CN110175574A - A kind of Road network extraction method and device - Google Patents

Abstract

The invention relates to a road network extraction method and device, belonging to the technical field of geographic data processing. The present invention first extracts the image node position and judges its spatial attribute, and obtains the road intersection node position and type; then performs image segmentation on the obtained intersection node, uses auxiliary data to construct a road search buffer, and extracts the direction texture in the buffer feature, determine the road direction and the number of branches passing through the intersection node according to the direction texture feature; then determine the center line of each road on the intersection node according to the road branch direction and the road branch number on each intersection node; then according to each intersection node The spatial position and adjacency relationship between intersection nodes determine the topology of the road network. The invention has a high degree of automation, can improve the extraction efficiency of the road network, and simultaneously improves the accuracy of the road network extraction through a small amount of artificial parameters.

Description

A road network extraction method and device

technical field

The invention relates to a road network extraction method and device, belonging to the technical field of geographical data processing.

Background technique

According to whether the road extraction algorithm needs human-computer interaction, the road extraction algorithm can be divided into semi-automatic road extraction and automatic road extraction. The semi-automatic road extraction method is currently more effective. First, the operator manually inputs initial information such as seed points or road templates, and then the computer uses the initial road information to complete road identification and positioning. The more commonly used semi-automatic road extraction methods include active contour models. method and template matching method. The fully automatic road extraction algorithm does not require any prior information and manual assistance in the road extraction stage, and independently completes road identification and positioning. Fully automatic road extraction is the ultimate goal of researching road extraction, but from the current research progress, The existing automatic road extraction algorithms are not robust, and the extraction results still need a lot of manual post-processing to meet the actual production needs, and there is still a long way to go to achieve this goal.

Due to the obvious features and rich details of high-resolution remote sensing images, the available information and applicable theoretical techniques are diverse. Mean and Poullis comprehensively analyzed and summarized the existing remote sensing image road extraction algorithms, and divided road extraction methods into three categories according to the level of information utilization in road extraction algorithms: pixel-based road extraction methods, and region-based road extraction methods and knowledge-based road extraction methods. The pixel-based road extraction method mainly processes and analyzes the pixel-level information to infer possible road information; the region-based road extraction method is currently the most commonly used high-resolution remote sensing image road network extraction method, first through image segmentation Algorithms or classification algorithms divide remote sensing images into different regions, and then extract road networks according to certain rules; knowledge-based road extraction algorithms need to integrate multiple data, and use multiple information to mine, summarize and analyze knowledge at a higher level. expression, and finally realize the recognition and extraction of road features.

Although the above-mentioned solutions can realize road extraction and recognition, they have the problems of single application image features and low intelligence.

Contents of the invention

The purpose of the present invention is to provide a road network extraction method to solve the problems of single application image features and low intelligence in the current road network extraction process; at the same time, the present invention also provides a road network extraction device to solve At present, there are problems in the process of road network extraction, such as single application of image features and low degree of intelligence.

The present invention provides a kind of road network extraction method for solving above-mentioned technical problem, and this extraction method comprises the following steps:

1) Obtain remote sensing images containing road features, and extract the location and type of road intersection nodes from the acquired remote sensing images;

2) Carry out image segmentation to the acquired intersection node, utilize auxiliary data to construct road search buffer zone, and extract the direction texture feature in the buffer zone, determine the road branch direction and road branch number of intersection node according to described direction texture feature ;

3) Determine the road centerline of the intersection node along the direction of each road branch according to the road branch direction and the number of road branches on each intersection node;

4) Judging the positional relationship between each intersection node and the centerline of each road, and determining the topological structure of the road network.

The present invention also provides a road network extraction device, the extraction device includes a memory and a processor, and a computer program stored in the memory and run on the processor, the processor is coupled to the memory , when the processor executes the computer program, the above road network extraction method is realized.

The present invention can be applied to the road information collection operation of high-resolution remote sensing images, changing the existing geographic information guarantee mode, completing the road information collection tasks with a certain width on high-resolution remote sensing images through a small amount of manual participation, and significantly shortening the mapping cycle , greatly improving the automatic processing of remote sensing images.

Further, in order to improve the accuracy of intersection node position and type identification, the extraction process of road intersection node position and type in described step 1) is:

A. Use the trained deep convolutional neural network to extract the features of the acquired remote sensing images to obtain the intersection node features;

B. Input the obtained features into the region generation network RPN to obtain the feature information of the candidate frame;

C. Classify the feature information of the candidate frame to determine the location and type of the intersection node.

Further, in order to avoid the interference caused by occlusion on the road and improve the accuracy of road centerline extraction, the step 3) adopts the road center point matching algorithm of mean shift to extract the road centerline.

Further, the present invention provides a specific road center point matching algorithm. The road center point matching algorithm based on mean shift includes extended Kalman filter and road center point matching based on Mean Shift to extract the road center line.

Further, the determination process of road network topology in said step 4) is as follows:

a. Take any road intersection node as the initial road intersection, search along any road branch direction of the initial road intersection and the corresponding road centerline until reaching the next road intersection node or area boundary, and complete the road section Repeat the process until all road branches on the road intersection node are searched;

b. Search for the next road intersection node in the direction of each road branch of the initial road intersection according to the method of step a, until all intersection nodes are traversed.

Description of drawings

Fig. 1 is the flowchart of road network extraction method of the present invention;

Fig. 2 is the flow chart of the overpass model training based on deep learning in the embodiment of the road network extraction method of the present invention;

Fig. 3 is a model diagram of a deep convolutional neural network in an embodiment of the road network extraction method of the present invention;

Fig. 4 is a flow chart of rapid target location and identification of the region generation network in the embodiment of the road network extraction method of the present invention;

Fig. 5 is the flow chart of the spatial attribute identification of road intersection nodes in the embodiment of the road network extraction method of the present invention;

Fig. 6-a is a direction texture rectangular diagram in an embodiment of the road network extraction method of the present invention;

Fig. 6-b is a distribution diagram of texture feature values in various directions in the embodiment of the road network extraction method of the present invention;

Fig. 7 is a flow chart of road centerline extraction in an embodiment of the road network extraction method of the present invention;

Fig. 8 is a road network flow chart in an embodiment of the road network extraction method of the present invention;

Fig. 9 is a flow chart of the device software architecture processing in the embodiment of the road network extraction device of the present invention.

Detailed ways

The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

method embodiment

The overall flow of the extraction method of the present invention is to first extract the position of the image node and judge its spatial attributes to obtain the position and type of the road intersection node; The direction texture feature in the buffer zone, determine the road direction and the number of branches passing through the intersection node according to the direction texture feature; then determine the center line of each road on the intersection node according to the road direction and the number of branches on each intersection node ; Then according to the spatial position and adjacent relationship between each intersection node, determine the topological structure of the road network; finally, the road network is trimmed by the post-processing method to obtain the final extraction result. The key technologies of this extraction method are road backbone node detection, road segment search and road network topology construction. The specific implementation process of this method is shown in Figure 1, and the specific process is as follows:

1. Obtain the location and type of road intersection nodes.

As a kind of deep learning model, deep convolutional neural network can automatically learn and extract features from data, and its generalization ability is significantly better than traditional methods. Deep convolutional neural network is a multi-layer supervised learning network with input Layer, hidden layer (including convolutional layer and downsampling layer) and output layer, the network structure is optimized through the error backpropagation algorithm, unknown parameters are solved, and finally a target model library containing the number of sample categories is formed, as shown in Figure 2. Taking the identification of overpasses as an example, first establish a database containing various overpass targets, use some samples in the database as training data sets, and another part of samples as verification data sets to verify and correct the trained deep convolutional neural network. First, use The training data set trains the deep convolutional neural network, obtains the corresponding training data set features, and inputs the training data set features into the classifier to train the classifier. The classifier can use SVM, MLP, etc.; then use the verification data Set the trained deep convolutional neural network to obtain the verification data set features, and input the obtained verification data set features into the trained classifier for verification. Through the above process, the overpass model library can be obtained, and the acquired remote sensing image can be input into the overpass model library to determine whether the remote sensing image contains the overpass and the characteristics of the overpass.

For the invention, the purpose is to identify road intersection nodes, and road intersection nodes, especially overpasses, as complex objects, require training with multiple convolutional neural network layers for recognition using deep convolutional neural networks. structure to improve the robustness of object detection. As shown in FIG. 3 , the deep convolutional neural network in this embodiment includes a network structure of 13 convolutional layers. After the convolutional features are obtained through the deep convolutional neural network, it needs to be classified.

In order to improve the efficiency of classification, it is necessary to provide a certain candidate area before classification. In this embodiment, the convolution feature is input to the RPN during the detection process to obtain the feature information of the candidate frame. For the features extracted from the candidate frame, use The classifier determines whether it belongs to a specific class, and for the candidate frame belonging to a certain feature, the regressor is used to further adjust its position to achieve rapid detection and positioning of the target. The whole processing process includes feature extraction of the image to be tested, generation of candidate windows by region generation network, feature classification, position refinement and other steps to complete the task of rapid positioning and recognition of the target. The technical process is shown in Figure 4.

2. Obtain the attribute information of the intersection node.

The attributes of an intersection node generally include the number of road branches on the intersection node and the direction of the road branch. Since the intersection node on the remote sensing image is a relatively complex geographical target, it is difficult to identify the attributes of the intersection node with only a single method. , especially for intersection nodes such as overpasses, their spatial attributes are more diverse and complex, and identification is more difficult. In order to completely and accurately obtain the attribute information of intersection nodes, especially the special intersection nodes such as overpasses, the present invention adopts a spatial attribute analysis and judgment method based on multivariate data. Taking an overpass as an example, the basic idea of recognition is to comprehensively utilize the gradient on the image. Features, texture features, spatial layout features, and other source data are analyzed and discriminated in a unified manner. The overpass is analyzed and processed in depth by formulating corresponding discrimination criteria, and the calculation of the number of road branches and the direction of the road branches of the overpass is realized. The road backbone network extracts and provides node information. The process is shown in Figure 5, and the process is as follows:

Segment the image of the overpass target determined in step 1, and then use the morphological method to filter out the interference factors in the scene, then construct the road search buffer according to the external auxiliary data, and extract the directional texture features in the buffer, and compare the different directions of the road The texture feature value on the road is used to judge the direction of the road and the number of branches (it can be seen from the road feature that the texture feature is stable along the road direction and the change is minimal). The auxiliary data includes GIS data, spatial information data, etc., and the search buffer refers to extending a certain distance along the two sides of the road as the search range, with the purpose of improving calculation efficiency.

Angular Texture Signature (Angular Texture Signature) refers to the texture feature value of a rectangular area with direction to represent the change of the gray level and texture of the area. For a point p on the image, define a function T(a,w,p), which represents the texture feature value (entropy, energy, mean, variance, etc.) ). Starting from the horizontal direction, rotate the angle at intervals of α to form a set of rectangular templates, and obtain a set of texture feature values {T(a0,w,p),T(a1,w,p),...,T(an,w ,p)}, use these values to describe the direction texture characteristics of the point. Taking the variance in the directional texture feature as an example to illustrate it is shown in Figure 6-a and Figure 6-b.

Figure 6-a is based on the center point p pixel as the axis rotation, starting from 0 degrees in the horizontal direction, and building rectangles at intervals of 20 degrees to form 18 rectangular frames. Figure 6-b is the variance calculated based on 18 rectangular templates Generated chart. From the chart drawn in Figure 6-b, it can be seen that the extreme points of the variance of the rectangular area around the road center point p appear in the two directions of 2 and 11, indicating that the gray scale changes in these two directions are the smallest. It can also be seen from Figure 6-a that these two directions are exactly the direction of the road, so it is theoretically reasonable to use directional texture features for road extraction.

3. Identify roads between adjacent intersection nodes.

The traditional road recognition algorithm based on template matching mainly uses the correlation coefficient as the similarity measure, which is very sensitive to the gross error of the observation value and has poor robustness. Suitable for high-resolution remote sensing images. For this reason, the present invention adopts the road center point matching algorithm based on mean value drift, and this algorithm algorithm adopts a kind of robust similarity measurement, solves the optimal matching point of road center point by mean value drift algorithm, has overcome traditional template matching to road surface The shortcomings of occlusion sensitivity; then use the Kalman filter to extract the road on the high-resolution remote sensing image by combining the road prior information and the current (matched) observation information. The matching process is shown in Figure 7.

Kalman filter state parameters are composed of road direction φ _k and road center coordinates (x _k , y _k ), in the state prediction process, the state prediction vector at k+dt Estimate the vector from the state at k Obtained by the state equation, the state equation is:

It can be seen from formula (2) that the state prediction process is not a linear process, and the extended Kalman filter needs to be used. The covariance matrix of the state prediction vector used by the extended Kalman filter is:

Among them, Φ _k represents the coefficient matrix after linearization of the state equation, Q _k represents the covariance matrix of the system noise, and P _k-1 is the covariance matrix of the state vector. After state prediction, the road center point matching algorithm is used to obtain observations, namely:

In the formula, φ _k represents the road direction at the coordinates (x _k , y _k ), which is calculated from the observed value of the road center coordinates and the estimated value of the road center coordinates of the previous state. The observation equation is:

Where A is a unit vector, then the state estimation vector at k is:

K _k is the gain of the Kalman filter, the system error covariance matrix Q and the observation error covariance matrix R play an important role in the Kalman filter. In the road extraction method of the present invention, the systematic error is mainly related to the curvature of the road to be extracted, so the systematic error covariance matrix Q should depend on the actual curvature of the road. For the observation value covariance matrix R, the following strategy is adopted: after matching a road center point, the algorithm will calculate the similarity between the matching point and the road template, and then determine R in real time according to the similarity. At the same time, the road centerline must be a smooth curve, so when the difference between the road direction in the observed value and the road direction in the previous state exceeds the threshold, then increase R.

Finally, the road center point is iteratively tracked according to equations (2)-(5) to realize the automatic matching search of road segments in high-resolution remote sensing images (this process has been described in the paper "Road Center Line of Remote Sensing Image Combined with Mean Shift and Kalman Filtering Tracking Algorithm" (Journal of Surveying and Mapping, February 2016) published in detail).

4. Road network.

The road network networking is based on the location and type of image nodes. The center of the homogeneous area of the road intersection can correspond to the vertices in the road network structure, and the road fitting curve connecting adjacent intersections corresponds to the road network lines in the structure. The road network topology construction process is divided into two steps: road network networking and road network regulation. The purpose of road network networking is to judge the spatial position and adjacent relationship between intersections, and to form a preliminary image road network topology. When networking, select an intersection as the initial position, use the road direction search algorithm to determine the exact direction of the road branch of the intersection, match and search road nodes along different road branches, and stop the current search at the next intersection and store the current vertex and Road section information, and then continue to search from the new intersection position until all intersections have been traversed; road network regulation is a post-processing process of the road network topology, and is a sorting out of road section nodes in the road network. Due to the complexity of the scene on the image, the interference factors are relatively large. Many, the road segment nodes obtained by using the matching search algorithm may extend to the surrounding scenes, making the nodes deviate from the road centerline. The purpose of the road network regulation is to retain the nodes on the road, and perform curve fitting on these nodes, and finally Vertices and road section information are stored in the corresponding table structure to form a complete road network topology. The process of road network topology construction is shown in Figure 8.

Among them, the road branch direction is obtained by calculating the directional texture eigenvalue of the road. The intersection branch direction detected by the previous algorithm is used as the initial direction, and the directional texture rectangular eigenvalue is calculated within a certain range of the initial direction neighborhood, and the texture eigenvalue appears extreme. The direction of the value is the direction of the road. Take any road intersection node as the initial road intersection, search along any road branch direction of the initial road intersection and the corresponding road centerline until reaching the next road intersection node or area boundary, and complete the construction of the road section , repeat this process until all road branches on the road intersection are searched; search for the next road intersection node in the direction of each road branch of the initial road intersection according to the above method, and respectively compare the vertices of each road section with the road section The information is saved in the topology table, and the processing continues from the next intersection until all intersections are traversed to form a preliminary road network topology.

Device embodiment

The road network extraction device of the present invention includes a memory and a processor, and a computer program stored in the memory and run on the processor, the processor is coupled with the memory, and the road network in the above method embodiment is realized when the processor executes the computer program Extraction Method. The software architecture processing flow of the device is shown in FIG. 9 . Processor can adopt single-chip microcomputer, DSP, PLC or MCU etc., memory can adopt RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, register, hard disk, mobile disk, CD-ROM or any other form known in the art The storage medium and the specific implementation means of each instruction have been described in the embodiments of the method, and will not be repeated here.

Claims (6)

1. A road network extraction method is characterized in that the extraction method comprises the following steps: 1) Obtain remote sensing images containing road features, and extract the location and type of road intersection nodes from the acquired remote sensing images; 2) Carry out image segmentation to the acquired intersection node, utilize auxiliary data to construct road search buffer zone, and extract the direction texture feature in the buffer zone, determine the road branch direction and road branch number of intersection node according to described direction texture feature ; 3) Determine the road centerline of the intersection node along the direction of each road branch according to the road branch direction and the number of road branches on each intersection node; 4) Judging the positional relationship between each intersection node and the centerline of each road, and determining the topological structure of the road network. 2. road network extraction method according to claim 1, is characterized in that, described step 1) in the extraction process of road intersection node position and type: A. Use the trained deep convolutional neural network to extract the features of the acquired remote sensing images to obtain the intersection node features; B. Input the obtained features into the region generation network RPN to obtain the feature information of the candidate frame; C. Classify the feature information of the candidate frame to determine the location and type of the intersection node. 3. The road network extraction method according to claim 1 or 2, characterized in that, said step 3) adopts the road center point matching algorithm of mean shift to extract the road centerline. 4. The road network extraction method according to claim 3, characterized in that, the road center point matching algorithm of the mean shift includes extended Kalman filtering and road center point matching based on Mean Shift to extract the road center line. 5. road network extraction method according to claim 1, is characterized in that, described step 4) in the determination process of road network topology structure is as follows: a. Take any road intersection node as the initial road intersection, search along any road branch direction of the initial road intersection and the corresponding road centerline until reaching the next road intersection node or area boundary, and complete the road section Repeat the process until all road branches on the road intersection node are searched; b. Search for the next road intersection node in the direction of each road branch of the initial road intersection according to the method of step a, until all intersection nodes are traversed. 6. A road network extracting device, characterized in that the extracting device comprises a memory and a processor, and a computer program stored on the memory and operated on the processor, the processor is connected to the memory coupled, the processor implements the road network extraction method according to any one of claims 1-5 when executing the computer program.