阅读说明：本技术 一种大场景道路基础设施数字化模型快速构建方法 (Rapid construction method of large-scene road infrastructure digital model ) 是由 李思李 田波 张盼盼 权磊 李立辉 谢晋德 何哲 于 2021-09-22 设计创作，主要内容包括：本发明公开一种大场景道路基础设施数字化模型快速构建方法,包括,获取道路的初始点云数据,其中道路包括路段,隧道,桥梁及颠坡；对初始点云数据通过进行预处理,获取若干个点云簇；构建检测模型,通过检测模型分别对若干个点云簇进行特征识别；基于识别结果,对点云簇进行标注,对标注好的点云簇进行三维建模,得到大场景道路基础设施数字化模型。本发明通过上述技术方案能够有效的对大场景道路基础设施数字化模型快速建模,同时建立的模型中能够准确的标识出不同的基础设施类别,具有较强的实用性。(The invention discloses a rapid construction method of a large-scene road infrastructure digital model, which comprises the steps of obtaining initial point cloud data of a road, wherein the road comprises road sections, tunnels, bridges and jolting slopes; preprocessing initial point cloud data to obtain a plurality of point cloud clusters; constructing a detection model, and respectively carrying out feature recognition on a plurality of point cloud clusters through the detection model; and marking the point cloud cluster based on the identification result, and performing three-dimensional modeling on the marked point cloud cluster to obtain a large-scene road infrastructure digital model. According to the technical scheme, the digital model of the infrastructure of the large-scene road can be effectively and quickly modeled, different infrastructure types can be accurately identified in the established model, and the method has high practicability.)

1. A method for quickly constructing a digital model of a large-scene road infrastructure is characterized by comprising the following steps:

acquiring initial point cloud data of a road, wherein the road comprises a road section, a tunnel, a bridge and a jolt slope;

preprocessing the initial point cloud data through voxel filtering and an Euclidean clustering algorithm to obtain a plurality of point cloud clusters;

constructing a detection model, and respectively carrying out feature recognition on a plurality of point cloud clusters through the detection model;

and marking a plurality of point cloud clusters based on the identification result, and performing three-dimensional modeling on the marked point cloud clusters to obtain a large-scene road infrastructure digital model.

2. The large-scale road infrastructure digital model rapid construction method according to claim 1,

the step of acquiring initial point cloud data of a road comprises the following steps:

scanning the road section and the tunnel through a vehicle-mounted laser radar to obtain point cloud data of the road section and the tunnel;

scanning a bridge floor of a bridge through a vehicle-mounted laser radar to obtain bridge floor point cloud data, scanning a bridge bottom of the bridge through a fixed laser radar to obtain bridge bottom point cloud data, and integrating the bridge floor point cloud data and the bridge bottom point cloud data to obtain bridge point cloud data;

scanning a slope by carrying a laser radar on an unmanned aerial vehicle to obtain slope-turning point cloud data;

and integrating the road section and tunnel point cloud data, the bridge point cloud data and the slope point cloud data to obtain initial point cloud data of the road.

3. The large-scale road infrastructure digital model rapid construction method according to claim 1,

the step of preprocessing by the voxel filtering comprises:

the method comprises the steps of constructing a three-dimensional voxel space based on initial point cloud data, constructing a three-dimensional voxel grid in the three-dimensional voxel space, putting the initial point cloud data into the corresponding three-dimensional voxel grid, obtaining a grid gravity center based on the three-dimensional voxel grid, and filtering the initial point cloud data based on the grid gravity center to obtain filtered point cloud data.

4. The large-scale road infrastructure digital model rapid construction method according to claim 3,

the step of preprocessing by the Euclidean clustering algorithm comprises the following steps:

converting the filtered point cloud data into a kd-tree data structure to obtain a point cloud data set;

selecting any point cloud data in the point cloud data set as an initial fixed point, performing Euclidean distance judgment and smooth parameter judgment on the point cloud data except the initial fixed point in the point cloud data set based on the initial fixed point, acquiring point cloud clusters based on a judgment result, repeating the steps based on the point cloud data except the point cloud clusters in the point cloud data set, and obtaining a plurality of point cloud clusters when the initial fixed point cannot be selected.

5. The large-scale road infrastructure digital model rapid construction method according to claim 1,

in the process of constructing the detection model, the detection model comprises a rotation alignment network, a global pooling layer, a first convolution layer, a second convolution layer, a first graph convolution layer, a second graph convolution layer and a third graph convolution layer.

6. The large-scale road infrastructure digital model rapid construction method according to claim 5,

the step of respectively carrying out feature recognition on a plurality of point cloud clusters through a detection model comprises the steps of,

respectively inputting the point cloud clusters into the detection model, aligning the point cloud clusters through the rotary alignment network, processing the aligned point cloud clusters through the first graph convolution layer to obtain first structure information, processing the aligned point cloud clusters through the first graph convolution layer to obtain first characteristic information, fusing the first structure information and the first characteristic information, processing a fusion result through the second graph convolution layer to obtain second structure information, processing the first characteristic information through the second convolution layer to obtain second characteristic information, fusing the second structure information and the second characteristic information again, inputting the re-fusion result into the third graph convolution layer to obtain a characteristic matrix, and aggregating the characteristic matrix through the global pooling layer to obtain an identification result, wherein, the identification result is the types of different infrastructures in the road.

7. The large-scale road infrastructure digital model rapid construction method according to claim 1,

the method further comprises the steps of performing feature recognition on a plurality of point cloud clusters through the detection model,

acquiring a point cloud data training set, wherein the point cloud data training set comprises all target features in the initial point cloud data;

and training the detection model through the point cloud data training set.

8. The large-scale road infrastructure digital model rapid construction method according to claim 1,

and performing three-dimensional modeling on the marked point cloud cluster through BIM software, and marking different infrastructures in the modeled model based on a marking result in the three-dimensional modeling process.

Technical Field

The invention relates to the technical field of digital model construction, in particular to a rapid construction method of a digital model of a large-scene road infrastructure.

Background

The digitization of road infrastructure belongs to an important component of an intelligent road, and the digitization of the road infrastructure needs to deeply apply a digital technology in a road traffic system and construct a digital road world with full perception, full connection, full scene and full intelligence by using a computer system. However, in the prior art, the technology for digitally modeling road infrastructure is mainly modeled by images captured by a GPS or GIS system, due to the problems of the accuracy limit of the shot images and the shooting angle, the infrastructure in the road scene cannot be accurately modeled in the modeling process based on the shot images, the prior art of digitally modeling road infrastructure also includes collecting point cloud data for modeling, which can collect more detailed data than image modeling, however, the conventional modeling needs to collect a large amount of point cloud data in a road infrastructure scene, which results in a slow modeling speed due to the excessive amount of point cloud data in the modeling process, meanwhile, in a model established by the point cloud data, direct facility type checking cannot be performed on virtual infrastructure in the model.

Disclosure of Invention

In order to solve the problems that the modeling speed is low, direct facility type checking cannot be performed on virtual infrastructures in a model and the like in the prior art, the invention provides the rapid construction method of the digital model of the large-scene road infrastructure, which can effectively improve the modeling speed and can directly check the facility types in model display.

In order to realize the technical effects, the invention provides the following technical scheme: comprises the steps of (a) preparing a mixture of a plurality of raw materials,

acquiring initial point cloud data of a road, wherein the road comprises a road section, a tunnel, a bridge and a jolt slope;

preprocessing the initial point cloud data through voxel filtering and an Euclidean clustering algorithm to obtain a plurality of point cloud clusters;

constructing a detection model, and respectively carrying out feature recognition on a plurality of point cloud clusters through the detection model;

and marking a plurality of point cloud clusters based on the identification result, and performing three-dimensional modeling on the marked point cloud clusters to obtain a large-scene road infrastructure digital model.

Optionally, the step of obtaining initial point cloud data of the road includes:

scanning the road section and the tunnel through a vehicle-mounted laser radar to obtain point cloud data of the road section and the tunnel;

scanning a bridge floor of a bridge through a vehicle-mounted laser radar to obtain bridge floor point cloud data, scanning a bridge bottom of the bridge through a fixed laser radar to obtain bridge bottom point cloud data, and integrating the bridge floor point cloud data and the bridge bottom point cloud data to obtain bridge point cloud data;

scanning a slope by carrying a laser radar on an unmanned aerial vehicle to obtain slope-turning point cloud data;

and integrating the road section and tunnel point cloud data, the bridge point cloud data and the slope point cloud data to obtain initial point cloud data of the road.

Optionally, the step of preprocessing by the voxel filtering includes:

the method comprises the steps of constructing a three-dimensional voxel space based on initial point cloud data, constructing a three-dimensional voxel grid in the three-dimensional voxel space, putting the initial point cloud data into the corresponding three-dimensional voxel grid, obtaining a grid gravity center based on the three-dimensional voxel grid, and filtering the initial point cloud data based on the grid gravity center to obtain filtered point cloud data.

Optionally, the step of performing preprocessing by the euclidean clustering algorithm includes:

converting the filtered point cloud data into a kd-tree data structure to obtain a point cloud data set;

selecting any point cloud data in the point cloud data set as an initial fixed point, performing Euclidean distance judgment and smooth parameter judgment on the point cloud data except the initial fixed point in the point cloud data set based on the initial fixed point, acquiring point cloud clusters based on a judgment result, repeating the steps based on the point cloud data except the point cloud clusters in the point cloud data set, and obtaining a plurality of point cloud clusters when the initial fixed point cannot be selected.

Optionally, in the process of constructing the detection model, the detection model includes a rotation alignment network, a global pooling layer, a first convolution layer, a second convolution layer, a first graph convolution layer, a second graph convolution layer, and a third graph convolution layer.

Optionally, the step of respectively performing feature recognition on a plurality of point cloud clusters through a detection model includes,

respectively inputting the point cloud clusters into the detection model, aligning the point cloud clusters through the rotary alignment network, processing the aligned point cloud clusters through the first graph convolution layer to obtain first structure information, processing the aligned point cloud clusters through the first graph convolution layer to obtain first characteristic information, fusing the first structure information and the first characteristic information, processing a fusion result through the second graph convolution layer to obtain second structure information, processing the first characteristic information through the second convolution layer to obtain second characteristic information, fusing the second structure information and the second characteristic information again, inputting the re-fusion result into the third graph convolution layer to obtain a characteristic matrix, and aggregating the characteristic matrix through the global pooling layer to obtain an identification result, wherein, the identification result is the types of different infrastructures in the road.

Optionally, before the feature recognition of the point cloud clusters by the detection model, the method further comprises,

acquiring a point cloud data training set, wherein the point cloud data training set comprises all target features in the initial point cloud data;

and training the detection model through the point cloud data training set.

Optionally, the labeled point cloud cluster is subjected to three-dimensional modeling through BIM software, and in the three-dimensional modeling process, different infrastructure in the modeled model is labeled based on a labeling result.

The invention has the following technical effects:

the method comprises the steps of acquiring initial point cloud data, and preprocessing the point cloud data to acquire a point cloud cluster; then respectively carrying out feature identification on different point cloud clusters through a detection model; and marking the point cloud clusters through the identified characteristics, and performing three-dimensional modeling on the marked point cloud clusters to obtain a digital model of the infrastructure of the large-scene road. According to the technical scheme, the digital model of the infrastructure of the large-scene road can be effectively and quickly modeled, different infrastructure types can be accurately identified in the established model, and the method has high practicability. Meanwhile, the invention uses voxel filtering, can effectively reduce the number of point cloud data, simultaneously retains the characteristics of the point cloud data, further accelerates the processing speed of the point cloud data, simultaneously uses a detection model to effectively detect the characteristic information of the point cloud cluster, accurately identifies the category of the point cloud cluster, and provides a marking data basis for marking in the later modeling.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

Fig. 1 is a schematic flow chart of a method according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to solve the problems that the modeling speed is low, direct facility type checking cannot be performed on virtual infrastructures in a model and the like in the prior art, the invention provides the following scheme:

as shown in fig. 1, the invention provides a method for quickly constructing a digital model of a large-scene road infrastructure, which comprises the following steps:

s1, acquiring initial point cloud data of a road, wherein the road comprises a road section, a tunnel, a bridge and a jolt slope; because the road contains different road structures, the point cloud data needs to be acquired by different acquisition modes aiming at different road structures, and the specific steps of acquiring the initial point cloud data of the road comprise:

s11, scanning the road section and the tunnel through a vehicle-mounted laser radar to obtain point cloud data of the road section and the tunnel; the highway section and the tunnel are comparatively gentle, can directly use on-vehicle laser radar to scan.

S12, scanning the bridge deck of the bridge through a vehicle-mounted laser radar to obtain bridge deck point cloud data, scanning the bridge bottom of the bridge through a fixed laser radar to obtain bridge bottom point cloud data, and integrating the bridge deck point cloud data and the bridge bottom point cloud data to obtain bridge point cloud data; because the bridge floor comparatively gently can directly use on-vehicle laser radar to scan, but the bridge bottom department can't scan because the restriction of position, and then need use fixed laser radar to scan to according to the actual position at bridge floor and bridge bottom, integrate above-mentioned point cloud data.

S13, scanning the slope through a laser radar carried by an unmanned aerial vehicle to obtain slope-climbing point cloud data; the pitch slope is comparatively jolted, and vehicle-mounted laser radar is at the pitch slope in-process of marcing, because the structural constraint on pitch slope, can't carry out effective scanning, so adopt unmanned aerial vehicle to carry out aerial scanning.

And S14, integrating the road section and tunnel point cloud data, the bridge point cloud data and the slope point cloud data to obtain initial point cloud data of the road. And integrating the data through the actual position to obtain initial point cloud data on the whole road.

The step of integrating the point cloud data comprises the following steps:

constructing a three-dimensional initial coordinate system, wherein the initial coordinate system can be constructed in a computer by combining a plurality of base stations based on a GPS;

scanning a road section and a tunnel through a vehicle-mounted laser radar, constructing a first measuring point coordinate system by taking the three-dimensional position of the vehicle-mounted laser radar as an origin, and carrying out position coordinate marking on point cloud data acquired in the scanning process of the vehicle-mounted laser radar in the first measuring point coordinate system in a polar coordinate mode under a coordinate axis;

marking the position of the vehicle-mounted laser radar when the vehicle-mounted laser radar needs to start scanning in an initial coordinate system, and mapping point cloud data in a first measuring point coordinate system of each frame in the scanning process to the initial coordinate system based on the position of the vehicle-mounted laser radar in the initial coordinate system in the moving and scanning processes of the vehicle-mounted laser radar;

constructing a first measuring point coordinate system for the bridge deck of the bridge through the vehicle-mounted laser radar, and mapping point cloud data of the bridge deck to the initial coordinate system;

scanning the bridge bottom of the bridge through a fixed laser radar, constructing a second measuring point coordinate system by taking the fixed laser radar as an origin, and carrying out position marking on point cloud data acquired by the fixed laser radar in the second measuring point coordinate system; mapping the point cloud data to an initial coordinate system;

when the laser radar carried by the unmanned aerial vehicle is used for scanning during slope changing, the position of the laser radar carried by the unmanned aerial vehicle is used as a third measuring point coordinate system, and point cloud data acquired by the laser radar carried by the unmanned aerial vehicle is subjected to coordinate marking under the third measuring point coordinate system; mapping the point cloud data to an initial coordinate system;

the vehicle-mounted laser radar and the laser radar carried by the unmanned aerial vehicle measure the linear acceleration and the inclination angle of the laser radar in each direction of an XYZ axis under a sensor coordinate in real time by carrying a high-precision inertial sensor, and the position of the laser radar is obtained by resolving through carrying out an inertial technology, and the specific steps comprise:

determining the initial position of a laser radar in an initial coordinate system, setting the initial speed to be zero, then measuring and collecting linear accelerations in different directions under sensor coordinates in real time when the laser radar moves, measuring the inclination angle of the laser radar in real time, correspondingly rotating the sensor coordinates through the inclination angle, wherein the rotated acceleration direction is the acceleration direction under the initial coordinate system, mapping the rotated acceleration direction from the sensor coordinates to the initial coordinate system, calculating acceleration data projected to the initial coordinate system through a quadratic integration method, drawing paths of the laser radar carried by the vehicle-mounted laser radar and the unmanned aerial vehicle at different time positions under the initial coordinate system in real time, corresponding position information at the same time with the collected point cloud data, and mapping the point cloud data to the initial coordinate system, the vehicle-mounted laser radar positioning device acquires the position information by loading the acceleration information acquired by the inertial sensor in real time at the laser radar of the vehicle, does not need to position transmission signals of external equipment outside the vehicle-mounted device, avoids the problem that a base station cannot accurately position the vehicle-mounted laser radar moving at a high speed, and ensures that the laser radar is accurately positioned under the high-speed movement.

And integrating and exporting all point cloud data in the initial coordinate system to obtain initial point cloud data.

The method comprises the steps of collecting point cloud data through mechanical scanning laser radar through scanning frequency rotation in the process of carrying out mobile collection on the point cloud data, determining the rotation speed of the radar in the laser radar rotation scanning process, determining the forward included angle between the laser radar and a vehicle according to the running time of the vehicle, integrating and exporting all point cloud data under an initial coordinate system according to the included angle of the running time at the current moment, carrying out rotation operation on the point cloud data under the initial coordinate system by taking the position of the vehicle in the initial coordinate system as the origin, carrying out rotation processing on all the point cloud data, integrating and exporting all the point cloud data under the initial coordinate system, and simultaneously selecting a high-precision laser radar by the vehicle-mounted laser radar to improve the precision of the point cloud data collection through the operation.

S2, preprocessing the initial point cloud data through voxel filtering and Euclidean clustering algorithm to obtain a plurality of point cloud clusters;

s21, the specific steps of processing by voxel filtering include,

and constructing a three-dimensional voxel space based on the initial point cloud data, constructing a three-dimensional voxel grid in the three-dimensional voxel space, putting the initial point cloud data into the corresponding three-dimensional voxel grid, and filtering the initial point cloud data based on the gravity center of the three-dimensional voxel grid to obtain filtered point cloud data.

The three-dimensional voxel space comprises a three-dimensional coordinate system, and when a three-dimensional voxel grid is constructed, the side length and the volume of the grid need to be selected according to the density of point cloud data, wherein the grid is a small cube. Placing the coordinates of the point cloud data into a coordinate system of a three-dimensional voxel space, ensuring that all the point cloud data are completely divided into all grids, calculating the side length of the grids, calculating the gravity center of the grids according to the side length of the grids, calculating the distance of different data points in the grids according to the gravity center of the grids, removing the point cloud data which do not accord with the threshold value by setting the threshold value, retaining the residual point cloud data, and obtaining the filtered point cloud data

S22, the specific steps of processing through the Euclidean clustering algorithm comprise:

converting the filtered point cloud data into a kd-tree data structure to obtain a point cloud data set; selecting any point cloud data in the point cloud data set as an initial fixed point, carrying out Euclidean distance judgment and smoothness judgment on the point cloud data except the initial fixed point in the point cloud data set based on the initial fixed point, obtaining a point cloud cluster based on a judgment result, repeating the steps on the data except the point cloud cluster in the point cloud data set, and obtaining a plurality of point cloud clusters when the initial fixed point cannot be selected.

Converting the filtered point cloud data into a kd-tree data structure to realize fast search based on neighborhood relations and facilitate subsequent fast calculation, selecting any point cloud data in a point cloud data set as an initial fixed point, selecting nearest neighbor point cloud data by setting an Euclidean distance threshold for the initial fixed point after the initial fixed point is selected, performing smooth parameter calculation on the nearest neighbor point cloud data, setting a smooth threshold after the calculation, removing the point cloud data exceeding the smooth threshold in the nearest neighbor point cloud data, reserving the remaining nearest neighbor point cloud data, integrating the remaining nearest neighbor point cloud data and the initial fixed point to obtain a point cloud cluster, continuously selecting the next fixed point for the filtered point cloud data, repeatedly selecting the nearest neighbor data and the Euclidean distance, and judging the smooth parameter, and obtaining a plurality of point cloud clusters until no redundant point cloud data is used as fixed points.

S3, constructing a detection model, and respectively carrying out feature recognition on a plurality of point cloud clusters through the detection model;

s31, constructing a detection model, wherein the detection model comprises a multilayer sensor, a global pooling layer, a first convolution layer, a second convolution layer, a first graph convolution layer, a second graph convolution layer and a third graph convolution layer. The rotary alignment network is respectively connected with the first convolution layer and the first graph convolution layer, and the first convolution layer is connected with the second convolution layer; the first graph convolution layer is connected with the second graph convolution layer, the second graph convolution layer and the second graph convolution layer are respectively connected with a third graph convolution layer, and the third graph convolution layer is connected with a global pooling layer, wherein the global pooling layer is set to be three layers in order to accurately output characteristics.

S32, respectively carrying out feature recognition on a plurality of point cloud clusters through the detection model, wherein the specific steps comprise respectively inputting the point cloud clusters into the detection model, aligning the point cloud clusters through a rotation alignment network, wherein the rotation alignment network comprises a T-Net and a matrix multiplication algorithm module which are used for carrying out normalized processing on the point cloud data, and the invariance of the geometric structure of the T-Net three-dimensional point cloud data is still kept when the T-Net three-dimensional point cloud data is subjected to matrix transformation. Meanwhile, a matrix multiplication algorithm is arranged in the matrix multiplication algorithm module, the T-net and the matrix multiplication algorithm module are connected in parallel to form a rotation alignment network, and the point cloud cluster outputs an affine transformation matrix after being input into the rotation alignment network.

In the detection model, processing the aligned point cloud cluster by a first graph volume layer to obtain first structure information, wherein the structure information is local structure information, processing the aligned point cloud cluster by the first volume layer to obtain first characteristic information, wherein the characteristic information is the characteristic information of an individual point, fusing the first structure information and the first characteristic information to ensure that the obtained characteristics simultaneously comprise the local structure information and the characteristic information of the individual point, processing the fusion result by a second graph volume layer to obtain second structure information, processing the first characteristic information by the volume layer to obtain second characteristic information, re-fusing the second structure information and the second characteristic information, inputting the re-fusion result into a third graph volume layer to obtain a characteristic matrix, wherein the characteristic matrix comprises the characteristic information of each point and the local structure information, and performing global characteristic description output by a global pool, the use of the global pooling layer can effectively reduce the complexity of model calculation and can more effectively improve the calculation speed, and the global pooling layer finally generates a one-dimensional vector to represent the corresponding facility category of the point cloud cluster, namely the identification result.

Before the feature recognition of the preprocessed point cloud data is carried out through the detection model, a point cloud data training set is obtained, wherein the point cloud data training set comprises all target features in the point cloud data; after the point cloud data training set can be scanned through the vehicle-mounted laser radar, the point cloud data training set is constructed by adding labels in a manual labeling mode. And training the detection model through a point cloud data training set. The point cloud data training set can effectively improve the identification accuracy of the detection model, and meanwhile, the acquired point cloud data training set can be divided into training samples and testing samples through ten-fold intersection, the testing samples are used for testing the detection model, and when a testing result reaches a certain precision, the testing samples are used again.

And S4, marking the point cloud cluster based on the identification result, and performing three-dimensional modeling on the marked point cloud cluster through BIM software to obtain a digital model of the large-scene road infrastructure.

Marking the point cloud cluster, marking different types, acquiring the point cloud cluster and corresponding labels, integrating the point cloud cluster and the corresponding labels, distinguishing different infrastructures in the model by using different colors in the process of modeling by using BIM software, and simultaneously adding the labels, so that the color distinction and the infrastructure in the digital model of the labels can be displayed more visually.

The foregoing illustrates and describes the principles, general features, and advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.