Automatic identification method of road in remote sensing image

Abstract

The application provides an automatic identification method for roads in remote sensing images, comprising the following steps: S1, pre-processing the remote sensing image; S2, obtaining a linear road feature map using a pre-trained road identification model; S3, performing repair, vectorization and regularization processing on the linear road feature map; S4, identifying an abnormal crossroad skeleton line according to a preset determination condition; S5, constructing a coordinate system based on the abnormal crossroad skeleton line, performing neighborhood analysis, and obtaining effective feature information after filtering; S6, taking the remote sensing image after encoding processing and label information as a sample set; S7, inputting the sample set into the pre-trained road identification model to obtain an optimized model; and S8, using the optimized model to perform automatic road identification on a new task remote sensing image. The optimized model improves the accuracy of road identification and can realize automatic identification of abnormal crossroads, reduces manual operation, and provides convenience for subsequent road post-processing.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image processing, and particularly relates to a method for automatically identifying a road in a remote sensing image. BACKGROUND

[0002] It is an economic and efficient means to extract road information using high-resolution remote sensing images, which can not only verify the extraction results very intuitively and conveniently through visual comparison, but also has a significantly lower cost than the road information acquisition method based on field measurement data or mobile trajectory data. However, the spectral information of ground objects in remote sensing images is rich and complex, and the phenomena of "same object with different spectra" and "different objects with similar spectra" are common. The spectral characteristics of the same type of ground object differ greatly, and the spectral characteristics of different types of ground objects are similar, and the spectral separability of the image is reduced. At present, there are still many difficulties and challenges in extracting road information from high-resolution remote sensing images.

[0003] Meanwhile, the intersection is an important part of the road network and is located at the junction of roads. Accurate intersection information is of great significance for image registration, intelligent transportation, GIS data updating and road change detection. Generally, urban roads are relatively straight, so the general structure of the urban road network can be determined by using intersections, such as connecting adjacent intersections to obtain the general layout of the road network in the area. Since the building materials of the road are similar to those of other surrounding ground objects, they have similar spectral characteristics. On the other hand, intersections are easily disturbed by surrounding buildings, parking lots, shadows, vegetation, vehicles, etc.

[0004] The tracking extraction algorithm of the road will be interrupted when encountering a crossroad, or will blindly select a branch to continue tracking, and after turning, it may circle a rectangular block and return to the previous road segment for repeated extraction. The tracking presents great uncontrollability when encountering a crossroad, resulting in abnormal extraction of the crossroad.

[0005] From the current research status, the semi-automatic method with artificial intervention to quickly and accurately extract the road is the most common choice. Through artificial intervention, the geometric features of the target object graph are analyzed and interpreted according to experience, and the image and processing algorithm that need to be processed are selected. In addition, through artificial intervention, the abnormal crossroads in the road are identified, which increases the time and cost of identifying abnormal crossroads. Moreover, the identification rules of artificial intervention are mostly based on simple and low-latitude features, which often lack completeness and strong generalization ability, and have low identification efficiency and poor accuracy. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an automatic identification method for roads in remote sensing images, which can automatically identify roads and abnormal crossroads, uses machine learning instead of manual intervention, obtains sufficient training samples to train a model, automatically obtains generalized features for identification of abnormal crossroads, reduces manual operation, improves the efficiency and accuracy of abnormal crossroad identification, and provides convenience for subsequent image preprocessing.

[0007] To achieve the above technical purposes, the technical scheme of the present application is as follows:

[0008] S1 obtains a remote sensing image and pre-processes it;

[0009] S2 uses a pre-trained road identification model to identify roads in the remote sensing image, obtaining a linear road feature map;

[0010] S3 repairs, vectorizes and regularizes the linear road feature map to obtain a regular planar road feature map and first feature information; the regular planar road feature map includes a skeleton line, and the skeleton line includes a crossroad skeleton line and a non-crossroad skeleton line; the crossroad skeleton line includes a crossroad skeleton line;

[0011] S4 determines the crossroad skeleton line according to a preset determination condition based on the first feature information, obtains label information of the crossroad skeleton line, and divides the crossroad skeleton line into an abnormal crossroad skeleton line and a normal crossroad skeleton line according to the label information;

[0012] S5 constructs a coordinate system based on the abnormal crossroad skeleton line, performs neighborhood analysis according to the coordinate system, filters out interference feature information of the abnormal crossroad skeleton line, and retains effective feature information;

[0013] S6 encodes the effective feature information, and uses the remote sensing image with the encoded effective feature information and label information as a sample set;

[0014] S7 inputs the sample set into the pre-trained road identification model, performs iterative training until a convergence condition is met, the model training is completed, and an optimized model is obtained;

[0015] S8 uses the optimized model to automatically identify roads in a new remote sensing image.

[0016] Specifically, step S4 includes:

[0017] S41 the preset determination condition is:

[0018] The first condition: the crossroad skeleton line has at least 3 lines;

[0019] The second condition is that two ends of one intersection skeleton line are both three intersection skeleton line coincidence points;

[0020] The third condition is that a ratio of a road line length and a width corresponding to the intersection skeleton line is less than 3;

[0021] The fourth condition is that a length of the intersection skeleton line is less than or equal to 100 m;

[0022] S42, judging the intersection skeleton line according to a preset judging condition;

[0023] If the intersection skeleton line meets the first condition, the second condition, the third condition and the fourth condition at the same time, a label information of the intersection skeleton line is N;

[0024] If the intersection skeleton line does not meet any one of the first condition, the second condition, the third condition and the fourth condition, the label information of the intersection skeleton line is Y;

[0025] S43, marking the intersection skeleton line with the label information N as an abnormal intersection skeleton line, and marking the intersection skeleton line with the label information Y as a normal intersection skeleton line.

[0026] Specifically, the step S5 includes:

[0027] S51, selecting two end points connected to form a line segment with three abnormal intersection skeleton line coincidence points as the end points;

[0028] S52, constructing a coordinate system with an extension direction of the line segment as an X axis and taking a midpoint of the line segment as a coordinate origin, and performing eight-neighborhood analysis to obtain second feature information;

[0029] S53, connecting the coincidence points of the abnormal intersection skeleton line and the non-intersection skeleton line as a demarcation point to form two regions;

[0030] S54, taking a region close to the coordinate origin side, i.e., the abnormal intersection skeleton line region, as an interference region, and taking a region far away from the coordinate point side, i.e., the non-intersection skeleton line region, as an effective region;

[0031] S55, filtering interference feature information of the interference region in the second feature information, and taking feature information of the demarcation point and the effective region as effective feature information.

[0032] Specifically, the second feature information in the step S4 includes a road length, a width and a direction, a skeleton line length and a direction, a distance of a skeleton line end point to a coordinate axis, and a quadrant where the skeleton line end point is located.

[0033] S31, performing smoothing processing on a contour boundary of a road, filling a hole of the road and repairing a road line gap to obtain a repaired linear road feature map;

[0034] S32 vectorize the repaired linear road feature map by using ArcGIS, and extract the skeleton line and the first feature information to obtain a planar road feature map;

[0035] S33 regularize the planar road feature map by using a simplification algorithm to obtain a regular planar road feature map.

[0036] Specifically, the road extraction model in step S2 is a U-Net model, and the model comprises a first convolution module, an encoding part, a decoding part, an intermediate bridging structure between the decoding parts, and a classifier connected in sequence.

[0037] Specifically, the encoding part comprises a four-level residual network module, and the residual network is a ResNet residual network.

[0038] Specifically, the preprocessing comprises grayscale processing and filtering processing.

[0039] Specifically, the simplification algorithm is Douglas-Poek.

[0040] The beneficial effects of the present application include:

[0041] (1) By constructing a coordinate system for the abnormal intersection skeleton line, performing eight-neighborhood analysis, obtaining effective feature information, screening out abnormal intersection sample sets, optimizing the road recognition model, realizing automatic recognition of abnormal intersections, reducing manual operation, and improving the efficiency and accuracy of abnormal intersection recognition;

[0042] (2) By using the optimized road recognition model, the road and abnormal intersection can be automatically recognized, and the characteristics of the road line, the width of the road line, the direction of the skeleton line and other characteristics can be obtained according to the model output, which provides convenience for the later road post-processing, realizes the rapid repair or connection of abnormal roads and intersections, and obtains a complete and accurate road network;

[0043] (3) The U-Net and ResNet model ideas are combined, the knowledge-driven U-Net model is used for road recognition, the network depth is deepened, the road recognition accuracy is improved, the multi-scale context information is fully utilized, and the loss of detail information is effectively avoided. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The method flowchart of the embodiment of the present application is shown in the figure;

[0045] Figure 2 The road recognition model network structure diagram of the embodiment of the present application is shown in the figure;

[0046] Figure 3A schematic diagram for constructing a coordinate system of an embodiment of the present application. Specific embodiments

[0047] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0048] The present application provides an automatic identification method for roads in remote sensing images, comprising the following steps: an automatic identification method for roads in remote sensing images, comprising the following steps:

[0049] Please refer to Figure 1 , S1 obtains a remote sensing image and pre-processes it;

[0050] Since the obtained road image often contains noise, it is necessary to pre-process the road image to eliminate the influence of noise on the road image. Specifically, the pre-processing includes grayscale processing and filtering processing, wherein a region of interest (ROI) in the remote sensing image is selected, and the region of interest is grayscale processed; wherein the region of interest contains road line information, and then the grayscale processed image is low-pass filtered. In addition, in order to eliminate the influence of light on the quality of the road image under different climate conditions. In an optional embodiment, the step of pre-processing the road image can also include adjusting the contrast of the low-pass filtered image. The specific grayscale processing and low-pass filtering processing mode is the same as the related prior art, and the embodiments of the present application will not be described again.

[0051] S2 uses a pre-trained road identification model to identify roads in the remote sensing image, and obtains a line-shaped road feature map;

[0052] The road identification model identifies the pre-processing result to form a line-shaped road feature map;

[0053] In the embodiments of the present application, a road identification model based on a U-Net network is used, please refer to Figure 2 , which includes a first convolution module X1, an encoding part C, a decoding part D, an intermediate bridging structure L0 between the decoding parts, a pyramid pooling part P, and a classifier E connected in sequence. The encoding and decoding parts are composed of BN (batch normalization), ReLU (activation function), and Conv (convolution operation).

[0054] The traditional U-Net network includes two parts, namely a feature extraction part and an up-sampling part. In the feature extraction part, there is one scale after each pooling layer, including the original image scale, a total of 5 scales. In the up-sampling part, each up-sampling is performed once, and the same scale corresponding to the feature extraction part is spliced.

[0055] The coding part C in the embodiment of the present application adopts a residual network to replace the feature extraction part in the traditional U-Net network, and each residual network module constitutes the main part of the improved U-Net downsampling. The coding part C includes four levels of residual networks X2-X5, and the residual network is a ResNet residual network. By using the residual network, the information loss in the coding process can be reduced, and the network depth can be increased. The decoding part D includes three levels of convolution modules Y1-Y3. The classifier E can be a random forest classifier. The pooling part P includes pyramid pooling modules Z1-Z3 corresponding to the convolution modules Y1-Y3 of each level constituting the decoding part D. The feature maps obtained by the convolution modules Y1-Y3 of each level are input into the pyramid pooling modules Z1-Z3, respectively. Furthermore, the feature maps output by the pyramid pooling modules are upsampled to the size of the original image and are superimposed with the input original image and then output to the classifier, and then the prediction result of each pixel is obtained through the classifier.

[0056] The improved U-Net network increases the diagonal connection structure on the basis of the traditional U-Net network, including the diagonal downward connection structure L1 and the diagonal upward connection structure L2, so that the network can fully utilize the context information of images of different scales. In addition to being connected to the residual network X2 through the bridging structure L0, the convolution module Y3 is also connected to the convolution module X1 through the downward connection structure L1 and to the residual network X3 through the upward connection structure L2. Similarly, in addition to being connected to the residual network X3 through the bridging structure L0, the convolution module Y2 is also connected to the convolution module X2 through the downward connection structure L1 and to the residual network X4 through the upward connection structure L2. In addition to being connected to the residual network X4 through the bridging structure L0, the convolution module Y1 is also connected to the convolution module X3 through the downward connection structure L1 and to the residual network X5 through the upward connection structure L2. The diagonal upward connection structure includes image upsampling and image fusion, and its purpose is to sample the image with a size smaller than the size of the corresponding scale image to the same size and then perform image fusion.

[0057] The remote sensing image containing real road labels is input into the network model as a training set for the network model to learn. The loss value is calculated through the loss function, and the prediction error loss value is fed back to the network for gradient feedback to correct the parameters of each module unit in the network. Through a predetermined number of loop calculations, the final prediction error is within the set threshold range, the prediction result accuracy is within the expected range, and a network model with accurate road recognition capability is obtained. Through the pre-trained road recognition model, the remote sensing image can be recognized to obtain the road recognition result.

[0058] Each residual network module used in the embodiment of the application is composed of two branches, the first branch is a main branch composed of BN, ReLU and Conv, and performs twice convolution; the second branch is a residual branch, and the two branches are superimposed at the output end. Assuming that the output is X1+1, the input is X1, and F is a function of the first branch, the expression is X1+1=F(X1)+X1. This mode of skip connection originating from ResNet can slow down the attenuation speed and improve the learning rate.

[0059] S3 repairs, vectorizes and regularizes the linear road feature map to obtain a regular planar road feature map and first feature information; the regular planar road feature map comprises a skeleton line, and the skeleton line comprises an intersection skeleton line and a non-intersection skeleton line; wherein the intersection skeleton line comprises a cross intersection skeleton line and a non-cross intersection skeleton line.

[0060] Further, step S3 comprises:

[0061] S31 smoothes the contour boundary of the road, fills the holes of the road and repairs the road line gap to obtain a repaired linear road feature map;

[0062] S32 vectorizes the repaired linear road feature map by using ArcGIS, extracts a skeleton line and first feature information, and obtains a planar road feature map;

[0063] The first feature information comprises road length, contour area, road width and road direction, and skeleton line length and skeleton line direction.

[0064] In the embodiment of the application, the repaired linear road feature map is converted into vector coordinate points by adding an X, Y data tool on the ArcGIS platform, the vector coordinate points are sequentially connected according to the collection serial number of the coordinate points by using a point set to line tool on the ArcGIS platform to obtain a road skeleton line; and the road skeleton line is subjected to buffer analysis by using a buffer analysis tool on the ArcGIS platform, the buffer radius is the road width, the obtained buffer area is extracted as a linear road element by using a feature to line tool on the ArcGIS platform, the linear road element is converted into a surface element, and finally a planar road feature map is obtained.

[0065] S33 utilizes the simplified algorithm to regularize the planar road feature map, and a regular planar road feature map is obtained. In the embodiment of the application, the road recognition result obtained in step S2 may cause the originally regular road to have sharp corners and missing parts. After the raster map is vectorized, the road in the vector layer is regularized from the perspective of the vector, mainly including: initial screening of the threshold for the fragment, preventing too many fragments from appearing non-road small map spots being regularized; regularization, since the road object is basically a regular rectangle and the prediction result is approximately rectangular, the road is found by using the matching algorithm and the key points are checked to directly simplify and regularize the road vector to achieve the geometric shape of regular appearance.

[0066] In the embodiment of the application, the Douglas-Peucker line simplification algorithm is improved to regularize the road recognition result. Since the algorithm needs to set a threshold to detect the key points of the road structure, it is more dependent on the setting of the manual threshold, and needs to weaken the influence of the threshold on the road shape. In this embodiment, the area of the vector layer after projection is calculated, a relatively small starting threshold is set, and then the actual area is used as the iteration end point. When the area of the road shape obtained by the algorithm is close to the original area, the simplification is stopped, and finally a simplified road contour with an area close to the true area is obtained.

[0067] The specific process is as follows: the road recognition result extracted in step S2 is denoised and vectorized to obtain a planar road feature map, the optimized Douglas-Peucker simplification algorithm is used to regularize the planar road feature map, when the area of the regularized planar road feature map is greater than or equal to 90% of the original road vector map, a regularized planar road feature map is obtained; otherwise, the optimized Douglas-Peucker simplification algorithm is iteratively used to continue to regularize the planar road feature map. When the area of the road vector map obtained by using the algorithm is greater than or equal to 95% of the area of the original planar road feature map, the most reasonable road contour can be obtained.

[0068] S4 determines the crossroad skeleton line according to a preset determination condition based on the first feature information, to obtain label information of the crossroad skeleton line, and divides the crossroad skeleton line into an abnormal crossroad skeleton line and a normal crossroad skeleton line according to the label information;

[0069] Specifically, step S4 includes:

[0070] S41, the preset determination condition is:

[0071] The first condition is that the crossroad skeleton line has at least three lines;

[0072] The second condition is that both ends of one of the crossroad skeleton lines are three crossroad skeleton line coincidence points;

[0073] The third condition is that the ratio of the length of the road line corresponding to the intersection skeleton line to the width is less than 3;

[0074] The fourth condition is that the length of the intersection skeleton line is less than or equal to 100 m;

[0075] S42, the intersection skeleton line is determined according to a preset determination condition;

[0076] If the intersection skeleton line meets the first condition, the second condition, the third condition and the fourth condition at the same time, the label information of the intersection skeleton line is N;

[0077] If the intersection skeleton line does not meet any one of the first condition, the second condition, the third condition and the fourth condition, the label information of the intersection skeleton line is Y;

[0078] S43, the intersection skeleton line with the label information N is marked as an abnormal intersection skeleton line, and the intersection skeleton line with the label information Y is marked as a normal intersection skeleton line.

[0079] S5, a coordinate system is constructed based on the abnormal intersection skeleton line, neighborhood analysis is performed according to the coordinate system, interference feature information of the abnormal intersection skeleton line is filtered, and effective feature information is reserved;

[0080] In the embodiment of the application, step S5 comprises:

[0081] S51, three abnormal intersection skeleton lines are selected as end points, and two end points are connected to form a line segment;

[0082] S52, a coordinate system is constructed by taking the extension direction of the line segment as the X axis and taking the midpoint of the line segment as the coordinate origin, eight-neighborhood analysis is performed, and second feature information is obtained;

[0083] S53, the intersection point of the abnormal intersection skeleton line and the non-intersection skeleton line is connected to form two regions;

[0084] S54, the region close to the coordinate origin, i.e., the abnormal intersection skeleton line region, is taken as an interference region, and the region far away from the coordinate point, i.e., the non-intersection skeleton line region, is taken as an effective region;

[0085] S55, in the second feature information, interference feature information of the interference region is filtered, and the feature information of the dividing point and the effective region is taken as effective feature information.

[0086] In the embodiment of the application, please refer to Figure 3, according to the preset determination condition, it is determined that L1, L2, L3, L4 and L are all abnormal intersection skeleton lines, and a label N is marked; the endpoint A of the L intersection skeleton line is the intersection point of the L1, L and L3 intersection skeleton lines, and the endpoint B is the intersection point of the L1, L and L4 intersection skeleton lines; connecting the A and B points forms a line segment, the line segment coincides with L, that is, the extension direction of L is taken as the X axis, the midpoint of L is taken as the coordinate origin, a coordinate system is constructed, and eight quadrants are divided, eight-neighborhood analysis is carried out based on the coordinate axes, and feature information is obtained; it is determined that C, D, E and F are boundary points, then two regions formed by connecting C, D, E and F are filtered to obtain feature information of the interference region, the feature information of the boundary points C, D, E and F and the effective region is reserved as effective feature information, and the effective feature information includes the distances C(x1, y1), D(x2, y2), E(x3, y3) and F(x4, y4) from the skeleton line endpoints to the coordinate axes and the quadrants in which the boundary points are located.

[0087] S6 encodes the effective feature information, and takes the remote sensing image with the encoded effective feature information and label information as a sample set;

[0088] S7 inputs the sample set into a pre-trained road recognition model, iteratively trains until a convergence condition is met, the model training is completed, and an optimized model is obtained;

[0089] S8 uses the optimized model to automatically recognize roads in a new remote sensing image.

[0090] In the embodiment of the application, the Focalloss function is introduced to calculate the loss value, and the calculation formula of the Focalloss is as follows:

[0091]

[0092] Wherein, L is the loss function value, y represents the category label, and y' is the output after the activation function. The modulation factor γ is a number greater than 0, which can reduce the influence of easily classified samples on the calculation of the loss value. When γ is 0, it is a cross-entropy loss function, and when γ increases, the influence of the adjustment factor also increases.

[0093] The sample set is input into the road recognition model for learning, a loss value is calculated through a loss function, and the prediction error loss value is fed back to the network for gradient feedback to correct the parameters of each module unit in the network. Through a predetermined number of loop calculations, the final prediction error is within a set threshold range, the prediction accuracy is within an expected range, a network model with accurate road recognition capability is obtained, remote sensing images can be subjected to road recognition, the road recognition result is output through a classifier in the model, the recognition result includes roads and abnormal crossroads, and the recognition result is post-processed. The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any changes or replacements that can be easily thought of by those skilled in the art within the technical range disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An automatic identification method of a road in a remote sensing image, comprising the following steps: S1: acquiring a remote sensing image and pre-processing the same; S2: using a pre-trained road identification model to identify the road in the remote sensing image, to obtain a linear road feature map; S3: performing repair, vectorization and regularization processing on the linear road feature map, to obtain a regular planar road feature map and first feature information; the regular planar road feature map comprises a skeleton line, and the skeleton line comprises an intersection skeleton line and a non-intersection skeleton line; wherein the intersection skeleton line comprises a crossroad intersection skeleton line; S4: according to the first feature information, determining the crossroad intersection skeleton line through a preset determination condition, to obtain label information of the crossroad intersection skeleton line, and dividing the crossroad intersection skeleton line into an abnormal crossroad intersection skeleton line and a normal crossroad intersection skeleton line according to the label information; S5: constructing a coordinate system based on the abnormal crossroad intersection skeleton line, performing neighborhood analysis according to the coordinate system, filtering interference feature information of the abnormal crossroad intersection skeleton line, and retaining effective feature information; S6: performing encoding processing on the effective feature information, and taking the remote sensing image with the effective feature information and the label information after the encoding processing as a sample set; S7: inputting the sample set into the pre-trained road identification model, performing iterative training until a convergence condition is met, ending the model training, and obtaining an optimized model; S8: using the optimized model to automatically identify the road in a new remote sensing image; The road identification model comprises a first convolution module X1, an encoding part C, a decoding part D, an intermediate bridging structure L0 between the encoding part C and the decoding part D, a pyramid pooling part P and a classifier E connected in sequence, wherein the encoding part C and the decoding part D are both composed of batch normalization, an activation function and a convolution operation; The encoding part C adopts an improved U-Net comprising four-level residual networks X2-X5, and the improved U-Net further comprises a diagonal lower connection structure L1 and a diagonal upper connection structure L2; The decoding part D comprises three-level convolution modules Y1-Y3, wherein, The convolution module Y3 is connected with the residual network X2 through the intermediate bridging structure L0, connected with the convolution module X1 through the lower connection structure L1, and connected with the residual network X3 through the upper connection structure L2; The convolution module Y2 is connected with the residual network X3 through the intermediate bridging structure L0, connected with the residual network X2 through the lower connection structure L1, and connected with the residual network X4 through the upper connection structure L2; The convolution module Y1 is connected with the residual network X4 through the intermediate bridging structure L0, connected with the residual network X3 through the lower connection structure L1, and connected with the residual network X5 through the upper connection structure L2.

2. The automatic identification method of a road in a remote sensing image according to claim 1, wherein the first feature information in step S3 comprises road length, contour area, road width, road direction, skeleton line length and skeleton line direction.

3. The method of claim 2, wherein step S4 comprises: S41, wherein the preset determination condition is: a first condition: the intersection skeleton line has at least three lines; a second condition: both ends of one of the intersection skeleton lines are coincident with three intersection skeleton lines; a third condition: the ratio of the length of the intersection skeleton line to the width of the corresponding road line is less than 3; a fourth condition: the length of the intersection skeleton line is less than or equal to 100 m; S42, wherein the intersection skeleton line is determined according to the preset determination condition: if the intersection skeleton line meets the first condition, the second condition, the third condition and the fourth condition, the label information of the intersection skeleton line is N; if the intersection skeleton line does not meet any one of the first condition, the second condition, the third condition and the fourth condition, the label information of the intersection skeleton line is Y; S43, wherein the intersection skeleton line with the label information N is marked as an abnormal intersection skeleton line, and the intersection skeleton line with the label information Y is marked as a normal intersection skeleton line.

4. The method of claim 3, wherein step S5 comprises: S51, wherein the coincident points of the three abnormal intersection skeleton lines are taken as end points, and two end points are connected to form a line segment; S52, wherein a coordinate system is constructed by taking the extension direction of the line segment as the X axis and taking the midpoint of the line segment as the coordinate origin, and eight-neighborhood analysis is performed to obtain second characteristic information; S53, wherein the coincident points of the abnormal intersection skeleton line and the non-intersection skeleton line are taken as boundary points, and the boundary points are connected to form two regions; S54, wherein the region close to the coordinate origin, i.e., the abnormal intersection skeleton line region, is taken as an interference region, and the region far from the coordinate point, i.e., the non-intersection skeleton line region, is taken as an effective region; S55, wherein in the second characteristic information, the interference characteristic information of the interference region is filtered, and the characteristic information of the boundary points and the effective region is taken as effective characteristic information.

5. The method of claim 1, wherein the effective characteristic information in step S4 comprises road length, width and direction, skeleton line length and direction, distance from the skeleton line end point to the coordinate axis, and the quadrant in which the skeleton line end point is located.

6. The method of claim 1, wherein step S3 comprises: S31, wherein the contour boundary of the road is smoothed, the holes of the road are filled, and the road line gaps are repaired to obtain a repaired linear road feature map; S32, wherein the repaired linear road feature map is vectorized by using ArcGIS, and the skeleton line and the first characteristic information are extracted to obtain a planar road feature map; S33, wherein the planar road feature map is regularized by using a simplification algorithm to obtain a regular planar road feature map.

7. The method of claim 1, wherein the preprocessing comprises grayscale processing and filtering processing.

8. The method of claim 6, wherein the simplification algorithm is Douglas-Puke.

Images