Semi-automatic machine learning denoising method for point cloud of mobile laser scanning in mountain tunnels

Abstract

The present invention discloses a semi-automatic machine learning denoising method for mountain tunnel mobile laser scanning point cloud. For the mountain tunnel three-dimensional point cloud data collected by the mobile laser scanning system, the probability density distribution is used to statistically calculate the noise distribution of the typical mountain tunnel mobile three-dimensional laser scanning point cloud, and the point cloud noise characteristics of different types of interference objects are explained. Then, a semi-multiscale convolutional network denoising model is constructed to achieve point cloud data denoising. In addition to quickly and efficiently achieving point cloud denoising and providing a basis for the full-space deformation extraction of mountain tunnels, the proposed typical mountain tunnel lining interference point cloud noise characteristics can provide reference for tunnel point cloud data denoising of different types and different operation stages, and provide a reference for expanding the application of three-dimensional laser scanning in typical tunnel cross-section conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel scanning, and more specifically, the present invention relates to a semi-automatic machine learning denoising method for a mountain tunnel mobile laser scanning point cloud. Background Art

[0002] During the operation of mountain tunnels, as the years of operation increase, more and more problems are exposed in the tunnel structure, such as surface damage, structural cracks, bulging and other diseases, and serious accidents such as block falling and landslides may occur. The convergence deformation of the tunnel lining is a key indicator reflecting the safety of the tunnel structure. The mobile laser scanning module of the tunnel intelligent comprehensive inspection vehicle can complete the collection of deformation information of the tunnel space at a speed of 60 kilometers per hour. The obtained point cloud data is "large in quantity and mixed in quality", including the deformation information of the tunnel lining, as well as the noise information of vehicles, flue plates, decorative layers, etc. in the tunnel. The noise characteristics of point clouds of different mountain tunnels vary greatly. If the mountain tunnel point cloud is denoised manually, there are problems such as slow denoising speed and huge workload. Summary of the invention

[0003] The present invention provides a semi-automatic machine learning denoising method for a mountain tunnel mobile laser scanning point cloud, aiming to improve the denoising speed of the mountain tunnel laser point cloud.

[0004] The present invention is implemented as follows: a semi-automatic machine learning denoising method for a mountain tunnel mobile laser scanning point cloud, the method specifically comprising the following steps:

[0005] S1. Scan the entire mountain tunnel based on a mobile laser scanning system to obtain point cloud data of the mountain tunnel, and divide the point cloud of the mountain tunnel into N segments;

[0006] S2, remove the discrete noise point cloud, and then use the two-dimensional probability density inversion method to obtain the noise point cloud in the i-section mountain tunnel point cloud and put it into the training set;

[0007] S3, training the multi-scale convolutional neural network Model_CNN based on the samples in the training set, and inputting the point cloud data of the i+1th mountain tunnel into the trained multi-scale convolutional neural network Model_CNN_i;

[0008] S4. Compare the output noise point cloud recognition result with the manually calibrated noise point cloud to obtain the recognition error error of the multi-scale convolutional neural network Model_CNN_i for the i+1 section of the mountain tunnel i+1 ;

[0009] S5. If the recognition error i+1If the error is less than the error threshold, the multi-scale convolutional neural network Model_CNN_i is used to identify the noise point cloud in the remaining N-(i+1) sections of the mountain tunnel point cloud. If the recognition error is greater than or equal to the error threshold, i=i+1 is set and step S2 is executed.

[0010] Furthermore, the two-dimensional probability density inversion method is based on the radial and circumferential distribution of the noise point cloud in the cross section of the mountain tunnel. θ,r To identify the noise point cloud in the point cloud, the probability value P θ,r Greater than or equal to ε noise The point cloud with a value of P is a noise point cloud, and vice versa, it is a tunnel lining point cloud. θ,r represents the probability value of the point cloud at the radial r and circumferential θ of the cross section of the mountain tunnel being a noise point cloud, ε noise is the probability threshold.

[0011] Furthermore, the radial and circumferential distribution patterns of noise point clouds in the cross section of mountain tunnels are θ,r The specific method of obtaining is as follows:

[0012] The interference object data set Noise_set inside the mountain tunnel includes: the tunnel mileage Noise_mileage where the interference object is located, the interference object type Noise_type, the interference object size Noise_size, and the spatial position Noise_position of the interference object in the tunnel;

[0013] For the multi-source data of mountain tunnels, we first remove the discrete noise point cloud, and based on the mountain tunnel mileage pile number Noise_mileage, we begin to annotate the actual tunnel lining location information Ponitcloud_lining and the interference location information Ponitcloud_noise in the point cloud set Ponitcloud_mixture.

[0014] Based on the different reflectivity and spatial morphology of the interference object and the lining, the radial and circumferential distribution law of the noise point cloud in the cross section of the mountain tunnel is obtained. θ,r .

[0015] Furthermore, all point cloud data of the mountain tunnel are divided into N segments based on the mileage information obtained by the mobile laser scanning system.

[0016] For the 3D point cloud data of mountain tunnels collected by the mobile laser scanning system, the probability density distribution is used to statistically analyze the noise distribution of the point cloud of typical mountain tunnel mobile 3D laser scanning, and the point cloud noise characteristics of different types of interference objects are explained. Then, a semi-multiscale convolutional network denoising model is constructed to achieve point cloud data denoising. In addition to quickly and efficiently achieving point cloud denoising and providing a basis for the full-space deformation extraction of mountain tunnels, the proposed point cloud noise characteristics of typical mountain tunnel lining interference objects can provide a reference for the denoising of tunnel point cloud data of different types and different operation stages, and provide a reference for expanding the application of 3D laser scanning in typical tunnel sections. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a semi-automatic machine learning denoising method for a mountain tunnel mobile laser scanning point cloud provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0019] Based on the analysis of the characteristics of three-dimensional point cloud data of mobile laser scanning in mountain tunnels, the present invention proposes a semi-automatic point cloud denoising method based on a multi-scale convolutional neural network. This method can reduce the workload of manual noise identification as much as possible on the basis of accurately identifying the point cloud characteristics of interference objects inside the tunnel. By optimizing the multi-scale convolution model through segmented training, the effective use of three-dimensional point cloud data can be achieved. The proposed point cloud denoising method is not affected by the cross-sectional form of the tunnel, the operating time, and the type of internal interference objects. It can be applied to the denoising of mobile point cloud data of different mountain tunnels, laying the foundation for the rapid extraction of tunnel full-space deformation; the tunnel intelligent comprehensive inspection vehicle can realize the rapid identification and analysis of tunnel deformation, improve the technical service capabilities of the tunnel intelligent inspection vehicle, and timely perceive the full-space convergence deformation of the tunnel, providing technical support for the intelligent maintenance of mountain tunnels.

[0020] Figure 1 A flow chart of a semi-automatic machine learning denoising method for a mountain tunnel mobile laser scanning point cloud provided by an embodiment of the present invention, the method specifically comprises the following steps:

[0021] S1. Scan the entire mountain tunnel based on a mobile laser scanning system to obtain point cloud data of the mountain tunnel, and divide the point cloud of the mountain tunnel into N segments;

[0022] According to the mileage information obtained by the mobile laser scanning system, all point cloud data are divided into N segments (20m or 50m is used as the segmentation unit according to the length, and the remaining part is taken as 1 segment separately). The position data of the inspection vehicle is obtained by inertial navigation. The mobile laser scanning system and the inertial navigation system are integrated on the intelligent comprehensive inspection vehicle.

[0023] S2, remove the discrete noise point cloud, and then use the two-dimensional probability density inversion method to obtain the noise point cloud in the i-section mountain tunnel point cloud and put it into the training set;

[0024] The two-dimensional probability density inversion method is based on the radial and circumferential distribution of noise point clouds in the cross section of mountain tunnels. θ,r To identify the noise point cloud in the point cloud, the probability value P θ,r Greater than or equal to ε noise The point cloud that satisfies formula (1) is a noise point cloud:

[0025] P θ,r ≥ε noise (1)

[0026] Among them, P θ,r represents the probability value of the point cloud at the radial r and circumferential θ of the cross section of the mountain tunnel being a noise point cloud, ε noise is the probability threshold and is the empirical parameter.

[0027] Radial and circumferential distribution of the cross section of a mountain tunnel based on noise point cloud θ,r The formation method is as follows:

[0028] According to the multi-source data such as mountain tunnel design documents and photos taken by the panoramic camera of the mobile laser scanning system, the mountain tunnel internal interference object dataset Noise_set is constructed by artificial processing, including: the tunnel mileage Noise_mileage where the interference object is located, the interference object type Noise_type, the interference object size Noise_size, and the spatial position Noise_position of the interference object in the tunnel. The point cloud formed by the laser scanning interference object is the noise point cloud;

[0029] For the overly discrete noise point clouds such as vehicles or pedestrians, they can be removed first according to the data denoising function of the software system. The difficulty of subsequent denoising lies in the point cloud data near the tunnel lining. Based on the tunnel mileage Noise_mileage in the interference object data set, the actual tunnel lining position information Ponitcloud_lining and the interference object position information Ponitcloud_noise in the point cloud set Ponitcloud_mixture are marked. The relationship between the three is shown in formula (2).

[0030] Ponitcloud_mixture=Ponitcloud_lining+Ponitcloud_noise (2)

[0031] Based on the different reflectivity and spatial morphology of the interference object and the lining, the radial and circumferential distribution law of the noise point cloud in the cross section of the mountain tunnel is obtained. θ,r .

[0032] S3. Train the multi-scale convolutional neural network Model_CNN based on the samples in the training set to obtain the multi-scale convolutional neural network Model_CNN_i, and input the point cloud data of the i+1th mountain tunnel into the multi-scale convolutional neural network Model_CNN_i;

[0033] S4. Compare the noise point cloud recognition result in the i+1 mountain tunnel point cloud output by the multi-scale convolutional neural network Model_CNN_i with the manually recognized noise point cloud to obtain the recognition error error of the multi-scale convolutional neural network Model_CNN_i for the i+1 mountain tunnel. i+1 ;

[0034] S5. If the recognition error i+1 If the error is less than the error threshold, a multi-scale convolutional neural network is used to identify the noise point cloud in the remaining N-(i+1) sections of the mountain tunnel point cloud. If the recognition error is greater than or equal to the error threshold, i=i+1 is set and step S2 is executed.

[0035] First, for a mountain tunnel, all the point cloud data of the mountain tunnel are divided into N segments according to the mileage information obtained by the mobile laser scanning system (20m or 50m is used as the segmentation unit according to the length, and the remaining part is taken as a separate segment). The above formula (1) is used to determine the noise point cloud in the first mountain tunnel and add it to the training set. The multi-scale convolutional neural network Model_CNN based on the geometric expansion convolution kernel is realized by combining the DCT layer and the convolution layer. The spatial scale expansion search replaces the traditional convolution kernel translation search. The noise point cloud sample set in the first mountain tunnel is used to train the model Model_CNN_1. Then, the model Model_CNN_1 is used to identify the noise point cloud in the second mountain tunnel point cloud, and the identified point cloud is added to the training set. The identification result is compared with the manual identification result to calculate whether the error meets the requirements. If so, the model Model_CNN_1 is used to continue to identify the point cloud data of the remaining sections; if not, the noise point cloud in the second section of the mountain tunnel point cloud is identified based on formula (2) and added to the training set, the training model is re-optimized and trained, and the model Model_CNN_2 is obtained to repeat the above process until a multi-scale convolution model Model_CNN_i with a certain error that meets the requirements is obtained. The above process can effectively reduce the workload of manual identification of noise point clouds, and utilize the characteristic of deep learning models that the more data, the more accurate it is, to determine the balance point between accuracy and recognition efficiency in manual identification and intelligent identification of noise point clouds.

[0036] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A semi-automatic machine learning denoising method for mountain tunnel mobile laser scanning point cloud, characterized in that: The method specifically comprises the following steps: S1. Scan the entire mountain tunnel based on a mobile laser scanning system to obtain point cloud data of the mountain tunnel, and divide the point cloud of the mountain tunnel into N segments; S2, remove the discrete noise point cloud, and then use the two-dimensional probability density inversion method to obtain the noise point cloud in the i-section mountain tunnel point cloud and put it into the training set; S3, training the multi-scale convolutional neural network Model_CNN based on the samples in the training set to obtain the multi-scale convolutional neural network Model_CNN_i; S4. Input the point cloud data of the i+1th mountain tunnel into the multi-scale convolutional neural network Model_CNN_i, output the noise point cloud recognition result and compare it with the manually calibrated noise point cloud to obtain the recognition error error of the multi-scale convolutional neural network Model_CNN_i for the i+1th mountain tunnel. i+1 ; S5. If the recognition error i+1 If the error is less than the error threshold, the multi-scale convolutional neural network Model_CNN_i is used to identify the noise point cloud in the remaining N-(i+1) sections of the mountain tunnel point cloud. If the recognition error is greater than or equal to the error threshold, i=i+1 and step S2 is executed; The two-dimensional probability density inversion method is based on the radial and circumferential distribution of noise point clouds in the cross section of mountain tunnels. θ,r To identify the noise point cloud in the point cloud, the probability value P θ,r Greater than or equal to ε noise The point cloud with a value of P is a noise point cloud, and vice versa, it is a tunnel lining point cloud. θ,r represents the probability value of the point cloud at the radial r and circumferential θ of the cross section of the mountain tunnel being a noise point cloud, ε noise is the probability threshold.

2. The semi-automatic machine learning denoising method for mountain tunnel mobile laser scanning point cloud as claimed in claim 1, characterized in that: The radial and circumferential distribution patterns of noise point clouds in the cross section of mountain tunnels θ,r The specific method of obtaining is as follows: The interference object data set Noise_set inside the mountain tunnel includes: the tunnel mileage Noise_mileage where the interference object is located, the interference object type Noise_type, the interference object size Noise_size, and the spatial position Noise_position of the interference object in the tunnel; For the multi-source data of mountain tunnels, we first remove the discrete noise point cloud, and based on the mountain tunnel mileage pile number Noise_mileage, we begin to annotate the actual tunnel lining location information Ponitcloud_lining and the interference location information Ponitcloud_noise in the point cloud set Ponitcloud_mixture. Based on the different reflectivity and spatial morphology of the interference object and the lining, the radial and circumferential distribution law of the noise point cloud in the cross section of the mountain tunnel is obtained. θ,r .

3. The semi-automatic machine learning denoising method for mountain tunnel mobile laser scanning point cloud as claimed in claim 1, characterized in that: Based on the mileage information obtained by the mobile laser scanning system, all point cloud data of the mountain tunnel are divided into N segments.

Images