Community road network extraction method combining trajectory position and neighborhood structure supervision

Abstract

The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, comprising the following steps: collecting trajectory superposition building footprint maps and building road maps of a plurality of community areas based on a third-party map software; taking the trajectory superposition building footprint maps as samples and corresponding building road maps as labels to generate input target image pairs, and constructing a training set and a verification set based on the input target image pairs; inputting the trajectory superposition building footprint maps in the verification set into a road generation model, and outputting corresponding result building road images by the road generation model, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training; and filtering and splicing the output result building road images to obtain a complete community road network. In the case of sparse and missing trajectories, the building footprint information is used for supervision to generate a community fine road with high precision and close to the real one.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geographic information, and in particular to a community road network extraction method combining trajectory position and neighborhood structure supervision. BACKGROUND

[0002] The urban community refined road network refers to the roads and branch elements in the urban community, and is the key information for the last mile of seamless navigation in the city. With the advancement of urbanization in China, the number of communities is gradually increasing, the scale is expanding, and the road network in the community is increasing and becoming more complex. Unlike the extraction of urban trunk road information, the shielding of multiple neighborhood elements in the community local environment and other factors cause the complexity of the community internal road scene to be high, and it is difficult to automatically extract it, which leads to the fact that the community-level road network information in many electronic maps is not complete enough, the information refinement degree is not enough or even missing, which causes the community navigation to have a "blind area", and causes great obstacles to the development of seamless navigation technology in the city. Therefore, accurately updating and enriching the community road network information is an urgent problem, which will help to alleviate the lost and detour behavior in the community scene, facilitate resident travel and save travel time.

[0003] In recent years, people's travel trajectory data is increasingly rich, and the travel trajectory records rich road geometry, topology and attribute information, which can be used to extract the spatial geometric characteristics of the road network. Therefore, using GPS trajectory data to extract the geometric structure and topological relationship of the road network has become a hot spot in the research of road information extraction. At present, there are three main methods for extracting road network based on trajectory data: one is the trajectory clustering method; the second is the trajectory incremental merging method; the third is the method of graph theory and image processing. However, these three kinds of community road extraction methods based on trajectory data respectively face the following challenges: (1) the shielding of multiple neighborhood elements in the community local environment and other factors, which leads to the problems of sparse trajectory data in the community, large density difference and high noise, which will seriously affect the road extraction result; (2) the algorithm needs manual intervention and cannot automatically and quickly extract the road network, which is time-consuming and laborious; (3) the adaptability of model algorithm parameters to different regions and different trajectory densities is poor, and the density, motion characteristics, noise distribution and other differences of trajectory in the road network space cannot be considered, which makes the accuracy and completeness of the extraction result not high. In summary, there are many difficulties in extracting community-level road network based on GPS trajectory data.

[0004] Therefore, how to avoid the interference of high noise in community trajectory data in the road network extraction process, and the time-consuming and laborious road network extraction, and the low accuracy and completeness of the extraction result, is still a problem to be solved by those skilled in the art. SUMMARY

[0005] The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, which is used to solve the problems of interference caused by high noise of community trajectory data in the existing road network extraction process, time and labor consumption of road network extraction, and low precision and completeness of extraction results.

[0006] The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, which comprises the following steps:

[0007] Based on third-party map software, trajectory superimposed building footprint maps and building road maps of multiple community areas are collected.

[0008] The trajectory superimposed building footprint maps are taken as samples, and the corresponding building road maps are taken as labels to generate input target image pairs, and a training set and a verification set are constructed based on the input target image pairs.

[0009] The trajectory superimposed building footprint maps in the verification set are input into a road generation model, and the road generation model outputs corresponding result building road images, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training.

[0010] The output multiple result building road images are filtered and spliced to obtain a complete community road network.

[0011] According to the community road network extraction method combining trajectory position and neighborhood structure supervision, trajectory superimposed building footprint maps and building road maps of multiple community areas are collected based on third-party map software, and the method specifically comprises the following steps:

[0012] Based on the positioning of each community center point coordinate to each community area range in the third-party map software, the trajectory superimposed building footprint maps and the building road maps in each community area range are intercepted, and the size of the trajectory superimposed building footprint maps and the building road maps is 256*256.

[0013] According to the community road network extraction method combining trajectory position and neighborhood structure supervision, the trajectory superimposed building footprint maps are taken as samples, and the corresponding building road maps are taken as labels to generate input target image pairs, and the method specifically comprises the following steps:

[0014] The trajectory superimposed building footprint maps and the building road maps intercepted in the same area are horizontally spliced into 512*256 images to obtain the input target image pairs.

[0015] The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, a MAC-GAN generative conditional adversarial network is composed of a generator and a discriminator, and a loss function in a training process of the MAC-GAN generative conditional adversarial network is composed of an adversarial loss of the generator and the discriminator and a reconstruction loss of a generated result building road image and a building road image label.

[0016] The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, in a detection model training process, the generator adopts a channel attention block with multi-scale skip connection and adaptive weighted fusion, and a MACU-Net integrated with an asymmetric convolution block is configured.

[0017] The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, the discriminator is configured with a 70*70 receptive field PatchGAN.

[0018] The application provides a community road network extraction method combining trajectory position and neighborhood structure supervision, the multiple output result building road images are filtered and spliced to obtain a complete community road network, and the method specifically comprises the following steps.

[0019] The multiple output result building road images are subjected to rank filtering processing, and the result building road images subjected to the rank filtering processing and belonging to a unified community are horizontally spliced to obtain a complete community road network of each community.

[0020] The application further provides a community road network extraction device combining trajectory position and neighborhood structure supervision, and the device comprises the following components.

[0021] A collection unit is configured to collect trajectory superimposed building footprint maps and building road maps of multiple community areas based on a third-party map software.

[0022] A construction unit is configured to generate input target image pairs by taking the trajectory superimposed building footprint maps as samples and corresponding building road maps as labels, and construct a training set and a verification set based on the input target image pairs.

[0023] A generation unit is configured to input the trajectory superimposed building footprint maps in the verification set into a road generation model, and the road generation model outputs corresponding result building road images, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training.

[0024] A splicing unit is configured to filter and splice the multiple output result building road images to obtain a complete community road network.

[0025] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the community road network extraction method supervised by joint trajectory position and neighborhood structure when executing the program.

[0026] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the community road network extraction method supervised by joint trajectory position and neighborhood structure.

[0027] The application provides a community road network extraction method supervised by joint trajectory position and neighborhood structure, which comprises the following steps: collecting trajectory superimposed building footprint maps and building road maps of a plurality of community areas based on third-party map software; taking the trajectory superimposed building footprint maps as samples and the corresponding building road maps as labels to generate input target image pairs, and constructing a training set and a verification set based on the input target image pairs; inputting the trajectory superimposed building footprint maps in the verification set into a road generation model, and outputting corresponding result building road images from the road generation model, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training; and filtering and splicing a plurality of output result building road images to obtain a complete community road network. In the case of sparse and missing trajectories, the community fine road with high precision and close to reality is generated by building footprint information supervision. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative effort.

[0029] Figure 1 The flowchart of the community road network extraction method supervised by joint trajectory position and neighborhood structure provided by the application is shown in the figure.

[0030] Figure 2 The schematic diagram of the multi-scale skip connection and channel attention block in the road generation model provided by the application is shown in the figure.

[0031] Figure 3 The flowchart of the community road network extraction method supervised by joint trajectory position and neighborhood structure MAC-GAN provided by the application is shown in the figure.

[0032] Figure 4A structural schematic diagram of a community road network extraction device provided by the present application is provided.

[0033] Figure 5 A structural schematic diagram of an electronic device provided by the present application is provided. DETAILED DESCRIPTION

[0034] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0035] In the prior art, there are problems such as interference caused by high noise of community trajectory data in the road network extraction process, time and labor consumption in road network extraction, and low accuracy and completeness of the extraction result. The present application will be described below in combination with Figure 1 The present application provides a community road network extraction method supervised by combined trajectory position and neighborhood structure. Figure 1 A flowchart of the community road network extraction method supervised by combined trajectory position and neighborhood structure provided by the present application is shown in Figure 1 The method comprises the following steps.

[0036] In step 110, based on third-party map software, trajectory superimposed building footprint maps and building road maps of a plurality of community areas are collected.

[0037] Specifically, third-party map software such as Baidu Map, Gaode Map, etc. is selected, and two types of image data sets of each community area are intercepted through the map API self-defined map style function of the third-party map software: the first type: a map retaining buildings and superimposed pedestrian trajectories; the second type: a map retaining only buildings and roads; the pedestrian trajectories are messy, sparse and uneven, and occasionally missing, and the purpose of the present application is to derive a map including community buildings and community roads based on the known map of buildings superimposed with pedestrian trajectories. Therefore, first, the map retaining buildings and superimposed pedestrian trajectories and the map retaining only buildings and roads of each community are collected to provide training samples and training labels for subsequent model training.

[0038] In step 120, the trajectory superimposed building footprint map is taken as a sample, the corresponding building road map is taken as a label to generate an input target image pair, and a training set and a verification set are constructed based on the input target image pair.

[0039] Specifically, the trajectory superimposed building footprint map is taken as a sample, the corresponding building road map is taken as a label, a building superimposed trajectory-building road is constructed as an input-target pair, which can also be called a sample-label pair, and a large number of input-target pairs are divided into a training set and a verification set in proportion, wherein the training set is used to train a model for subsequently generating a building road result picture, and the verification set is used to input a building superimposed trajectory picture into the trained model to output a building road result picture.

[0040] In step 130, the trajectory superimposed building footprint map in the verification set is input into a road generation model, and the road generation model outputs a corresponding result building road picture, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training.

[0041] Specifically, the training set is input into the MAC-GAN generative conditional adversarial network, and the spatial relationship between the trajectory position and the building footprint is taken as prior knowledge, so that the MAC-GAN generative conditional adversarial network gradually approaches the optimal result by learning and optimizing the model parameters of the generative-adversarial recurrent learning conditional generative adversarial network through sufficient learning of the ternary spatial features of “road-trajectory-building footprint”, and applying the MAC-GAN model optimized by learning to the verification set to convert the input trajectory superimposed building footprint map into a building road map to extract the roads in each region in the community.

[0042] In step 140, the output multiple result building road pictures are filtered and spliced to obtain a complete community road network.

[0043] Specifically, the generated series of result pictures are filtered and spliced to obtain a community overall road picture, thereby completing the extraction of the community road network.

[0044] Based on the above embodiment, in the method, the trajectory superimposed building footprint map and the building road map of a plurality of community regions are collected based on a third-party map software, and specifically include:

[0045] Based on the positioning of each community center point coordinate to each community region range in the third-party map software, the trajectory superimposed building footprint map and the building road map in each community region range are intercepted, and the size of the trajectory superimposed building footprint map and the building road map is 256*256.

[0046] Specifically, taking Baidu Map as an example, a trajectory superimposed building footprint map and a building road map are extracted: first, two kinds of map styles are defined by using Baidu Map API: (1) keep buildings and superimpose trajectories; (2) keep buildings and roads. Among them, the building footprint is represented by orange pixels, the trajectory is represented by black pixels, and the road is represented by green pixels, while the road grade is distinguished by line width. Since the geometric shapes of building footprints and roads in the community are small, in order to capture the context features while maintaining high visual clarity, the community center coordinates are used to locate the specific community area range in the map, and the 19th level map image size of 1024*512 target area is intercepted. Then, the target area is divided into 2 512*512 images, and then uniformly scaled to 256x256 size.

[0047] Based on any one of the above embodiments, in the method, the input target image pair is generated by taking the trajectory superimposed building footprint map as a sample and the corresponding building road map as a label.

[0048] The trajectory superimposed building footprint map and the building road map intercepted in the same area are horizontally spliced into a 512*256 image to obtain the input target image pair.

[0049] Specifically, the input-target pair is obtained according to the following manner: the building superimposed trajectory image and the building road are horizontally spliced into an image with a size of 512*256, the training set and the validation set are both 512*256, and they are both input-target image pairs. However, the target image is not used in the validation of the validation set, and the target image is used in the calculation of the evaluation index; preferably, the division ratio of the training set and the validation set is 10:1.

[0050] Based on any one of the above embodiments, in the method, the MAC-GAN generative conditional adversarial network is composed of a generator and a discriminator, and the loss function in the training process of the MAC-GAN generative conditional adversarial network is composed of the adversarial loss of the generator and the discriminator and the reconstruction loss of the generated result building road image and the building road image label.

[0051] Specifically, the MAC-GAN is designed to receive a 256*256 conditional image as input, and in each forward pass, the generator converts the input into a generated image filled with road data by fully capturing the features of the input image (building footprint superimposed trajectory) to "deceive" the discriminator to classify the generated image as "real", while the discriminator will learn to classify the generated image as "fake" and the ground truth target as "real". The discriminator is configured by a Markov discriminator (PatchGAN), and the generated image and the real image are respectively connected to the input image after the conditional image is input, and the probability (1 or 0) of the image being true or false is calculated through convolution operation. The overall loss function is designed as the sum of the adversarial loss of the generator and the discriminator and the reconstruction loss of the generated image and the target image, and the weights of the generator and the discriminator are updated through the back propagation algorithm. Through such continuous cyclic alternation learning and optimization, the training is completed until the discriminator cannot distinguish whether the picture generated by the generator is fake, and the trained model file is saved.

[0052] Based on any of the above embodiments, in the method, the generator is configured with a channel attention block with multi-scale skip connection and adaptive weighted fusion, and a MACU-Net integrated with an asymmetric convolution block.

[0053] Specifically, the generator is configured with a channel attention block with multi-scale skip connection and adaptive weighted fusion, and a MACU-Net integrated with an asymmetric convolution block ACB. The asymmetric convolution block is constructed using three branches of 3*3 square convolution, 1*3 convolution (horizontal kernel), and 3*1 convolution (vertical kernel). Figure 2 The multi-scale skip connection and channel attention block in the road generation model provided by the present application are shown in the schematic diagram Figure 2 As shown, the multi-scale context skip connection structure specifically includes: the feature maps of the same level encoder layer (i.e. ) are directly connected first; then, the fine-grained detailed road geometry and topology information contained in the low-level decoder layer (i.e. and ) are transmitted through transposed convolution and ACB; finally, the coarse-grained road context semantic information contained in the higher-level encoder layer (i.e. and ) is transmitted by the maximum pooling layer and ACB. As shown in Figure 2After the multi-scale skip connection of the five feature maps with the same size and resolution, the channel attention block CAB further reduces the number of channels and rearranges the channel features. Taking the channel attention block CAB3 as an example, 1x1 convolution and 128 filters are used to reduce the initial number of channels. Then, the spatial dimension is compressed by using the operation of average pooling and maximum pooling at the same time. The compressed channel of the feature map is compressed to 1 / 16 of the original size through two convolution layers with 8 filters and ReLU activation function. Then the number of channels is recovered by using two convolution layers with 128 filters. Finally, the two layers are activated by the sigmoid function, multiplied by the output of the first convolution and passed through two ACB convolution layers, and the final output is obtained They constitute a jointly working encoding-decoding structure.

[0054] Based on any of the above embodiments, in the method, the discriminator selects a PatchGAN configuration with a 70x70 receptive field.

[0055] Specifically, the discriminator selects a PatchGAN configuration with a 70x70 receptive field, which includes five convolution layers. Except that the last two convolution layers have a step size of 1 and a convolution kernel size of 4x4, the other convolution layers have a step size of 2 and a convolution kernel size of 4x4. A Leaky-ReLU activation function with regular instantiation and a parameter of 0.2 is set between every two convolution layers.

[0056] The application firstly defines two types of map styles through Baidu map API, positions the community center point coordinates to a specific community area range in the map, and intercepts a 1024*512 target area 19th level map image dataset: (1) a map retaining buildings and superimposing trajectories, and (2) a map retaining only roads. Then, the target area is evenly divided into two 512*512 images, and then uniformly scaled to 256x256 size. Using the two types of image datasets, the same position of the two images is horizontally spliced to construct an input-target pair: building superimposed trajectory-building road, and divided into a training set and a validation set in a ratio of 10:1, wherein the training set includes 750 images and the validation set includes 74 images; then the training set is input into the MAC-GAN conditional generative adversarial network, the spatial relationship between the trajectory position and the building footprint is used as prior knowledge, so that the generator and the discriminator learn the "road-trajectory-building footprint" ternary spatial features through sufficient learning, and the "generation-game" continuously cyclic learning optimization conditional generative adversarial network model parameters, until gradually approaching the optimal result, and saving the trained MAC-GAN model file; then the MAC-GAN model optimized by learning is applied to the validation set to convert the input image into a building road image to extract the roads in each area of the community; finally, the generated series of result images are processed by rank filtering and splicing to obtain the community overall road image, thereby completing the extraction of the community road network.

[0057] Based on the above embodiment, the application provides a MAC-GAN community road network extraction method combining trajectory position and neighborhood structure supervision, Figure 3 The flowchart of the MAC-GAN community road network extraction method combining trajectory position and neighborhood structure supervision provided by the application is shown in Figure 3

[0058] The method comprises the following steps:

[0059] Step 1), using community center point coordinates, defining map styles through Baidu map API to intercept image datasets of each community area, which are (1) a map retaining buildings and superimposing trajectories, and (2) a map retaining only roads. Using the two types of image datasets to construct an input-target pair: building superimposed trajectory-building road, and dividing into a training set and a validation set in a ratio;

[0060] The specific process of the dataset is as follows:

[0061] ​The data set includes a training data set and a validation data set; first, the community center point coordinates are used to locate a specific community area range in the map, and the 19th level map image data set of the target area of 1024*512 is intercepted: (1) the map retaining buildings and superimposing trajectories, (2) the map retaining only roads. Then the target area is evenly divided into 2 512*512 images, and then uniformly scaled to 256x256 size. Using these two types of image data sets, two images at the same position are horizontally spliced to construct input-target pairs: building superimposed trajectory-building road, and divided into training set and validation set in the ratio of 10:1, wherein the training set includes 750, and the validation set includes 74.

[0062] Step 2) input the training set into the MAC-GAN generative conditional adversarial network, use the spatial relationship between the trajectory position and the building footprint as prior knowledge, so that the generator and the discriminator can learn the "road-trajectory-building footprint" ternary spatial features through sufficient learning, and continuously cycle learning and optimization of the conditional generative adversarial network model parameters, gradually approaching the optimal result;

[0063] The specific structure of the MAC-GAN model is as follows:

[0064] The MAC-GAN is mainly composed of a generator G and a discriminator D. The method uses a channel attention block with multi-scale skip connection and adaptive weighted fusion, and integrates a MACU-Net with an asymmetric convolution block ACB to configure the generator. The asymmetric convolution block uses three branches of 3x3 square convolution, 1x3 convolution (horizontal kernel), and 3x1 convolution (vertical kernel) to construct. The multi-scale context skip connection structure is as follows: the feature maps of the same level encoder layer (i.e. ) are first directly connected; then, the fine-grained detailed road geometry and topological information contained in the low-level decoder layer (i.e. and ) is transmitted through transposed convolution and ACB; finally, the coarse-grained road context semantic information contained in the higher level encoder layer (i.e. and ) is transmitted by the maximum pooling layer and ACB. As Figure 4After the five feature maps of the same size and resolution are connected by multi-scale skip connection, the channel attention block CAB further reduces the number of channels and rearranges the channel features. Taking the channel attention block CAB3 as an example, 1x1 convolution and 128 filters are used to reduce the initial number of channels. Then, the spatial dimension is compressed by using the operation of average pooling and maximum pooling at the same time. The compressed feature map channel is compressed to 1 / 16 of the original size through two convolution layers with 8 filters and ReLU activation function. Then the number of channels is restored using two convolution layers and 128 filters. Finally, the two layers are activated by the sigmoid function, multiplied by the output of the first convolution, and passed through two ACB convolution layers to obtain the final output. They form a jointly working encoding-decoding structure. The PatchGAN with a 70x70 receptive field is selected to configure the discriminator of the MAC-GAN.

[0065] The specific working process of the MAC-GAN model is as follows:

[0066] The MAC-GAN is designed to receive a 256x256 conditional image as input. In each forward pass, the generator converts the input into a generated image filled with road data by fully capturing the features of the input image (building footprint superimposed trajectory), in order to "deceive" the discriminator and make it classify the generated image as "real". The discriminator will learn to classify the generated image as "fake" and the ground truth target as "real". The discriminator is configured by the Markov discriminator (PatchGAN). The generated image and the real image are connected with the conditional image input image, and then input through convolution operation to calculate the probability (1 or 0) of the image being true or false. The adversarial loss of the generator and the discriminator and the reconstruction loss of the generated image and the target image are designed, and the weights of the generator and the discriminator are updated through the back propagation algorithm. Through such continuous cycle alternation learning and optimization, the training is completed when the discriminator cannot distinguish whether the image generated by the generator is fake, and the trained model file is saved.

[0067] Step 3), then apply the MAC-GAN model optimized by learning to the validation set to convert the input image into a building road image to extract the road in each area of the community;

[0068] The specific implementation is as follows:

[0069] The validation set prepared in step 1) is input into the saved MAC-GAN model to obtain a series of result images generated by the model.

[0070] Step 4) filter and splice the generated series of result images to obtain the overall road image of the community, thereby completing the extraction of the road network of the community.

[0071] The specific implementation is as follows: the result image generated in step 3) is post-processed by rank filtering, the 256*256 road images generated in a community area are horizontally spliced into 512*256, and finally the community overall road image is obtained.

[0072] The community road network extraction device provided by the application is described below, and the community road network extraction device described below can be correspondingly referred to the community road network extraction method described above.

[0073] Figure 4 The structure diagram of the community road network extraction device provided by the application is shown in the drawing, and the device comprises an acquisition unit 410, a construction unit 420, a generation unit 430 and a splicing unit 440. Figure 4

[0074] The acquisition unit 410 is configured to acquire the trajectory superimposed building footprint map and the building road map of a plurality of community areas based on a third-party map software.

[0075] The construction unit 420 is configured to generate an input target image pair by taking the trajectory superimposed building footprint map as a sample and the corresponding building road map as a label, and construct a training set and a verification set based on the input target image pair.

[0076] The generation unit 430 is configured to input the trajectory superimposed building footprint map in the verification set into a road generation model, and the road generation model outputs a corresponding result building road image, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training.

[0077] The splicing unit 440 is configured to filter and splice the output plurality of result building road images to obtain a complete community road network.

[0078] ​The application provides a community road network extraction device combined with trajectory position and neighborhood structure supervision, which collects trajectory superposition building footprint maps and building road maps of a plurality of community areas based on a third-party map software; takes the trajectory superposition building footprint maps as samples, corresponding building road maps as labels to generate input target image pairs, constructs a training set and a verification set based on the input target image pairs; inputs the trajectory superposition building footprint maps in the verification set into a road generation model, and the road generation model outputs corresponding result building road images, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training; and filters and splices a plurality of output result building road images to obtain a complete community road network. In the case of sparse and missing trajectories, building footprint information is used for supervision to generate a community fine road with high precision and close to reality.

[0079] Based on the above-mentioned embodiments, in the device, during the training process of the detection model, the trajectory superposition building footprint maps and the building road maps of a plurality of community areas are collected based on the third-party map software, and specifically include:

[0080] Based on the positioning of each community center point coordinate to each community area range in the third-party map software, the trajectory superposition building footprint maps and the building road maps in each community area range are intercepted, and the size of the trajectory superposition building footprint maps and the building road maps is 256*256.

[0081] Based on the above-mentioned embodiments, in the device, the trajectory superposition building footprint maps are taken as samples, and corresponding building road maps are taken as labels to generate input target image pairs, and specifically include:

[0082] The trajectory superposition building footprint maps and the building road maps intercepted in the same area are horizontally spliced into 512*256 images to obtain the input target image pairs.

[0083] Based on the above-mentioned embodiments, in the device, the MAC-GAN generative conditional adversarial network is composed of a generator and a discriminator, and the loss function in the training process of the MAC-GAN generative conditional adversarial network is composed of the adversarial loss of the generator and the discriminator, and the reconstruction loss of the generated result building road images and the building road image labels.

[0084] Based on the above-mentioned embodiments, in the device, the generator adopts a channel attention block with multi-scale skip connection and adaptive weighted fusion, and a MACU-Net integrated with an asymmetric convolution block for configuration.

[0085] On the basis of the above embodiment, in the device, the discriminator selects a PatchGAN configuration with a 70x70 receptive field.

[0086] On the basis of the above embodiment, in the device, the filtering and splicing of the output multiple result building road images to obtain a complete community road network specifically includes:

[0087] The output multiple result building road images are processed by rank filtering, and the result building road images belonging to a unified community after the rank filtering are horizontally spliced to obtain a complete community road network of each community.

[0088] Figure 5 An example of an entity structure diagram of an electronic device is shown in FIG. 1. Figure 5 As shown in FIG. 1, the electronic device can include a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 complete mutual communication through the communications bus 540. The processor 510 can invoke a logical instruction in the memory 530 to execute a joint trajectory position and neighborhood structure supervised community road network extraction method, which includes: based on a third-party map software, collecting trajectory superimposed building footprint maps and building road maps of multiple community areas; taking the trajectory superimposed building footprint maps as samples and the corresponding building road maps as labels to generate input target image pairs, constructing a training set and a verification set based on the input target image pairs; inputting the trajectory superimposed building footprint maps in the verification set into a road generation model, and the road generation model outputs corresponding result building road images, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training; and filtering and splicing the output multiple result building road images to obtain a complete community road network.

[0089] In addition, the logical instructions in the memory 530 described above can be implemented in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0090] In another aspect, the present application also provides a computer program product, which comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions, when the program instructions are executed by a computer, the computer can execute the community road network extraction method provided by the above-mentioned method, which comprises: based on a third-party map software, collecting a plurality of community area trajectory superimposed building footprint maps and building road maps; taking the trajectory superimposed building footprint map as a sample and the corresponding building road map as a label to generate an input target image pair, and constructing a training set and a verification set based on the input target image pair; inputting the trajectory superimposed building footprint map in the verification set into a road generation model, and the road generation model outputs a corresponding result building road image, wherein the road generation model is obtained by training the training set into a MAC-GAN generative conditional adversarial network; filtering and splicing the output multiple result building road images to obtain a complete community road network.

[0091] In yet another aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the above-provided community road network extraction method supervised by joint trajectory position and neighborhood structure, the method comprising: collecting, based on a third-party map software, trajectory superimposed building footprint maps and building road maps of a plurality of community areas; generating input target image pairs by taking the trajectory superimposed building footprint maps as samples and corresponding building road maps as labels, constructing a training set and a verification set based on the input target image pairs; inputting the trajectory superimposed building footprint maps in the verification set into a road generation model, the road generation model outputting corresponding result building road images, wherein the road generation model is obtained by inputting the training set into a MAC-GAN generative conditional adversarial network for training; and filtering and splicing the output result building road images to obtain a complete community road network.

[0092] The device embodiments described above are merely illustrative, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0093] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0094] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for extracting community road networks under joint trajectory location and neighborhood structure supervision, characterized in that, include: Based on third-party map software, track data from multiple community areas were collected and overlaid with building footprint maps and building road maps. Using the trajectory overlaid with the building footprint map as a sample, and the corresponding building and road map as a label, input target image pairs are generated, and a training set and a validation set are constructed based on the input target image pairs; The trajectory overlaid with the building footprint map in the validation set is input into the road generation model, and the road generation model outputs the corresponding result building and road image. The road generation model is obtained by training the training set into the MAC-GAN generative conditional adversarial network. The output images of buildings and roads are filtered and stitched together to obtain a complete community road network. The MAC-GAN generative conditional adversarial network consists of two parts: a generator and a discriminator. The loss function during the training process of the MAC-GAN generative conditional adversarial network is composed of the adversarial loss of the generator and the discriminator, and the reconstruction loss of the generated building and road images and building and road image labels. The generator is configured with a channel attention block with multi-scale skip connections and adaptive weighted fusion, and a MACU-Net with integrated asymmetric convolutional blocks. The discriminator is configured with a 70×70 receptive field PatchGAN. The specific structure of the MAC-GAN model is as follows: MAC-GAN is mainly constructed from two parts: a generator (G) and a discriminator (D). The method uses a channel attention block with multi-scale skip connections and adaptive weighted fusion, and integrates an asymmetric convolutional block (ACB) into MACU-Net to configure the generator. The asymmetric convolutional block is constructed using three branches: 3×3 square convolution, 1×3 convolution, and 3×1 convolution. The multi-scale context skip connection structure is as follows: feature maps of the same level encoder layer are directly connected; road geometry and topology information contained in the lower-level decoder layer is passed through transposed convolution and ACB; road context semantic information contained in the higher-level encoder layer is transmitted by max pooling layer and ACB; after connecting five feature maps of the same size and resolution through multi-scale skip connections, the channel attention block (CAB) further reduces the number of channels and rearranges the channel features. A PatchGAN with a 70×70 receptive field is selected to configure the discriminator of MAC-GAN. The specific working process of the MAC-GAN model is as follows: MAC-GAN is designed to receive a 256×256 conditional image as input. In each forward pass, the generator transforms the input image into a generated image filled with road data by fully capturing the features of the input image. The discriminator is configured by a Markov discriminator. The generated image and the real image are concatenated with the conditional image and the input image, respectively, and then input after convolution operation to calculate the probability of the image being real or fake.

2. The community road network extraction method based on joint trajectory location and neighborhood structure supervision according to claim 1, characterized in that, Based on third-party mapping software, trajectory data from multiple community areas were collected and overlaid with building footprint maps and building road maps, specifically including: Based on the coordinates of the center point of each community, the area of ​​each community in the third-party map software is located, and the trajectory overlaid with the building footprint map and the building road map within the area of ​​each community is extracted. The size of the trajectory overlaid with the building footprint map and the building road map is 256*256.

3. The community road network extraction method based on joint trajectory location and neighborhood structure supervision according to claim 2, characterized in that, The step of generating input target image pairs by using the trajectory overlaid with the building footprint map as a sample and the corresponding building and road map as a label specifically includes: By overlaying the traces captured from the same area onto the building footprint map and the building road map and horizontally stitching them together to form a 512*256 image, the input target image pair is obtained.

4. The community road network extraction method based on joint trajectory location and neighborhood structure supervision according to claim 1, characterized in that, The process of filtering and stitching together the output images of multiple buildings and roads to obtain a complete community road network specifically includes: The output building and road images are processed by rank filtering. The resulting building and road images belonging to the same community after rank filtering are horizontally stitched together to obtain the complete community road network of each community.

5. A community road network extraction device with joint trajectory location and neighborhood structure supervision, characterized in that, include: The data collection unit is used to collect trajectory overlay maps of buildings and road maps of multiple community areas based on third-party map software. The construction unit is used to generate input target image pairs by using the trajectory overlaid with the building footprint map as a sample and the corresponding building and road map as a label, and to construct a training set and a validation set based on the input target image pairs; The generation unit is used to input the trajectory overlaid with the building footprint map in the validation set into the road generation model, and the road generation model outputs the corresponding result building road image. The road generation model is obtained by training the training set into the MAC-GAN generative conditional adversarial network. The stitching unit is used to filter and stitch together multiple output building and road images to obtain a complete community road network; The MAC-GAN generative conditional adversarial network consists of two parts: a generator and a discriminator. The loss function during the training process of the MAC-GAN generative conditional adversarial network is composed of the adversarial loss of the generator and the discriminator, and the reconstruction loss of the generated building and road images and building and road image labels. The generator is configured with a channel attention block with multi-scale skip connections and adaptive weighted fusion, and a MACU-Net with integrated asymmetric convolutional blocks. The discriminator is configured with a 70×70 receptive field PatchGAN. The specific structure of the MAC-GAN model is as follows: MAC-GAN is mainly constructed from two parts: a generator (G) and a discriminator (D). The method uses a channel attention block with multi-scale skip connections and adaptive weighted fusion, and integrates an asymmetric convolutional block (ACB) into MACU-Net to configure the generator. The asymmetric convolutional block is constructed using three branches: 3×3 square convolution, 1×3 convolution, and 3×1 convolution. The multi-scale context skip connection structure is as follows: feature maps of the same level encoder layer are directly connected; road geometry and topology information contained in the lower-level decoder layer is passed through transposed convolution and ACB; road context semantic information contained in the higher-level encoder layer is transmitted by max pooling layer and ACB; after connecting five feature maps of the same size and resolution through multi-scale skip connections, the channel attention block (CAB) further reduces the number of channels and rearranges the channel features. A PatchGAN with a 70×70 receptive field is selected to configure the discriminator of MAC-GAN. The specific working process of the MAC-GAN model is as follows: MAC-GAN is designed to receive a 256×256 conditional image as input. In each forward pass, the generator transforms the input image into a generated image filled with road data by fully capturing the features of the input image. The discriminator is configured by a Markov discriminator. The generated image and the real image are concatenated with the conditional image and the input image, respectively, and then input after convolution operation to calculate the probability of the image being real or fake.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the community road network extraction method with joint trajectory location and neighborhood structure supervision as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the community road network extraction method with joint trajectory location and neighborhood structure supervision as described in any one of claims 1 to 4.

Images