Mine laneway environment data acquisition and summarization system based on 3D laser radar

Abstract

The invention discloses a mine laneway environment data acquisition and summarization system based on a 3D laser radar, and the system comprises a data acquisition and processing module which is responsible for receiving data from the 3D laser radar and an auxiliary sensor and processing the data, the processing comprises denoising, registration and reconstruction, the denoising is used for eliminating interference points in the data, and the registration is used for registering the data; the registration is used for aligning a plurality of pieces of scanning data to generate a complete roadway three-dimensional model, the registration process comprises an initial registration stage and a fine registration stage, in the initial registration stage, feature points and feature vectors thereof are extracted from point cloud data by using SIFT, then, the preliminarily extracted features are further optimized by using a deep learning model, and the feature vectors of the feature points and the feature vectors of the feature points are further optimized; the deep learning model automatically learns feature representation, the mismatching rate is reduced by learning a matching strategy, and in the fine registration stage, a dynamic iteration factor is introduced, and iterative optimization is performed on a preliminary registration result in a mode of gradually approaching an optimal solution.

Description

Technical Field

[0001] The present invention relates to the technical field of mine roadway environment data acquisition, and particularly to a mine roadway environment data acquisition and summary system based on 3D lidar. Background Art

[0002] Mine roadways are an important part of coal mine production, and their stability and safety are directly related to the production efficiency and personnel safety of coal mines. By scanning mine roadways with 3D lidar, three-dimensional point cloud data of the roadway surface is obtained, and then three-dimensional reconstruction of the roadway is realized. This helps coal mine enterprises to more intuitively understand the roadway structure and provide a scientific basis for roadway planning, design, and maintenance. Using the roadway environment data collected by 3D lidar, safety indicators such as roadway deformation and displacement can be monitored in real time, and potential safety hazards can be discovered in time. This helps coal mine enterprises to take measures in advance for repair and reinforcement to ensure the safety and stability of the roadway.

[0003] The mine roadway environment is complex and changeable, and traditional registration methods often struggle to adapt to this change, resulting in insufficient registration accuracy and robustness. The registration process consumes a large amount of computing resources and time, and how to improve the registration efficiency is also an urgent problem to be solved. Therefore, a mine roadway environment data acquisition and summary system based on 3D lidar is proposed. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems existing in the prior art, namely, the mine roadway environment is complex and changeable, traditional registration methods often struggle to adapt to this change, resulting in insufficient registration accuracy and robustness. The registration process consumes a large amount of computing resources and time, and how to improve the registration efficiency is also an urgent problem to be solved. A mine roadway environment data acquisition and summary system based on 3D lidar is proposed.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A mine roadway environment data acquisition and summary system based on 3D lidar, comprising:

[0007] 3D lidar acquisition device: Using a multi-beam 3D lidar as the core sensor, responsible for emitting laser beams and receiving reflected signals to obtain three-dimensional point cloud data in the roadway environment, which contains key information such as the three-dimensional shape, size, and deformation of the roadway;

[0008] Auxiliary sensors: Used to measure the attitude, acceleration, moving distance, and direction of the device;

[0009] Data Acquisition and Processing Module: Responsible for receiving data from 3D lidar and auxiliary sensors and processing it. The processing includes denoising, registration, and reconstruction. The denoising is used to eliminate interference points in the data. The registration is used to align multiple scan data to generate a complete 3D model of the roadway. The registration process includes an initial registration stage and a fine registration stage. In the initial registration stage, SIFT is used to extract feature points and their feature vectors from the point cloud data. Then, a deep learning model is used to further optimize the initially extracted features. The deep learning model automatically learns feature representations and reduces the false matching rate by learning a matching strategy. In the fine registration stage, a dynamic iteration factor is introduced, and the initial registration result is iteratively optimized by gradually approaching the optimal solution.

[0010] Data Summarization and Analysis Module: The processed 3D point cloud data is summarized in the data summarization and analysis module for visual display and quantitative analysis.

[0011] The above technical solution further includes:

[0012] Further, the 3D lidar acquisition device emits laser beams through a multi-beam 3D lidar and receives the reflected signals to generate 3D point cloud data of the roadway. The lidar emits laser beams through an internal laser emitter. The laser beams are reflected back after hitting the roadway surface and are captured by the lidar's receiver. By measuring the time difference between the emission and reception of the laser beams, the distance between the laser beams and the roadway surface is calculated. Combining the attitude information of the lidar, the 3D coordinates of the roadway surface are obtained. The calculation formula is expressed as where d represents the distance between the laser beam and the roadway surface, c represents the speed of light, and t represents the time difference between the emission and reception of the laser beam.

[0013] Further, the data acquisition and processing module includes a data receiving unit, a preprocessing unit, a denoising unit, a registration unit, and a reconstruction unit. The data receiving unit is responsible for receiving the raw data from the 3D lidar and auxiliary sensors, including laser ranging information, device attitude information, acceleration information, etc. The data receiving unit is directly connected to the 3D lidar and auxiliary sensors to receive and store data in real time. The preprocessing unit preprocesses the received raw data and calibrates the 3D lidar, including internal parameter calibration (such as transmitter position, angle, etc.) and external parameter calibration (such as relative position relationship with the auxiliary sensor). The preprocessing unit receives the raw data transmitted by the data receiving unit and transfers the data to the denoising unit after processing. The denoising unit eliminates the interference points in the data and transfers the data to the registration unit after denoising. The registration unit aligns multiple scan data to generate a complete three-dimensional model of the roadway. The registration process includes an initial registration stage and a fine registration stage. In the initial registration stage, the SIFT algorithm is used to extract feature points and their feature vectors from the point cloud data, and the deep learning model is used to further optimize the features. In the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the preliminary registration result. After the registration unit performs the registration process, it transfers the data to the reconstruction unit, and the reconstruction unit generates a three-dimensional model of the roadway based on the registered data.

[0014] Further, the external parameter calibration includes joint calibration. The joint calibration fuses the data of the 3D lidar and the auxiliary sensor to verify the accuracy of the calibration result. This is achieved by collecting a set of samples containing the data of the 3D lidar and the auxiliary sensor and comparing their corresponding relationships. If the corresponding relationships are inconsistent, the calibration is performed again.

[0015] Further, the specific steps for the denoising unit to eliminate the interference points in the data are as follows:

[0016] Data loading: The denoising unit loads the preprocessed data received from the preprocessing unit.

[0017] Noise detection: Identify and eliminate the noise points in the data through a statistics-based filtering method. Set a threshold, and consider the points whose distance from the average value exceeds a certain range as noise points, expressed as Noise_Point = {p_i | ∥p_i - μ∥ > σ × Threshold}, where p_i is the point in the point cloud, μ is the average value of the point cloud, σ is the standard deviation, and Threshold is the set threshold.

[0018] Noise elimination: Once the noise points are identified, the denoising unit deletes or replaces them from the data and replaces the noise points with the average value or median of the surrounding points to maintain the continuity of the data.

[0019] Post - processing: After eliminating the noise points, the denoising unit performs post - processing steps such as smoothing or resampling to further improve the data quality. These steps help reduce the artifacts generated by noise elimination and improve the overall readability of the data.

[0020] Furthermore, in the initial registration stage, the SIFT algorithm is used to extract feature points and their feature vectors from the point cloud data, including the following steps:

[0021] Constructing the scale space: Perform scale - space transformation on the point cloud data to generate a series of images with different scales. The purpose of the scale - space transformation is to detect key points at different scales, thereby enhancing the scale invariance of the algorithm. The scale - space transformation is expressed as L(x, y, σ) = G(x, y, σ) * I(x, y), where L is the scale - space image, G is the Gaussian function, I is the original image, σ is the scale parameter, and x and y are the data coordinates of the point cloud data.

[0022] Detecting key points: In the scale space, use the Difference of Gaussians (DoG) function to detect key points. Key points are points that are significant and stable at different scales, such as corner points, edge points, etc. The DoG function is expressed as D(x, y, σ) = (G(x, y, kσ) - G(x, y, σ)) * I(x, y), where D is the DoG image and k is a constant factor.

[0023] Calculating the direction information of key points: For each detected key point, calculate its gradient direction and gradient magnitude. The gradient direction is used to determine the direction of the key point, and the gradient magnitude is used for subsequent feature description.

[0024] Generating feature descriptors: Centered on the key points, divide several small regions (such as 4x4 small blocks) in their neighborhoods. Calculate the gradient direction and gradient magnitude for the points in each small region and perform statistics to generate a multi - dimensional (128 - dimensional for 4x4 small blocks) feature vector. This feature vector is the SIFT feature descriptor, which contains the local shape information of the key points.

[0025] Furthermore, using the deep - learning model to further optimize the features includes the following steps:

[0026] Data preparation: Collect the extracted feature points and their feature descriptors, and prepare label data. The label data is the known point - cloud registration result, that is, the correspondence of feature points between different scanning stations. This correspondence is obtained through manual annotation or pre - calculated using other registration algorithms.

[0027] Model training: Use the extracted feature points, their feature descriptors, and label data to train a deep learning model. During the training process, the deep learning model learns the mapping from feature descriptors to higher-level feature representations, thereby improving the matching accuracy of feature points;

[0028] Model inference: For new 3D lidar scan data, use the SIFT algorithm to generate feature points and their feature descriptors. Input the extracted feature descriptors into the trained deep learning model to obtain optimized feature representations. These optimized feature representations have stronger robustness and accuracy and can better adapt to the complexity of the mine roadway environment;

[0029] Feature matching: Use the optimized feature representations for feature matching;

[0030] Registration algorithm: Based on the matched feature point pairs, use the registration algorithm to calculate the transformation parameters between different scan sites, thereby achieving alignment.

[0031] Furthermore, the specific steps for the fine registration stage to introduce a dynamic iteration factor to iteratively optimize the preliminary registration result:

[0032] Introduce a dynamic iteration factor: Based on the initial registration stage, the fine registration stage introduces a dynamic iteration factor to iteratively optimize the registration result. The dynamic iteration factor α is used to adjust the transformation parameters in each iteration, gradually reducing the error between point clouds;

[0033] Iterative optimization: Let the transformation parameter after initial registration be T0 and the dynamic iteration factor be α. Then the transformation parameter Ti after the i-th iteration is calculated by Ti = Ti-1 + α * ΔTi, where ΔTi is the parameter adjustment amount in the i-th iteration, which is solved by minimizing the error function. The error function is usually defined based on the geometric distance or feature difference between point clouds and aims to measure the degree of mismatch between two point clouds;

[0034] Calculation of the parameter adjustment amount ΔTi: Obtained by using the gradient descent method to solve the minimized error function;

[0035] Convergence judgment: Set a convergence condition to determine when to stop the iteration. The convergence condition includes the error being less than a certain threshold and the number of iterations reaching a preset upper limit. When the convergence condition is met, the iteration process ends and the final registration result is obtained.

[0036] The present invention has the following beneficial effects:

[0037] In the present invention, the registration process is divided into a preliminary registration stage and a fine registration stage. In the preliminary registration stage, the SIFT algorithm is used to extract feature points and their feature vectors from the point cloud data, and a deep learning model is used to further optimize the features. By combining the SIFT algorithm and the deep learning model, accurate extraction and optimization of the point cloud data features can be achieved, thereby improving the registration accuracy. In the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the preliminary registration result, accelerating the convergence speed of the registration, and thus improving the accuracy, robustness, and efficiency of the registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 FIG. is a system block diagram of a mine roadway environment data acquisition and summary system based on a 3D lidar proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] Please refer to Figure 1 As shown, the present invention is a mine roadway environment data acquisition and summary system based on a 3D lidar, including:

[0041] 3D lidar acquisition device: Using a multi-beam 3D lidar as the core sensor, responsible for emitting laser beams and receiving reflected signals to obtain three-dimensional point cloud data in the roadway environment. These data contain key information such as the three-dimensional shape, size, and deformation of the roadway;

[0042] Auxiliary sensor: Used to measure the attitude, acceleration, moving distance, and direction of the device;

[0043] Data acquisition and processing module: Responsible for receiving data from the 3D lidar and the auxiliary sensor and performing processing. The processing includes denoising, registration, and reconstruction. The denoising is used to eliminate interference points in the data, and the registration is used to align multiple scan data to generate a complete three-dimensional model of the roadway. The registration process includes a preliminary registration stage and a fine registration stage. In the preliminary registration stage, the SIFT is used to extract feature points and their feature vectors from the point cloud data. Then, a deep learning model is used to further optimize the preliminarily extracted features. The deep learning model automatically learns feature representations and reduces the false matching rate by learning a matching strategy. In the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the preliminary registration result by gradually approaching the optimal solution;

[0044] Data Summarization and Analysis Module: The processed 3D point cloud data is summarized in the data summarization and analysis module for visual display and quantitative analysis.

[0045] In one embodiment, for the above 3D lidar acquisition device, the 3D lidar acquisition device emits laser beams through a multi-beam 3D lidar and receives reflected signals to generate 3D point cloud data of the roadway. The lidar emits laser beams through an internal laser emitter. After the laser beams encounter the roadway surface, they will be reflected back and captured by the receiver of the lidar. By measuring the time difference between the emission and reception of the laser beams, the distance between the laser beams and the roadway surface is calculated. Combining with the attitude information of the lidar, the 3D coordinates of the roadway surface are obtained. The calculation formula is expressed as where d represents the distance between the laser beam and the roadway surface, c represents the speed of light, and t represents the time difference between the emission and reception of the laser beam.

[0046] In one embodiment, for the above data acquisition and processing module, the data acquisition and processing module includes a data reception unit, a preprocessing unit, a denoising unit, a registration unit, and a reconstruction unit. The data reception unit is responsible for receiving the raw data from the 3D lidar and auxiliary sensors, including laser ranging information, device attitude information, acceleration information, etc. The data reception unit is directly connected to the 3D lidar and auxiliary sensors to receive and store data in real time. The preprocessing unit preprocesses the received raw data and calibrates the 3D lidar, including internal parameter calibration (such as emitter position, angle, etc.) and external parameter calibration (such as the relative position relationship with the auxiliary sensor). The preprocessing unit receives the raw data transmitted by the data reception unit and transfers the data to the denoising unit after processing. The denoising unit eliminates the interference points in the data and transfers the data to the registration unit after denoising. The registration unit aligns multiple scan data to generate a complete 3D model of the roadway. The registration process includes an initial registration stage and a fine registration stage. In the initial registration stage, the SIFT algorithm is used to extract feature points and their feature vectors from the point cloud data, and the features are further optimized using a deep learning model. In the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the preliminary registration result. After the registration unit performs the registration process, it transfers the data to the reconstruction unit. The reconstruction unit generates a 3D model of the roadway based on the registered data.

[0047] In one embodiment, for the above external parameter calibration, the external parameter calibration includes joint calibration. The joint calibration fuses the data of the 3D lidar and the auxiliary sensor to verify the accuracy of the calibration result. This is achieved by collecting a set of samples containing the data of the 3D lidar and the auxiliary sensor and comparing their corresponding relationships. If the corresponding relationships are inconsistent, the calibration is performed again.

[0048] In one embodiment, for the above denoising unit, the specific steps for the denoising unit to eliminate interference points in the data are as follows:

[0049] Data loading: The denoising unit loads the preprocessed data received from the preprocessing unit;

[0050] Noise detection: Identify and eliminate noise points in the data through a statistics-based filtering method. Set a threshold, and consider points whose distance from the average value exceeds a certain range as noise points, expressed as Noise_Point = {p_i ∣ ∥p_i - μ∥ > σ × Threshold}, where p_i is a point in the point cloud, μ is the average value of the point cloud, σ is the standard deviation, and Threshold is the set threshold;

[0051] Noise elimination: Once the noise points are identified, the denoising unit deletes or replaces them from the data, and replaces the noise points with the average value or median of the surrounding points to maintain the continuity of the data;

[0052] Post-processing: After eliminating the noise points, the denoising unit performs post-processing steps, such as smoothing or resampling, to further improve the data quality. These steps help reduce artifacts generated by noise elimination and improve the overall readability of the data.

[0053] In one embodiment, for the above initial registration stage, the initial registration stage uses the SIFT algorithm to extract feature points and their feature vectors from the point cloud data, including the following steps:

[0054] Construct a scale space: Perform a scale space transformation on the point cloud data to generate a series of images with different scales. The purpose of the scale space transformation is to detect key points at different scales, thereby enhancing the scale invariance of the algorithm. The scale space transformation is expressed as L(x, y, σ) = G(x, y, σ) * I(x, y), where L is the scale space image, G is the Gaussian function, I is the original image, σ is the scale parameter, and x and y are the data coordinates of the point cloud data;

[0055] Detect key points: In the scale space, use the Difference of Gaussians function to detect key points. Key points are points that are significant and stable at different scales, such as corner points, edge points, etc. The Difference of Gaussians function is expressed as D(x, y, σ) = (G(x, y, kσ) - G(x, y, σ)) * I(x, y), where D is the Difference of Gaussians image and k is a constant factor;

[0056] Calculate the direction information of the key points: For each detected key point, calculate its gradient direction and gradient magnitude. The gradient direction is used to determine the direction of the key point, and the gradient magnitude is used for subsequent feature description;

[0057] Generate feature descriptors: centered on the key points, divide several small regions (such as 4x4 small blocks) in their neighborhoods, calculate the gradient direction and gradient magnitude for the points in each small region, and perform statistics to generate a multi-dimensional (128-dimensional for 4x4 small blocks) feature vector, which is the SIFT feature descriptor and contains the local shape information of the key points.

[0058] In one embodiment, for the above-mentioned further optimization of features using a deep learning model, the further optimization of features using a deep learning model includes the following steps:

[0059] Data preparation: Collect the extracted feature points and their feature descriptors, and prepare label data, which is the known point cloud registration result, that is, the correspondence of feature points between different scanning stations, and the correspondence is obtained by manual annotation or pre-calculated using other registration algorithms;

[0060] Model training: Use the extracted feature points and their feature descriptors and label data to train the deep learning model. During the training process, the deep learning model learns the mapping from the feature descriptor to a more advanced feature representation, thereby improving the matching accuracy of the feature points;

[0061] Model inference: For new 3D lidar scan data, use the SIFT algorithm to generate feature points and their feature descriptors, input the extracted feature descriptors into the trained deep learning model, and obtain optimized feature representations. These optimized feature representations have stronger robustness and accuracy and can better adapt to the complexity of the mine roadway environment;

[0062] Feature matching: Use the optimized feature representations for feature matching;

[0063] Registration algorithm: Based on the matched feature point pairs, use the registration algorithm to calculate the transformation parameters between different scanning stations, so as to achieve alignment.

[0064] In one embodiment, for the above-mentioned fine registration stage, the specific steps of introducing a dynamic iteration factor to iteratively optimize the preliminary registration result:

[0065] Introduce a dynamic iteration factor: On the basis of the initial registration stage, the fine registration stage introduces a dynamic iteration factor to iteratively optimize the registration result. The dynamic iteration factor α is used to adjust the transformation parameters in each iteration, gradually reducing the error between the point clouds;

[0066] Iterative optimization: Let the transformation parameters after initial registration be T0, and the dynamic iteration factor be α. Then the transformation parameters Ti after the i-th iteration are calculated by the following formula: Ti = Ti-1 + α * ΔTi, where ΔTi is the parameter adjustment amount during the i-th iteration, which is solved by minimizing the error function. The error function is usually defined based on the geometric distance or feature difference between point clouds, aiming to measure the degree of mismatch between two point clouds;

[0067] Calculation of the parameter adjustment amount ΔTi: It is obtained by using the gradient descent method to solve the minimized error function;

[0068] Convergence judgment: Set a convergence condition to determine when to stop the iteration. The convergence condition includes that the error is less than a certain threshold and the number of iterations reaches a preset upper limit. When the convergence condition is met, the iteration process ends and the final registration result is obtained.

[0069] Suppose there are two adjacent scanning stations A and B in a mine roadway, and the point cloud data PA and PB are obtained respectively. After initial registration, there is still a certain error between PA and PB. At this time, the dynamic iteration factor α can be introduced, and the iterative optimization formula can be used for fine registration.

[0070] In each iteration, according to the current transformation parameters Ti-1 and the dynamic iteration factor α, the parameter adjustment amount ΔTi is calculated. Then, ΔTi is added to Ti-1 to obtain the new transformation parameters Ti. Repeat this process until the convergence condition is met.

[0071] Finally, when the iteration process ends, the final transformation parameters Tn will be obtained, which can align PA and PB with high precision under a common coordinate system. In this way, a complete 3D model of the roadway can be generated, providing strong support for the safe production and management of the mine.

[0072] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A mine roadway environment data acquisition and summary system based on 3D lidar, characterized in that, Including: 3D LiDAR acquisition device: Using a multi-beam 3D LiDAR as the core sensor, responsible for emitting laser beams and receiving reflected signals to obtain three-dimensional point cloud data in the roadway environment; Auxiliary sensors: Used to measure the attitude, acceleration, moving distance, and direction of the device; Data acquisition and processing module: Responsible for receiving data from the 3D LiDAR and auxiliary sensors and performing processing. The processing includes denoising, registration, and reconstruction. The denoising is used to eliminate interference points in the data, the registration is used to align multiple scan data to generate a complete three-dimensional model of the roadway. The registration process includes an initial registration stage and a fine registration stage. In the initial registration stage, SIFT is used to extract feature points and their feature vectors from the point cloud data. Then, a deep learning model is used to further optimize the initially extracted features. The deep learning model automatically learns feature representations and reduces the false matching rate by learning a matching strategy. In the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the initial registration result by gradually approaching the optimal solution; Data summarization and analysis module: The processed three-dimensional point cloud data is summarized in the data summarization and analysis module for visual display and quantitative analysis.

2. The mine roadway environment data acquisition and summary system based on 3D lidar according to claim 1, characterized in that, The 3D lidar acquisition device emits laser beams through a multi-beam 3D lidar and receives reflected signals to generate three-dimensional point cloud data of the roadway. The lidar emits laser beams through an internal laser emitter. After the laser beams encounter the roadway surface, they will be reflected back and captured by the receiver of the lidar. By measuring the time difference between the emission and reception of the laser beams, the distance between the laser beams and the roadway surface is calculated. The calculation formula is expressed as where d represents the distance between the laser beam and the roadway surface, c represents the speed of light, and t represents the time difference between the emission and reception of the laser beam.

3. The mine roadway environment data acquisition and summary system based on 3D lidar according to claim 1, characterized in that, The data acquisition and processing module includes a data reception unit, a preprocessing unit, a denoising unit, a registration unit, and a reconstruction unit. The data reception unit is responsible for receiving the raw data from the 3D LiDAR and auxiliary sensors. The data reception unit is directly connected to the 3D LiDAR and auxiliary sensors to receive and store data in real time. The preprocessing unit preprocesses the received raw data and calibrates the 3D LiDAR, including internal parameter calibration and external parameter calibration. The preprocessing unit receives the raw data transmitted by the data reception unit and transmits the data to the denoising unit after processing. The denoising unit eliminates the interference points in the data and transmits the data to the registration unit after denoising. The registration unit aligns multiple scan data to generate a complete three-dimensional model of the roadway. The registration process includes an initial registration stage and a fine registration stage. In the initial registration stage, the SIFT algorithm is used to extract feature points and their feature vectors from the point cloud data, and a deep learning model is used to further optimize the features. In the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the initial registration result. After the registration unit performs registration processing, it transmits the data to the reconstruction unit. The reconstruction unit generates a three-dimensional model of the roadway based on the registered data.

4. A mine roadway environment data acquisition and summary system based on a 3D lidar according to claim 3, characterized in that, The external parameter calibration includes joint calibration. The joint calibration fuses the data of the 3D LiDAR and the auxiliary sensors to verify the accuracy of the calibration result. It is achieved by collecting a set of samples containing the data of the 3D LiDAR and the auxiliary sensors and comparing their corresponding relationships. If the corresponding relationships are inconsistent, calibration is performed again.

5. The mine roadway environment data acquisition and summary system based on 3D lidar according to claim 3, characterized in that, The specific steps for the denoising unit to eliminate interference points in the data: Data loading: The denoising unit loads the preprocessed data received from the preprocessing unit; Noise Detection: Identify and eliminate noise points in the data through a statistics-based filtering method. Set a threshold, and consider points that deviate from the average by more than a certain range as noise points, expressed as Noise_Point = {p_i | ∥p_i - μ∥ > σ × Threshold}, where p_i is a point in the point cloud, μ is the average of the point cloud, σ is the standard deviation, and Threshold is the set threshold; Noise Elimination: Once the noise points are identified, the denoising unit deletes or replaces them from the data, and replaces the noise points with the average or median of the surrounding points; Post-processing: After eliminating the noise points, the denoising unit performs post-processing steps.

6. The mine roadway environment data acquisition and summary system based on 3D lidar according to claim 3, wherein, In the initial registration stage, the SIFT algorithm is used to extract feature points and their feature vectors from the point cloud data, including the following steps: Construct a scale space: Perform a scale space transformation on the point cloud data to generate a series of images with different scales. The scale space transformation is expressed as L(x, y, σ) = G(x, y, σ) * I(x, y), where L is the scale space image, G is the Gaussian function, I is the original image, σ is the scale parameter, and x and y are the data coordinates of the point cloud data; Detect key points: In the scale space, use the Difference of Gaussians function to detect key points. The Difference of Gaussians function is expressed as D(x, y, σ) = (G(x, y, kσ) - G(x, y, σ)) * I(x, y), where D is the Difference of Gaussians image and k is a constant factor; Calculate the direction information of the key points: For each detected key point, calculate its gradient direction and gradient magnitude. The gradient direction is used to determine the direction of the key point, and the gradient magnitude is used for subsequent feature description; Generate a feature descriptor: Centered on the key point, divide several small regions in its neighborhood. Calculate the gradient direction and gradient magnitude for the points in each small region and perform statistics to generate a multi-dimensional feature vector. The feature vector is the SIFT feature descriptor, which contains the local shape information of the key point.

7. The mine roadway environment data acquisition and summary system based on 3D lidar according to claim 3, characterized in that, The further optimization of the features using a deep learning model includes the following steps: Data preparation: Collect the extracted feature points and their feature descriptors, and prepare label data. The label data is the known point cloud registration result, that is, the correspondence between feature points between different scanning stations. The correspondence is obtained through manual annotation or pre-calculation using other registration algorithms; Model training: Use the extracted feature points, their feature descriptors, and label data to train the deep learning model. During the training process, the deep learning model learns the mapping from the feature descriptor to a higher-level feature representation; Model inference: For new 3D lidar scan data, use the SIFT algorithm to generate feature points and their feature descriptors. Input the extracted feature descriptors into the trained deep learning model to obtain an optimized feature representation; Feature matching: Use the optimized feature representation for feature matching; Registration algorithm: Based on the matched feature point pairs, use the registration algorithm to calculate the transformation parameters between different scanning stations.

8. The mine roadway environment data acquisition and summary system based on 3D lidar according to claim 3, wherein In the fine registration stage, introduce a dynamic iteration factor and perform iterative optimization on the preliminary registration result. The specific steps are as follows: Introduce a dynamic iteration factor: Based on the initial registration stage, in the fine registration stage, a dynamic iteration factor is introduced to iteratively optimize the registration result. The dynamic iteration factor α is used to adjust the transformation parameters in each iteration, gradually reducing the error between point clouds; Iterative optimization: Let the transformation parameter after initial registration be T0, and the dynamic iteration factor be α. Then the transformation parameter Ti after the i-th iteration is calculated by the formula Ti = Ti-1 + α * ΔTi, where ΔTi is the parameter adjustment amount in the i-th iteration, which is solved by minimizing the error function; Calculation of the parameter adjustment amount ΔTi: Obtained by using the gradient descent method to solve the minimized error function; Convergence judgment: Set a convergence condition to determine when to stop the iteration. The convergence condition includes that the error is less than a certain threshold and the number of iterations reaches a preset upper limit. When the convergence condition is met, the iteration process ends, and the final registration result is obtained.

Images