阅读说明：本技术 一种地铁隧道限界检测方法 (Subway tunnel clearance detection method ) 是由 郭春生 王维 付和宽 刘蝶 王令文 高志强 谢海燕 于 2019-11-19 设计创作，主要内容包括：本发明公开了一种地铁隧道限界检测方法,解决了目前对于隧道侵界的检测效率低、成本高的弊端,其技术方案要点是通过三维激光扫描检测装置扫描获取隧道断面点云,生成断面点云图的外接矩形框,将矩形框内的断面点云转换成断面图像；标记断面图像中隧道特征点以获得样本集；搭建基于卷积神经网络的回归模型,通过标记获得的样本集对回归模型进行训练、测试,再通过回归模型进行预测；获取轨道车的轮廓线,识别特征点以统一轨道车坐标与断面点云坐标并叠加,基于回归模型对轨道车在隧道内是否侵界进行判断,本发明的一种地铁隧道限界检测方法,能通过模型计算将轨道车和断面点云进行坐标系的统一,高效便捷的进行侵界的判断。(The invention discloses a subway tunnel limit detection method, which solves the defects of low detection efficiency and high cost of tunnel invasion at present, and adopts the technical scheme that a three-dimensional laser scanning detection device is used for scanning to obtain tunnel section point clouds to generate an external rectangular frame of a section point cloud image, and the section point clouds in the rectangular frame are converted into a section image; marking tunnel characteristic points in the cross-section image to obtain a sample set; building a regression model based on a convolutional neural network, training and testing the regression model through a sample set obtained by marking, and predicting through the regression model; the contour line of the rail car is obtained, the feature points are identified to unify the rail car coordinates and the point cloud coordinates of the cross section and are overlapped, and whether the rail car invades the boundary in the tunnel is judged based on the regression model.)

1. A subway tunnel limit detection method is characterized by comprising the following steps:

scanning and acquiring point clouds of a tunnel section by a three-dimensional laser scanning detection device, generating an external rectangular frame of a section point cloud picture, and converting the section point clouds in the rectangular frame into a section image;

marking tunnel characteristic points in the cross-section image to obtain a sample set;

building a regression model based on a convolutional neural network, training and testing the regression model through a sample set obtained by marking, and predicting through the regression model;

and acquiring the contour line of the rail car, identifying the feature points to unify the coordinates of the rail car and the point cloud coordinates of the section, superposing the coordinates, and judging whether the rail car invades the boundary in the tunnel based on a regression model.

2. A method of detecting the boundary of a subway tunnel according to claim 1, wherein the specific steps of marking the acquired sample set are:

selecting characteristic points in the cross-section image;

determining the origin of coordinates of the sectional image, and obtaining the coordinates of the characteristic points in the corresponding sectional image;

acquiring the coordinate proportion of the characteristic points in the section image;

and dividing the samples in the sample set into a training set and a testing set according to a set proportion.

3. A subway tunnel clearance detection method as claimed in claim 2, wherein: the selected characteristic points in the section image comprise the lowest point of the contact net at the top of the tunnel and the central points of the two steel rails.

4. A method as claimed in claim 3, wherein the training of the regression model specifically comprises:

training the regression model through a training set in the sample set;

performing parameter optimization by using a random gradient algorithm with self-adaptive learning rate;

the training times of the regression model are more than ten thousand times of the total amount of samples in the training set.

5. A subway tunnel clearance detection method as claimed in claim 4, wherein: the test process of the regression model through the sample set specifically comprises the following steps:

inputting the section image corresponding to the test set into a regression model for testing;

calculating and outputting the corresponding characteristic point coordinate proportion by the regression model;

restoring the coordinates of the feature points in the cross-section image according to the output feature point coordinate proportion;

selecting a precision index, comparing and calculating the coordinates of the characteristic points obtained by restoring in the current section image with the coordinates of the characteristic points obtained by corresponding marks, and judging whether the prediction of the current section image is accurate or not;

testing all section images of the test set, counting the accuracy, and judging that the regression model is accurate and can be used for prediction when the accuracy is greater than a set threshold; otherwise, the parameters are adjusted to continue training.

6. A subway tunnel clearance detection method as claimed in claim 5, wherein: the prediction and invasion judgment process specifically comprises the following steps:

selecting a cross-section image outside the sample set, and inputting the cross-section image into a regression model for calculation;

calculating and outputting the coordinate proportion of the corresponding characteristic points through a regression model;

calculating to obtain the coordinates of the characteristic points on the cross-section image according to the input cross-section image;

obtaining the coordinates of the feature points in a point cloud coordinate system according to the external rectangular frame coordinates corresponding to the section images;

acquiring a track wheel profile characteristic line, and matching the central coordinates of the track wheel in the track wheel profile characteristic line with the central points of the two steel rails in the point cloud coordinate to obtain the coordinates of the track wheel profile in the point cloud coordinate system;

detecting the coordinate relation between the lowest point of the tunnel top touch net and the contour of the rail car in the characteristic points, and judging that the boundary is invaded if the lowest point of the tunnel top touch net is invaded; if the lowest point of the tunnel top touch net does not invade the boundary, sequentially judging whether each point on the section is positioned in the railcar contour, and if any one point is positioned in the railcar contour, judging that the boundary is invaded; otherwise, it is not invaded.

Technical Field

The invention relates to tunnel detection, in particular to a subway tunnel clearance detection method.

Background

Along with the increase of the service time of the subway tunnel, the damage of the safe operation of the train caused by the invasion of the railway area by the auxiliary facilities in the tunnel needs to be carried out on the clearance of the subway tunnel in an irregular way to judge whether the clearance is invaded. The traditional invasion boundary detection method mainly adopts the means of a feeler lever, a tape measure, a total station and the like for measurement, and has low efficiency and high cost. In recent years, students at home and abroad carry out related research on rapid detection of tunnel limits. Chinese patent No. CN110006396 discloses a tunnel section and limit scanning detection device and method, which mainly adopts three-dimensional laser scanner, inertial navigation system, GPS antenna and encoder for positioning and detection, but the subway tunnel communication condition is not good, so it is difficult to have practicability.

Disclosure of Invention

The invention aims to provide a subway tunnel clearance detection method, which can unify a rail car and a section point cloud by a coordinate system through model calculation, and efficiently and conveniently judge the invasion.

The technical purpose of the invention is realized by the following technical scheme:

a subway tunnel clearance detection method comprises the following steps:

scanning and acquiring point clouds of a tunnel section by a three-dimensional laser scanning detection device, generating an external rectangular frame of a section point cloud picture, and converting the section point clouds in the rectangular frame into a section image;

marking tunnel characteristic points in the cross-section image to obtain a sample set;

building a regression model based on a convolutional neural network, training and testing the regression model through a sample set obtained by marking, and predicting through the regression model;

and acquiring the contour line of the rail car, identifying the feature points to unify the coordinates of the rail car and the point cloud coordinates of the section, superposing the coordinates, and judging whether the rail car invades the boundary in the tunnel based on a regression model.

Preferably, the specific steps of labeling the obtained sample set are as follows:

selecting characteristic points in the cross-section image;

determining the origin of coordinates of the sectional image, and obtaining the coordinates of the characteristic points in the corresponding sectional image;

acquiring the coordinate proportion of the characteristic points in the section image;

and dividing the samples in the sample set into a training set and a testing set according to a set proportion.

Preferably, the selected feature points in the cross-sectional image comprise the lowest point of the contact net at the top of the tunnel and the center points of the two steel rails.

Preferably, the training of the regression model specifically includes:

training the regression model through a training set in the sample set;

performing parameter optimization by using a random gradient algorithm with self-adaptive learning rate;

the training times of the regression model are more than ten thousand times of the total amount of samples in the training set.

Preferably, the test process of the regression model by the sample set specifically includes:

inputting the section image corresponding to the test set into a regression model for testing;

calculating and outputting the corresponding characteristic point coordinate proportion by the regression model;

restoring the coordinates of the feature points in the cross-section image according to the output feature point coordinate proportion;

selecting a precision index, comparing and calculating the coordinates of the characteristic points obtained by restoring in the current section image with the coordinates of the characteristic points obtained by corresponding marks, and judging whether the prediction of the current section image is accurate or not;

testing all section images of the test set, counting the accuracy, and judging that the regression model is accurate and can be used for prediction when the accuracy is greater than a set threshold; otherwise, the parameters are adjusted to continue training.

Preferably, the prediction and boundary violation judgment process specifically comprises the following steps:

selecting a cross-section image outside the sample set, and inputting the cross-section image into a regression model for calculation;

calculating and outputting the coordinate proportion of the corresponding characteristic points through a regression model;

calculating to obtain the coordinates of the characteristic points on the cross-section image according to the input cross-section image;

according to the external rectangular frame coordinates corresponding to the section image, obtaining the predicted coordinates of the feature points in a point cloud coordinate system;

acquiring a track wheel profile characteristic line, and matching the central coordinates of the track wheel in the track wheel profile characteristic line with the central points of the two steel rails in the point cloud coordinate to obtain the coordinates of the track wheel profile in the point cloud coordinate system;

detecting the coordinate relation between the lowest point of the tunnel top touch net and the contour of the rail car in the characteristic points, and judging that the boundary is invaded if the lowest point of the tunnel top touch net is invaded; if the lowest point of the tunnel top touch net does not invade the boundary, sequentially judging whether each point on the section is positioned in the railcar contour, and if any one point is positioned in the railcar contour, judging that the boundary is invaded; otherwise, it is not invaded.

In conclusion, the invention has the following beneficial effects:

the tunnel section point cloud is directly obtained through scanning and converted into an image, the identification matching of the feature points is carried out through the built regression model, the section point cloud and the rail car coordinate system are unified, whether the rail car invades the boundary can be directly judged through the geometric relation, the efficiency is higher, and the operation is convenient and fast.

Drawings

FIG. 1 is a cross-sectional image obtained by transformation;

FIG. 2 is a schematic diagram of a cross-sectional image and feature points;

fig. 3 is a schematic diagram of a track wheel contour line and a cross-sectional image coordinate system after superposition.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings.

The method for detecting the boundary of the subway tunnel disclosed in the embodiment includes the following specific steps.

1. And acquiring the point cloud of the tunnel section by adopting a three-dimensional laser scanning detection device, and converting the acquired point cloud of the section into a tunnel section picture. After the cross-section point cloud picture is obtained, as shown in fig. 1, an external rectangular frame of the cross-section point cloud is generated, and the cross-section point cloud in the rectangular frame is converted into a cross-section image.

The detection device can be a static tripod or a platform, and can also be a detection vehicle driven by manpower or a motor.

2. And acquiring a tunnel boundary feature point sample set for convolutional neural network training. The tunnel boundary characteristic points include a plurality of points, wherein, as shown in fig. 2, a tunnel top touching net lowest point a, two steel rail center position points B1 and B2 are mainly selected as the characteristic points. The specific process of obtaining the sample set is as follows:

a) and sample marking: selecting the lower left corner of the image as the origin of coordinates, and obtaining the coordinates of the tunnel top net touching lowest point A, two steel rail central position points B1 and B2 which are selected characteristic points in the tunnel section image, so that one image corresponds to the coordinate proportion of three points in the format of

Where W represents the width of the current sectional image and H represents the height of the current sectional image. The original section image set is S, the section image set marked with the characteristic points is L, and the original section image set and the section image set are equal in total amount and correspond to each other one by one;

b) dividing a sample set: dividing a sample set (S, L) into a training set T₁And test set T₂Preferably, the training set is divided into 70% of the total number of the sample set, and the test set is divided into 30% of the total number of the sample set.

3. Building a regression Model based on a convolutional neural network, and specifically operating as follows:

selecting a deep learning framework: using a deep learning framework such as Tensorflow, PyTorch, etc.;

building a model: the method is realized by improving a ResNet-101 network architecture, a last 1000-d full connection layer and a softmax layer are removed, and a low-dimensional full connection layer of a 6-d full connection layer and a Sigmoid layer of a Sigmoid activation function are added again;

4. and (5) training, testing and predicting the regression model.

a) And training: the data set is a training set T₁The number of training times is at least the training set T₁10000 times of the total amount, using a random gradient algorithm with self-adaptive learning rate to optimize parameters, adopting a batch normalization method for accelerating learning, relieving gradient dissipation, and adopting GPU hardware acceleration to accelerate calculation; the Adam algorithm is preferably adopted, is an optimization algorithm for replacing the traditional random gradient descent SGD process, and is a common learning rate self-adaptive optimization algorithm in the field of deep learning at present.

b) And testing: the data set is a test set T₂And calculating by using the Euclidean distance as an accuracy index through a regression model, wherein when the Euclidean distance meets a set certain threshold value, the current image is accurately predicted and recorded as 1, otherwise, the Euclidean distance is 0, and the optimal threshold value is set to be less than or equal to 0.1. When all data are tested, counting the accuracy;

where, Σ T represents the number of times the image is predicted accurately, Count (T)₂) Representing the test set size.

When the accuracy is greater than a certain threshold, the method can be used for predicting whether the parameters are adjusted to continue training; the threshold value of the accuracy is preferably set to 95% or more.

The specific test procedure is as follows:

inputting the section image corresponding to the test set into a regression model for testing;

calculating and outputting the corresponding characteristic point coordinate proportion through a regression model;

restoring the coordinates of the feature points in the cross-section image according to the output feature point coordinate proportion;

selecting a precision index, comparing and calculating the coordinates of the characteristic points obtained by restoring in the current section image with the coordinates of the characteristic points obtained by corresponding marks, and judging whether the prediction of the current section image is accurate or not;

testing all section images of the test set, counting the accuracy, and judging that the regression model is accurate and can be used for prediction when the accuracy is greater than a set threshold; otherwise, the parameters are adjusted to continue training.

c) And predicting: selecting tunnel section images which are not in the training set and the test set as input, calculating through a regression model, and outputting corresponding coordinate proportion in a format of

And (3) knowing W, H of the input sectional image, reversely deducing coordinates of the characteristic points A ', B ' 1 and B ' 2, knowing coordinates of a circumscribed rectangle under a point cloud coordinate system, and reversely calculating the coordinates of the characteristic points A ', B ' 1 and B ' 2 to new coordinates A ', B ' 1 and B ' 2 under the point cloud coordinate system.

5. And superposing the rail car and the point cloud coordinate system to judge whether to invade the boundary. When the feature line of the wheel profile of the rail car is known, as shown in fig. 3, the wheel center coordinates of the rail car are matched with B "1 and B" 2 of the feature points obtained by calculation under the new coordinates, so that the coordinates of the wheel profile of the rail car under the point cloud coordinate system can be obtained. Generally, whether a coordinate of a lowest point A' of a touching net at the top of the tunnel is invaded is detected, if so, the invaded boundary is directly judged, if not, whether each point on the cross section is in the track wheel contour range is sequentially compared, and if so, the invaded boundary is considered.

The present embodiment is only for explaining the present invention, and it is not limited to the present invention, and those skilled in the art can make modifications of the present embodiment without inventive contribution as needed after reading the present specification, but all of them are protected by patent law within the scope of the claims of the present invention.