting three-dimensional data of underground pipe gallery equipment based on SLAM and AI includes the following steps:

[0037] S1: Select a fire partition, scan it with a laser scanner and obtain high-precision point cloud data;

[0038] S2: Manually mark the equipment type and location in the high-precision point cloud data and export it as a labeled dataset;

[0039] S3: Use the labeled data set to train the AI ​​model and form a trained AI model;

[0040] S4: For the remaining fire partitions, use a handheld SLAM instrument to scan and obtain low-precision point cloud data;

[0041] S5: matching feature points in the low-precision point cloud data with feature points in the high-precision point cloud data and correcting errors to form corrected point cloud data;

[0042] S6: Use the trained AI model to repair the corrected point cloud data and form the repaired point cloud data;

[0043] S7: Input the repaired point cloud data into the trained AI model to automatically mark the device type and location.

[0044] Specifically, in step S1, "select a fire partition" selects a typical fire partition. "Typical" can be understood as a representative fire partition, which is determined by technical personnel in this field through analysis of design drawings and on-site investigations. For example, the equipment in the fire partition is diverse, the structure within the fire partition is complex, the fire partition needs to connect multiple important functional areas, the fire partition has obvious spatial changes, and the fire partition is clearly marked as a standard section.

[0045] In step S1, the high-precision point cloud data obtained is three-dimensional data. The following describes in detail what data the three-dimensional data includes, including position coordinates, color information, intensity values, surface details, texture information, equipment layout, and space dimensions. The position coordinates are the three-dimensional coordinates (x, y, z) of each point, describing the specific position of the point in space; the color information is the color value of each point (if the laser scanner is equipped with a color camera), which helps to better understand and distinguish different objects; the intensity value is the reflection intensity value of each point, which indicates the intensity of the laser return signal and is often used to distinguish materials and surface characteristics; the surface details are the detailed structural information of the surfaces of walls, floors, ceilings, pipes, equipment, etc.; the texture information is the texture of the surface material, which helps to identify and distinguish different equipment and facilities; the equipment layout is to accurately record the location and layout of various types of equipment in the fire partition, such as pipes, cables, valves, cameras, wireless AP antennas, etc.; the space dimensions are the length, width, height and other size information of the partition, which is helpful for subsequent design and planning.

[0046] In step S1, the method of acquiring data after the laser scanner scans is: according to the size and complexity of the scanned partition, the appropriate resolution and range of the laser scanner is selected, and the laser scanner is placed at a certain position in the fire partition to ensure the horizontal stability of the laser scanner; the laser scanner rotates in steps at a fixed angle to collect data of the entire fire partition, and the data of multiple scanning positions are spliced ​​and aligned to generate high-precision point cloud data.

[0047] Although high-precision point cloud data can accurately represent the geometric shape of the device, it lacks semantic information, so it needs to be manually labeled in step S2; and the resulting labeled dataset is crucial, which can help the model learn how to identify and classify different devices from high-precision point cloud data. The labeled dataset can be saved as a JSON or CSV file.

[0048] In step S6, a convolutional neural network may be used for repair.

[0049] Optionally, in step S3, the labeled data set includes geometric features, color features and texture features, and the AI ​​model is trained by a learning algorithm and using the geometric features, color features and texture features, so that the trained AI model can learn the characteristics of the device and mark the information of the device.

[0050] Specifically, the learning algorithm can be TensorFlow or PyTorch. During training, the labeled data set is divided into a training set and a validation set, and then a suitable learning algorithm is selected for model training. The training set is used for training, and the validation set is used to evaluate and optimize the performance of the model, finally forming a trained AI model.

[0051] Optionally, in step S5,

[0052] Extract corner points, edge points, and surface points from low-precision point cloud data and high-precision point cloud data and match them through matching algorithms.

[0053] According to the matching results, the ICP algorithm and RANSAC algorithm are used to estimate the transformation matrix from low-precision point cloud data to high-precision point cloud data. According to the matching results, the ICP algorithm and RANSAC algorithm are used to align the low-precision point cloud data to the high-precision point cloud data. The corresponding transformation results are obtained according to the alignment and transformation matrix operations.

[0054] The error is corrected according to the transformation result, and the correction error includes position correction and attitude correction.

[0055] Specifically, corner points, edge points, and surface points are points of geometric features in low-precision point cloud data and high-precision point cloud data. In actual application, other points of geometric features can also be selected. The extraction method can be ISS key point detection or Harris corner point detection. After extracting the points of geometric features, the points of the set features are described by FPFH or SHOT method for subsequent matching. The purpose of matching is to find the corresponding point pairs of geometric features in low-precision point cloud data and high-precision point cloud data. The matching methods include nearest neighbor search, FLANN, and RANSAC. Since the handheld slam instrument will produce cumulative errors during the movement, it will result in the acquisition of The low-precision point cloud data obtained deviates from the actual position, so the position error can be corrected by matching and transformation; similarly, there is an error in the posture (rotation angle) of the handheld SLAM instrument, which causes the obtained low-precision point cloud data to rotate or tilt, so the posture error can be corrected by matching and transformation, and after processing in step S5, it is ensured that the corrected point cloud data and the high-precision point cloud data are in the same coordinate system, so that step S6 can be performed more effectively; the transformation matrix includes a rotation matrix and a translation vector, which represents the rigid transformation relationship between the low-precision point cloud data and the high-precision point cloud data. The obtained transformation matrix is ​​applied to the low-precision point cloud data, and then aligned to correct the posture and position errors.

[0056] Optionally, in step S6, the trained AI model repair content includes:

[0057] Fill in missing details in the corrected point cloud data and improve edge resolution and clarity;

[0058] Identify and remove noisy points in the rectified point cloud data.

[0059] Specifically, to fill in the missing details, interpolation methods, deep learning models or local surface reconstruction techniques can be used to generate new points in the missing areas; edge details in the corrected point cloud data can be enhanced through edge detection or super-resolution technology; noise points in the corrected point cloud data can be identified and removed by applying filtering algorithms, which can be Gaussian filtering or median filtering.

[0060] Optionally, in step S1, the laser scanning instrument is a Faro S350 laser scanning instrument.

[0061] Optionally, in step S2, manual marking is performed using point cloud processing software.

[0062] Specifically, the point cloud processing software may be Faro Scene, which is manually marked by a technician in this field to ensure the accuracy of the marking. The manual marking is in the form of adding a label to each corresponding device.

[0063] This embodiment solves the problem of low data accuracy in traditional acquisition methods through SLAM and AI technologies. It uses laser scanning in combination with a handheld SLAM instrument to obtain the initial point cloud data of the entire corridor, and uses AI technology to repair the low-precision point cloud data to obtain the final point cloud data of the entire corridor, which reduces the labor cost of traditional laser scanning instruments to obtain point cloud data of the entire corridor, and also avoids the problem of low accuracy of handheld SLAM instruments to obtain point cloud data of the entire corridor. Using laser scanning instruments to scan and obtain high-precision point cloud data can improve data accuracy and integrity, and using handheld SLAM instruments to scan and obtain low-precision point cloud data can obtain data in real time, quickly, flexibly, and comprehensively, improving work efficiency. Combining AI technology can achieve the effect of repairing, analyzing, and identifying data. The fusion of the three data can form a relatively complete unit model, providing more comprehensive and accurate spatial information.

[0064] Embodiment 2: This embodiment includes all the contents of embodiment 1, and provides a method for collecting three-dimensional data of underground pipe gallery equipment based on SLAM and AI. Figure 5 shown.

[0065] The method for collecting three-dimensional data of underground pipe gallery equipment based on SLAM and AI includes the following steps:

[0066] S1: Select a fire partition, scan it with a laser scanner and obtain high-precision point cloud data;

[0067] S2: Manually mark the equipment type and location in the high-precision point cloud data and export it as a labeled dataset;

[0068] S3: Use the labeled data set to train the AI ​​model and form a trained AI model;

[0069] S4: For the remaining fire partitions, use a handheld SLAM instrument to scan and obtain low-precision point cloud data;

[0070] S5: Match the feature points in the low-precision point cloud data with the feature points in the high-precision point cloud data and correct the errors to form the corrected point cloud data, use the Open3d library to detect the edge points of the corrected point cloud data and obtain the total number of edge points in the point cloud data before AI repair, use the filter to detect the noise points of the corrected point cloud data and obtain the total number of noise points in the point cloud data before AI repair;

[0071] S6: Use the trained AI model to repair the corrected point cloud data and form the repaired point cloud data, use the Open3D library to detect the edge points of the repaired point cloud data and obtain the total number of edge points in the point cloud data after AI repair, use the filter to detect the noise points of the repaired point cloud data and obtain the total number of noise points in the point cloud data after AI repair;

[0072] S7: Input the repaired point cloud data into the trained AI model to automatically mark the device type and location;

[0073] S8: Select the corresponding feature points in the corrected point cloud data and the repaired point cloud data through the key point detection algorithm and the matching algorithm, and obtain the total number of feature points. According to the total number of feature points, use the Open3d library to calculate the normal vectors of the corresponding feature points and obtain the normal vectors of each feature point in the point cloud data before AI repair and the normal vectors of each feature point in the point cloud data after AI repair.

[0074] S9: Obtain a repair qualification factor according to the total number of edge points in the point cloud data after AI repair, the total number of edge points in the point cloud data before AI repair, the total number of noise points in the point cloud data before AI repair, the total number of noise points in the point cloud data after AI repair, the total number of feature points, the normal vector of each feature point in the point cloud data before AI repair, and the normal vector of each feature point in the point cloud data after AI repair;

[0075] S10: Determine whether the AI ​​repair effect is qualified through the repair qualification factor.

[0076] Specifically, when the repair qualification factor is greater than or equal to the selection threshold of the repair qualification factor, it indicates that the effect of AI repair is qualified; when the repair qualification factor is less than the selection threshold of the repair qualification factor, it indicates that the effect of AI repair is unqualified; the selection threshold of the repair qualification factor is set by technical personnel in this field according to the repair standard.

[0077] In step S8, the key point detection algorithm can be an ISS algorithm to extract feature points from the corrected point cloud data and the repaired point cloud data, and then use a matching algorithm (nearest neighbor search algorithm) to match the feature points corresponding to the corrected point cloud data and the repaired point cloud data.

[0078] Optionally, in step S9, when calculating the repair qualification factor, the following formula is satisfied:

[0079]

[0080] Among them, Clarity is the repair qualification factor, epn after is the total number of edge points in the point cloud data after AI repair, epn b e f ore is the total number of edge points in the point cloud data before AI repair, qn b e f ore is the total number of noise points in the point cloud data before AI repair, qn a ft er is the total number of noise points in the point cloud data after AI repair, A is the total number of feature points,

[0081] The normal vector of the ath feature point in the point cloud data before AI repair, It is the normal vector of the ath feature point in the point cloud data after restoration by ai.

[0082] Specifically, "edge points" can be understood as points located at the edge of an object in point cloud data, which usually represent the geometric features of the object. When the edge points in the point cloud data increase after AI restoration, it means that the details are enhanced; the more obvious the change in the normal vector of the feature point in the point cloud data before and after AI restoration, the clearer the edge; It is expressed as the average difference between the normal vectors of the feature points in the point cloud data before and after AI restoration. It is expressed as the Euclidean distance between the normal vectors of the ath feature point in the point cloud data before and after AI restoration. The following principles are used for calculation: in They represent the x, y, and z components of the normal vector of the ath feature point in the point cloud data before AI restoration. They respectively represent the x, y, and z components of the normal vector of the ath feature point in the point cloud data after AI restoration; "noise points" can be understood as abnormal points in the point cloud data that do not conform to the overall distribution, which are usually caused by environmental interference, equipment errors and other problems. The total number of noise points in the point cloud data after AI restoration decreases, indicating that the restoration effect is good.

[0083] This embodiment adds the judgment of repair eligibility to solve the problem that the traditional collection method has relatively single functions. By judging whether the effect of AI repair is qualified through the repair qualification factor, the staff can make timely adjustments to improve the accuracy of data collection.

[0084] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention description and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.

Claims

1. A three-dimensional data acquisition method for underground pipe gallery equipment based on SLAM and AI, characterized in that: The following steps are involved: S1: Select a fire partition, scan it with a laser scanner and obtain high-precision point cloud data; S2: Manually mark the equipment type and location in the high-precision point cloud data and export it as a labeled dataset; S3: Use a labeled data set to train an AI model and form a trained AI model, wherein the labeled data set includes geometric features, color features, and texture features. The AI ​​model is trained by a learning algorithm and using geometric features, color features, and texture features, so that the trained AI model can learn the characteristics of the device and mark the information of the device; S4: For the remaining fire partitions, use a handheld SLAM instrument to scan and obtain low-precision point cloud data; S5: Match the feature points in the low-precision point cloud data with the feature points in the high-precision point cloud data and correct the errors to form the corrected point cloud data. Extract corner points, edge points, and surface points from low-precision point cloud data and high-precision point cloud data and match them through matching algorithms. According to the matching results, the ICP algorithm and RANSAC algorithm are used to estimate the transformation matrix from low-precision point cloud data to high-precision point cloud data. According to the matching results, the ICP algorithm and RANSAC algorithm are used to align the low-precision point cloud data to the high-precision point cloud data. The corresponding transformation results are obtained according to the alignment and transformation matrix operations. Correct errors according to the transformation results, and the correction errors include position correction and attitude correction; S6: Use the trained AI model to repair the corrected point cloud data and form the repaired point cloud data; S7: Input the repaired point cloud data into the trained AI model to automatically mark the device type and location.

2. The 3D data acquisition method for underground pipe gallery equipment based on SLAM and AI as claimed in claim 1 is characterized in that: In step S6, the trained AI model repair content includes: Fill in missing details in the corrected point cloud data and improve edge resolution and clarity; Identify and remove noisy points in the rectified point cloud data.

3. The 3D data acquisition method for underground pipe gallery equipment based on SLAM and AI as claimed in claim 2 is characterized in that: In step S1 , the laser scanning device is a Faro S350 laser scanning device.

4. The 3D data acquisition method for underground pipe gallery equipment based on SLAM and AI as claimed in claim 3 is characterized in that: In step S2, manual labeling is performed using point cloud processing software.

Images