Road network generation method and device, electronic equipment and storage medium

Abstract

Embodiments of the present disclosure provide a road network generation method and device, electronic equipment and storage medium. The road network generation method comprises: performing feature extraction on trajectories contained in a grid in a to-be-processed region to obtain trajectory features of the grid; inputting the trajectory features of the grid into a pre-trained road network generation model, sequentially performing encoding operation, multi-scale hollow convolution operation and decoding operation on the trajectory features to obtain road segmentation information corresponding to the grid output by the road network generation model; and generating road network data corresponding to the to-be-processed region based on the road segmentation information. In the embodiments of the present disclosure, the field of view can be expanded to the entire input feature through the hollow convolution operation, and multi-scale features can be integrated in the center part, so as to adapt to the narrowness, complexity, connectivity and large span of the road network, solve the problems of poor connectivity of road network data and difficulty in separating multiple lanes, and improve the accuracy of road network data.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and in particular, to a road network generation method and device, electronic equipment and storage medium. BACKGROUND

[0002] An electronic map, i.e., a digital map, is a map stored and consulted in a digital way by using computer technology. The electronic map can be applied to many aspects such as urban planning and construction, transportation, tourism, car navigation, and the like, and provides great convenience for people.

[0003] In the process of generating electronic map data, road network data is generated to represent the relationship between roads and lanes. The road network is the lifeblood of a city, and accurate, comprehensive and real-time updated road network data is crucial for the development of smart cities. With the widespread use of GPS (Global Positioning System) embedded devices, different types of mobile objects generate a large amount of GPS trajectories, which provides a new opportunity for the extraction of road network data.

[0004] The trajectory-based road network generation method based on deep learning has the advantages of not needing to set experience parameters and being able to utilize the prior knowledge of existing road networks. However, due to the narrowness, complexity, connectivity and large span of road networks, the network based on deep learning in the prior art cannot guarantee a sufficient field of view to learn the topological relationship of the road network, thereby leading to the problems of poor connectivity and difficulty in separating multiple lanes in the generated road network data. SUMMARY

[0005] In view of the above problems, the embodiments of the present disclosure provide a road network generation method, device, electronic equipment and storage medium to overcome the above problems or at least partially solve the above problems.

[0006] According to a first aspect of an embodiment of the present disclosure, a road network generation method is provided, comprising:

[0007] performing feature extraction on trajectories contained in a grid in a to-be-processed region to obtain trajectory features of the grid;

[0008] inputting the trajectory features of the grid into a pre-trained road network generation model, and sequentially performing an encoding operation, a multi-scale dilated convolution operation and a decoding operation on the trajectory features to obtain road segmentation information corresponding to the grid output by the road network generation model;

[0009] generating road network data corresponding to the to-be-processed region based on the road segmentation information.

[0010] Optionally, the track features include track features of a target channel pre-selected from a plurality of preset channels, and a cumulative weight proportion of the target channel in the plurality of channels is greater than a preset threshold.

[0011] Optionally, the target channel is selected by: performing first training on a first initial model including an attention network by using first sample data; the first sample data includes multi-channel track features of a first sample grid in a first sample region and a first label of the first sample grid, and the first label indicates whether the first sample grid actually contains road network data; each channel in the attention network is assigned a weight; and after the first training is completed, a weight proportion of each channel in the attention network is counted, and a channel with a cumulative weight proportion greater than a preset threshold is selected as the target channel.

[0012] Optionally, the road segmentation information includes road surface segmentation information and road centerline segmentation information.

[0013] Optionally, the road network generation model is trained by: clustering tracks in a second sample region to obtain clustering clusters, selecting a representative track from each clustering cluster, and taking a grid in the second sample region that successfully binds a road with the representative track as a second sample grid; performing second training on a second initial model including an encoding network, a multi-scale hollow convolution network and a decoding network connected in sequence by using second sample data; the second sample data includes track features of the second sample grid and a second label of the second sample grid, and the second label indicates actual road segmentation information of the second sample grid; and after the second training is completed, the second trained model is taken as the road network generation model.

[0014] Optionally, before the second trained model is taken as the road network generation model, the method further includes: predicting a grid in a third sample region by using the second trained model to obtain a third sample grid that is predicted to contain a road but actually does not contain a road; performing third training on the second trained model by using third sample data; the third sample data includes track features of the third sample grid and a third label of the third sample grid, and the third label indicates actual road segmentation information of the third sample grid; and taking the third trained model as the road network generation model includes: after the third training is completed, taking the third trained model as the road network generation model.

[0015] Optionally, before the road network data corresponding to the to-be-processed region is generated based on the road segmentation information, the road segmentation information corresponding to the grid and the track feature of the grid are input into a pre-trained road network optimization model to obtain the optimized road segmentation information of the grid output by the road network optimization model; and the road network data corresponding to the to-be-processed region is generated based on the road segmentation information, including generating the road network data corresponding to the to-be-processed region based on the optimized road segmentation information.

[0016] Optionally, the road network optimization model is trained in the following manner: a fourth sample grid is obtained, and the road network data in the fourth sample grid is eroded and dilated in multiple lanes to obtain changed road segmentation information of the fourth sample grid; a fourth initial model is trained using fourth sample data; the fourth sample data includes the changed road segmentation information of the fourth sample grid and the track feature of the fourth sample grid, and a fourth label of the fourth sample grid, the fourth label indicating actual road segmentation information of the fourth sample grid; and after the fourth training is completed, a model trained through the fourth training is taken as the road network optimization model.

[0017] According to a second aspect of embodiments of the present disclosure, a road network generation apparatus is provided, including:

[0018] The extraction module is configured to perform feature extraction on the track contained in the grid in the to-be-processed region to obtain the track feature of the grid.

[0019] The prediction module is configured to input the track feature of the grid into a pre-trained road network generation model, and sequentially perform an encoding operation, a multi-scale hollow convolution operation and a decoding operation on the track feature to obtain road segmentation information of the grid corresponding to the road network generation model output by the road network generation model.

[0020] The generation module is configured to generate road network data corresponding to the to-be-processed region based on the road segmentation information.

[0021] Optionally, the track feature includes a track feature of a target channel preselected from a plurality of channels, and an accumulated weight proportion of the target channel in the plurality of channels is greater than a preset threshold.

[0022] Optionally, the target channel is screened through the following modules: a first training module, configured to perform first training on a first initial model containing an attention network by using first sample data; the first sample data includes multi-channel trajectory features of a first sample grid in a first sample region and a first label of the first sample grid, the first label indicating whether the first sample grid actually contains road network data; the attention network assigns weights to the multi-channels respectively; and a first determination module, configured to, after the first training is completed, count weight proportions of the channels in the attention network, and select a channel with a cumulative weight proportion greater than a preset threshold as the target channel.

[0023] Optionally, the road segmentation information includes road surface segmentation information and road centerline segmentation information.

[0024] Optionally, the road network generation model is trained through the following modules: a first acquisition module, configured to cluster trajectories in a second sample region to obtain clustering clusters, select representative trajectories from the clustering clusters, and select a grid in the second sample region as a second sample grid if the grid is successfully bound to the representative trajectories for road matching; a second training module, configured to perform second training on a second initial model containing an encoding network, a multi-scale hollow convolution network and a decoding network connected in sequence by using second sample data; the second sample data includes trajectory features of the second sample grid and a second label of the second sample grid, the second label indicating actual road segmentation information of the second sample grid; and a second determination module, configured to, after the second training is completed, take a model trained by the second training as the road network generation model.

[0025] Optionally, the device further includes: a second acquisition module, configured to predict a grid in a third sample region by using the model trained by the second training, and acquire a grid predicted to contain a road but actually not containing a road as a third sample grid; a third training module, configured to perform third training on the model trained by the second training by using third sample data; the third sample data includes trajectory features of the third sample grid and a third label of the third sample grid, the third label indicating actual road segmentation information of the third sample grid; and the second determination module, specifically configured to, after the third training is completed, take a model trained by the third training as the road network generation model.

[0026] Optionally, the device further includes: an optimization module, configured to input road segmentation information corresponding to the grid and trajectory features of the grid into a pre-trained road network optimization model to obtain optimized road segmentation information of the grid output by the road network optimization model; and the generation module, specifically configured to generate road network data corresponding to the to-be-processed region based on the optimized road segmentation information.

[0027] Optionally, the road network optimization model is trained by a third obtaining module, configured to obtain a fourth sample grid, perform erosion and multi-lane dilation on road network data in the fourth sample grid, and obtain changed road segmentation information of the fourth sample grid; a fourth training module, configured to perform fourth training on a third initial model by using fourth sample data; the fourth sample data includes the changed road segmentation information of the fourth sample grid and trajectory features of the fourth sample grid, and a fourth label of the fourth sample grid, the fourth label indicating actual road segmentation information of the fourth sample grid; and a third determining module, configured to take the fourth training completed model as the road network optimization model after the fourth training is completed.

[0028] According to a third aspect of embodiments of the present disclosure, an electronic device is provided, comprising: one or more processors; and one or more computer-readable storage media having stored thereon instructions that, when executed by the one or more processors, cause the processors to perform the road network generation method according to any one of the preceding aspects.

[0029] According to a fourth aspect of embodiments of the present disclosure, a computer-readable storage medium is provided, having stored thereon a computer program which, when executed by a processor, causes the processor to perform the road network generation method according to any one of the preceding aspects.

[0030] Embodiments of the present disclosure provide a road network generation method and device, an electronic device and a storage medium. The trajectory features of a grid in a to-be-processed region are extracted to obtain trajectory features of the grid. The trajectory features of the grid are input into a pre-trained road network generation model. The trajectory features are sequentially subjected to an encoding operation, a multi-scale hollow convolution operation and a decoding operation to obtain road segmentation information corresponding to the grid output by the road network generation model. The road network data corresponding to the to-be-processed region is generated based on the road segmentation information. As can be seen, in the embodiments of the present disclosure, the trajectory features are sequentially subjected to an encoding operation, a multi-scale hollow convolution operation and a decoding operation in the road network generation model. The field of view can be expanded to the entire input feature by the hollow convolution operation, and multi-scale features can be integrated in the center part. Therefore, the road network generation model can adapt to the narrowness, complexity, connectivity and large span of the road network, solve the problems of poor connectivity and difficulty in separating multi-lane of the road network data, and improve the accuracy of the road network data. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present disclosure. Obviously, the drawings in the following description are only some of the drawings of the embodiments of the present disclosure, and other drawings can also be obtained by those of ordinary skill in the art without any creative effort based on these drawings.

[0032] Figure 1 is a step flow chart of a road network generation method of an embodiment of the present disclosure.

[0033] Figure 2 is a flow chart of a road network generation method of an embodiment of the present disclosure.

[0034] Figure 3 is a structural block diagram of a road network generation device of an embodiment of the present disclosure. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present disclosure will be described clearly and completely in the following with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only some of the embodiments of the present disclosure, not all the embodiments of the present disclosure. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without any creative effort belong to the scope of protection of the present disclosure.

[0036] Generally, the road network generation model includes an encoding network and a decoding network. The encoding network is used to encode the trajectory features, learn the corresponding encoding features, and then the decoding network is used to decode the encoding features to obtain the road segmentation result.

[0037] However, the existing road network generation model cannot guarantee sufficient field of view to learn the topology of the road network, resulting in poor connectivity of the segmented road network and insufficient multi-lane resolution. The non-linear capability of the convolutional neural network increases with the increase of the network layers, and another factor that determines the learning capability of the network is the field of view of the network. The convolution operation has a fixed field of view, and its translational invariance is also because the convolution has a certain field of view. In the trajectory generation road network task, the road has the characteristics of narrowness, large span, connectivity, complexity, etc. Therefore, the feature with a larger field of view has a certain sensitivity to the position information of the road, and can learn a certain road network topology. At the same time, it is also necessary to maintain the resolution of the feature, so as not to lack the abstraction ability of the narrow road in the high-level feature.

[0038] Therefore, in the embodiments of the present disclosure, a multi-scale dilated convolution network is added between the encoding network and the decoding network in the road network generation model, the receptive field is expanded through dilated convolution, and the multi-scale features are integrated in the center part to adapt to the narrowness, connectivity, complexity and large span of the road network, etc.

[0039] Referring to Figure 1 , a step flowchart of a road network generation method of an embodiment of the present disclosure is shown.

[0040] As Figure 1 shown, the road network generation method can include the following steps:

[0041] Step 101, feature extraction is performed on trajectories contained in a grid in a to-be-processed region, to obtain trajectory features of the grid.

[0042] The to-be-processed region can refer to a region for which road network data needs to be generated. Optionally, the coverage of the to-be-processed region can include but is not limited to: a province, a city, a district, a street, a county, a township, etc.

[0043] A large number of trajectories contained in the to-be-processed region are obtained. Optionally, the trajectories can include but are not limited to at least one of the following: a deliveryman's riding trajectory, a courier's riding trajectory, a ride-hailing vehicle's driving trajectory, a bus's driving trajectory, a private car's driving trajectory, etc.

[0044] Grid-based trajectory feature extraction is performed on the to-be-processed region, to convert vector data into a grid image. The to-be-processed region is divided into a plurality of grids, and for each grid, feature extraction is performed on trajectories contained in the grid, to obtain trajectory features of the grid. The range covered by one grid is not limited in the embodiment of the present disclosure, for example, the range covered by one grid can be 1 square kilometer, 2 square kilometers, etc.

[0045] Optionally, the trajectory features of the grid can include but are not limited to at least one of the following: point density, line density, average speed, direction features of eight-directional projections of speed, and neighborhood features of eight directions. The eight directions are north, northwest, west, southwest, south, southeast, east, and northeast, respectively. The neighborhood features refer to whether there is a trajectory in eight neighborhood directions centered on the pixel (i.e., the grid), and the value is 0 or 1.

[0046] Step 102, the trajectory features of the grid are input into a pre-trained road network generation model, and an encoding operation, a multi-scale dilated convolution operation, and a decoding operation are sequentially performed on the trajectory features, to obtain road segmentation information corresponding to the grid output by the road network generation model.

[0047] The road network generation model in the embodiment of the present disclosure can include an encoding network, a multi-scale dilated convolution network, and a decoding network connected in sequence.

[0048] For each grid, the trajectory feature of the grid is input into a pre-trained road network generation model, and in the road network generation model, an encoding network, a multi-scale hollow convolution network and a decoding network connected in sequence are used to sequentially perform an encoding operation, a multi-scale hollow convolution operation and a decoding operation on the trajectory feature of the grid, to obtain road segmentation information corresponding to the grid output by the road network generation model. The road segmentation information can represent which part of the grid is a road and which part is not a road.

[0049] In step 103, the road network data corresponding to the to-be-processed region is generated based on the road segmentation information.

[0050] After obtaining the road segmentation information corresponding to each grid, the road network data corresponding to the to-be-processed region can be generated based on the road segmentation information corresponding to these grids. The specific process will be described in detail in the embodiments below.

[0051] In the embodiments of the present disclosure, the encoding operation, the multi-scale hollow convolution operation and the decoding operation are sequentially performed on the trajectory feature in the road network generation model. The field of view can be expanded to the entire input feature through the hollow convolution operation, and the multi-scale features can be integrated in the center part, so as to adapt to the narrowness, complexity, connectivity and large span of the road network, solve the problems of poor connectivity of road network data and difficulty in separating multi-lane, and improve the accuracy of road network data.

[0052] Generally, the trajectory feature includes 19 channels of features such as point density, line density, average speed, eight-directional projection direction features, and eight-directional neighborhood features. However, through analysis of the trajectory feature, it can be found that the features of different channels are highly correlated, which can easily lead to feature redundancy. In addition, due to the uneven quality of trajectories, the quality of multi-channel features generated from original GPS trajectories is also uneven. For example, the instantaneous speed and direction error of GPS trajectory points are large, which can easily introduce noise interference under the condition of feature redundancy, increase the storage pressure, and reduce the anti-interference ability of the model. In addition, generally, the road segmentation information output by the road network generation model can have problems of multi-lane adhesion and poor road connectivity to a certain extent.

[0053] In this embodiment, to address the feature redundancy problem, the extracted trajectory features are preprocessed, and an attention mechanism is added at the channel level to rank and filter features based on their importance, selecting typical features to reduce redundancy. To address connectivity and multi-lane separation issues, a multi-scale dilated convolution is introduced between the encoding and decoding parts to expand the field of view to the entire feature map without reducing feature resolution, adapting to the narrowness, connectivity, complexity, and long span characteristics of the road network. Through a multi-task learning approach, road surface information is added as an auxiliary information flow, fusing encoded information and road surface information during road centerline extraction to simultaneously generate both the road surface and the road centerline, thus alleviating the difficulty of high-dimensional feature encoding caused by the narrowness of the centerline. To address the issues of multi-lane congestion and missing intersection connections, a transfer learning approach is employed. Large-scale initial labels are generated using clustering for pre-training, and active learning-based data cleaning and labeling are used to fine-tune the road network generation model, thereby improving these problems. Furthermore, a neural network-based post-processing optimization method is proposed to replace the rule-based topology optimization, allowing the generated road network to extend along the direction of the original trajectory, further enhancing the connectivity of the road network.

[0054] Reference Figure 2 The diagram shows a flowchart of a road network generation method according to an embodiment of the present disclosure.

[0055] like Figure 2 As shown, the road network generation method may include the following steps:

[0056] Step 201: Obtain the trajectory.

[0057] For the areas that need to generate road network data, acquire massive amounts of trajectories within the areas to be processed.

[0058] Step 202, trajectory preprocessing.

[0059] Optionally, trajectory preprocessing may include, but is not limited to: removing stray lines based on speed and distance detection, smoothing the trajectory using Kalman filtering, filtering cluttered trajectories within buildings, removing dwell points based on speed, and breaking the trajectory based on the length of trajectory points, etc. The specific process of trajectory preprocessing can be implemented based on practical experience, and will not be discussed in detail here.

[0060] Step 203: Rasterize and extract features.

[0061] In this embodiment, to address the feature redundancy problem, channels can be pre-selected based on an attention mechanism. Figure 2 (The process within the dashed box) involves selecting important target channels from a preset multi-channel array (such as the 19 channels mentioned above) so that feature extraction can be performed based on the channel selection results during the rasterization feature extraction process.

[0062] Optionally, the target channel is screened by the following steps A1-A2:

[0063] Step A1, first training a first initial model containing an attention network by using first sample data.

[0064] The structure of the first initial model can adopt any applicable image segmentation network structure, which can include but is not limited to FCN (Fully Convolutional Networks), FPN (Feature Pyramid Networks), UNet (U-Networks), etc.

[0065] In the first initial model, a first layer embeds an attention network structure, so as to add an attention mechanism at the channel level to the above-mentioned multiple channels (such as the above-mentioned 19 channels), respectively assign weights to the above-mentioned multiple channels in the attention network, and set the channel compression ratio of the attention network to 1. The attention network can adopt any applicable attention mechanism network structure, which can include but is not limited to SENet (Squeeze-and-Excitation Networks), GAT (Graph Attention Network), etc.

[0066] A large number of trajectories contained in a first sample area are obtained. The first sample area can be any area with initial road network data (i.e., road network base map).

[0067] The first sample area is rasterized to extract trajectory features, and vector data is converted into raster images. The first sample area is divided into multiple grids, and the grids in the first sample area are taken as first sample grids. For each first sample grid, the trajectories contained in the first sample grid are extracted to obtain the multi-channel trajectory features (such as the above-mentioned 19 channels) of the first sample grid. For each first sample grid, according to the initial road network data corresponding to the first sample area, it can be determined whether the first sample grid contains road network data, and the first sample grid is set with a first label indicating whether the first sample grid actually contains road network data. Based on this, the first sample data is constructed, which includes the multi-channel trajectory features of the first sample grids in the first sample area and the first labels of the first sample grids.

[0068] In the process of first training the first initial model by using the first sample data, the multi-channel track features of the first sample grid are taken as the input of the first initial model, and the road network segmentation result of the first sample grid output by the first initial model is obtained. The road network segmentation result indicates whether the first sample grid contains road network data.

[0069] Based on the road network segmentation result of the first sample grid and the first label of the first sample grid, the loss function of the first training can be calculated. When the loss function of the first training reaches the convergence condition, it is determined that the first training is completed. Optionally, the loss function of the first training can be selected from any suitable loss function. For example, cross-entropy loss function, exponential loss function, Dice loss function, intersection over union loss function, etc.

[0070] Step A2, after the first training is completed, the weight proportion of each channel in the attention network is counted, and the channel with cumulative weight proportion greater than a preset threshold is selected as the target channel.

[0071] The weight parameters of the added attention mechanism of the input channel are visualized. After the first training is completed, the weight proportion of each channel in the attention network in all channels is counted. The weight proportion of any one channel in all channels can be the ratio between the weight of the channel and the total weight of all channels. Then, the channels are sorted in descending order according to the weight proportion, and the cumulative weight proportion is obtained in turn from the first channel (i.e. the channel with the highest weight proportion). It is determined whether the cumulative weight proportion is greater than a preset threshold. Finally, the least number of channels with cumulative weight proportion greater than the preset threshold are selected as the target channels.

[0072] For the preset threshold, any suitable value can be set according to actual experience. For example, the preset threshold is 95%, the cumulative weight proportion of the first channel obtained for the first time is 80%, which is less than 95%; the cumulative weight proportion of the first channel and the second channel obtained for the second time (i.e. the sum of the weight proportions of the first channel and the second channel) is 90%, which is less than 95%; the cumulative weight proportion of the first channel, the second channel and the third channel obtained for the third time (i.e. the sum of the weight proportions of the first channel, the second channel and the third channel) is 96%, which is greater than 95%, so the first channel, the second channel and the third channel can be selected as the target channels.

[0073] For example, for the above 19 channels, the first three channels can be selected as the target channels after the above processing process, so the target channels are the point density channel, the line density channel and the average speed channel.

[0074] After rasterization, the to-be-processed region is divided into a plurality of grids. When performing feature extraction on the trajectories contained in any one grid, the feature extraction is performed based on a target channel pre-selected from the preset plurality of channels, and the trajectory features of the target channel are extracted as the trajectory features corresponding to the grid, so as to overcome feature redundancy and relieve memory pressure.

[0075] In an optional embodiment, the trajectory features of the plurality of channels (such as the 19 channels described above) can be extracted from the trajectories contained in the grid, specifically, the point density, line density, average speed, direction features of eight direction projections of speed, and eight direction neighborhood features are extracted from the original GPS trajectories, where the eight directions are north, northwest, west, southwest, south, southeast, east, and northeast, and the neighborhood features refer to whether there is a trajectory in the eight neighborhood directions centered on the grid (that is, the pixel), and the value is 0 or 1. Then, the trajectory features of the target channel are selected from the trajectory features of the plurality of channels, and the trajectory features of the target channel are taken as the trajectory features corresponding to the grid.

[0076] In another optional embodiment, the trajectory features of the plurality of channels (such as the 19 channels described above) can not be extracted, but the trajectory features of the target channel can be directly extracted from the trajectories contained in the grid, and the trajectory features of the target channel are taken as the trajectory features corresponding to the grid.

[0077] Step 204, the road network generation model predicts the road segmentation information.

[0078] The road network generation model for generating road network data based on trajectories is pre-trained.

[0079] Optionally, the road network generation model is trained by the following steps B1-B3:

[0080] Step B1, clustering the trajectories in the second sample region to obtain clustering clusters, selecting a representative trajectory from each of the clustering clusters, and taking the grid in the second sample region that is successfully bound to the representative trajectory as a second sample grid.

[0081] A large number of trajectories contained in a second sample region are obtained. The second sample region can be any region with initial road network data (that is, road network base map), and the second sample region can be the same as or different from the first sample region described above.

[0082] The trajectories in the second sample region are clustered to obtain a plurality of cluster groups. Optionally, the clustering manner can include, but is not limited to, density clustering, k-means clustering, etc. The density clustering can adopt a DBSCAN (Density-Based Spatial Clustering of Applications with Noise) manner, etc.

[0083] For each cluster group, a representative trajectory is selected from the cluster group according to a preset trajectory selection rule. Optionally, the trajectory selection rule can include, but is not limited to, randomly selecting a trajectory, selecting a trajectory of a core point, etc.

[0084] The second sample region is divided into a plurality of grids. For each grid in the second sample region, the grid is matched with the representative trajectory selected from the cluster groups, and the grid in the second sample region that is successfully matched with the representative trajectory is taken as a second sample grid. The process of the matching can be that the representative trajectory is matched with a road in the initial road network data corresponding to the grid. Optionally, the matching manner can include, but is not limited to, model matching, distance calculation matching, etc.

[0085] Step B2, the second initial model is second trained by using the second sample data.

[0086] The structure of the second initial model can adopt any applicable image segmentation network structure. Optionally, the second initial model in the embodiment can adopt a neural network model based on multi-task learning to perform two tasks of road surface segmentation and road centerline segmentation. For example, the second initial model can adopt a T2RNet (Track to Road Networks) or the like.

[0087] The second initial model can include an encoding network, a multi-scale hollow convolution network and a decoding network connected in sequence. The parameters of the encoding network are shared by the two tasks, the parameters of the decoding network are independent for the two tasks, and the basic framework is a UNet. In the road surface segmentation task, the input of each decoding network layer includes the output of the previous decoding network layer and the output of the corresponding encoding network layer. In the centerline segmentation task, the input of each decoding network layer includes the output of the previous decoding network layer and the output of the corresponding encoding network layer, and further includes the output of the corresponding decoding network layer in the road surface segmentation task. Optionally, the multi-scale hollow convolution network can select the middle part of a D-LinkNet.

[0088] For each second sample grid, feature extraction is performed on the trajectories contained in the second sample grid to obtain trajectory features of the second sample grid, and the trajectory features of the second sample grid can also be normalized. For each second sample grid, based on the initial road network data corresponding to the second sample region, actual road segmentation information of the second sample grid can be obtained, and a second label is set for the second sample grid, which indicates the actual road segmentation information of the second sample grid. Based on this, the second sample data is constructed, which includes the trajectory features of the second sample grid and the second label of the second sample grid.

[0089] In the process of performing the second training on the second initial model by using the second sample data, the trajectory features of the second sample grid are taken as the input of the second initial model, and the encoding network of the second initial model is used for encoding operation and feature learning (where the input size can be 1024*1024*3). The features learned by the encoding network are input into the multi-scale hollow convolution network, the parallel structure of the hollow convolution step lengths of 1, 2, 4, 8, and 16 is integrated, and multi-scale hollow convolution operation and feature learning are performed to widen the field of view of the network to the entire input feature, so as to capture the topological properties such as the position and connectivity of the road network. The features learned by the multi-scale hollow convolution network are input into the decoding network for decoding operation and feature learning, and are decoded into predicted road segmentation information. The second initial model outputs the predicted road segmentation information of the second sample grid.

[0090] Based on the predicted road segmentation information of the second sample grid and the second label of the second sample grid, a loss function of the second training can be calculated, and when the loss function of the second training reaches a convergence condition, it is determined that the second training is completed. Optionally, the loss function of the second training can use any suitable loss function. For example, the loss function of the second training can be the weighted sum of the Dice loss function and the cross-entropy loss function, etc.

[0091] Optionally, the trajectory features of the second sample grid can include the above-mentioned multi-channel trajectory features, or the trajectory features of the second sample grid can include the trajectory features of the selected target channel. The actual road segmentation information of the second sample grid can include the actual road surface segmentation information and the actual road center line segmentation information of the second sample grid. The predicted road segmentation information of the second sample grid can include the predicted road surface segmentation information and the predicted road center line segmentation information of the second sample grid.

[0092] Step B3, after the second training is completed, the model of the second training is completed is taken as the road network generation model.

[0093] In an optional embodiment, after the second training is completed, the model trained by the second training can not be used as the road network generation model temporarily, and a fine-tuning sample based on active learning is obtained by labeling, and the model trained by the second training is further trained by using the fine-tuning sample. The purpose of labeling is to identify internal roads and multi-lane roads, so that the network has a sensitive response to relatively sparse trajectories, while retaining strong noise suppression capability.

[0094] Therefore, after the second training is completed, the following steps C1-C3 can be performed:

[0095] Step C1, using the model trained by the second training to predict the grid in the third sample area, and obtaining the grid predicted to contain roads but actually not containing roads as the third sample grid.

[0096] Obtain the trajectories contained in the third sample area. The third sample area can be any area with initial road network data (i.e. road network base map), and the third sample area is different from the second sample area.

[0097] Divide the third sample area into a plurality of grids. For each grid in the third sample area, the trajectory feature of the grid is obtained by extracting the trajectory feature contained in the grid, and the trajectory feature of the grid can also be normalized. The trajectory feature of the grid is input into the model trained by the second training to obtain the predicted road segmentation information of the grid output by the model trained by the second training. According to the predicted road segmentation information of the grid and the initial road network data corresponding to the grid, it can be determined whether the grid is a grid predicted to contain roads but actually not containing roads. The grid in the third sample area predicted to contain roads but actually not containing roads is taken as the third sample grid.

[0098] Step C2, using the third sample data to perform third training on the model trained by the second training.

[0099] For each third sample grid, the trajectory feature of the third sample grid is obtained by extracting the trajectory feature contained in the third sample grid, and the trajectory feature of the third sample grid can also be normalized. For each third sample grid, according to the initial road network data corresponding to the third sample area, the actual road segmentation information of the third sample grid can be obtained, the third sample grid is set with a third label, and the third label indicates the actual road segmentation information of the third sample grid. Based on this, the third sample data is constructed as a fine-tuning sample, and the third sample data includes the trajectory feature of the third sample grid and the third label of the third sample grid.

[0100] The second trained model is third trained by using third sample data, and the third training process is basically the same as the above-mentioned second training process, and specific reference can be made to the related description of step B2 above, and detailed description is not given here.

[0101] Optionally, the track features of the third sample grid can include the track features of the multiple channels described above, or the track features of the third sample grid can include the track features of the target channel screened out described above. The actual road segmentation information of the third sample grid can include actual road surface segmentation information and actual road center line segmentation information of the third sample grid. The predicted road segmentation information of the third sample grid can include predicted road surface segmentation information and predicted road center line segmentation information of the third sample grid.

[0102] Step C3, after the third training is completed, the third trained model is used as the road network generation model.

[0103] Based on the trained road network generation model, the track features of the grid contained in the to-be-processed region obtained in step 203 are normalized and input into the road network generation model. The track features of the grid are sequentially executed in the road network generation model. The encoding operation, the multi-scale hollow convolution operation and the decoding operation are performed to obtain the road segmentation information corresponding to the grid output by the road network generation model.

[0104] Optionally, the track features of the grid can include the track features of the multiple channels described above, or the track features of the grid can include the track features of the target channel screened out described above. The road segmentation information of the grid can include road surface segmentation information and road center line segmentation information of the grid.

[0105] In the embodiment, the road network generation module is based on a multi-task learning framework, multi-scale hollow convolution is added to expand the field of view to the entire feature map; the grid matched with the base map road network after track clustering is used as a pre-trained label to perform a large-scale pre-training process; then, a small-scale data based on active learning is used for fine-tuning. Since the segmentation mode of internal road sparse tracks is similar to the segmentation mode of sparse tracks at intersections, the fine-tuned model has greatly improved the connection of intersections and multi-lane recall.

[0106] Step 205, the road network optimization model optimizes the road segmentation information.

[0107] Since the road segmentation information obtained in step 204 still has a certain degree of multi-lane adhesion and poor road connectivity, the embodiment proposes a deep learning post-processing optimization method combined with original tracks.

[0108] A road network optimization model for optimizing road segmentation information is pre-trained.

[0109] Optionally, the road network optimization model is trained through steps D1-D3 as follows:

[0110] In step D1, a fourth sample grid is obtained, and the road network data in the fourth sample grid is eroded and multi-lane expanded to obtain the changed road segmentation information of the fourth sample grid.

[0111] Optionally, the fourth sample grid can include the grid filtered from the fourth sample area according to step B1 above and the grid filtered from the fifth sample area according to step C1 above; or, the fourth sample grid can include the second sample grid and the third sample grid above. The fourth sample area and the fifth sample area can be any area with initial road network data (i.e., road network base map).

[0112] For each fourth sample grid, the initial road network data in the fourth sample grid is randomly eroded and randomly expanded in the multi-lane part to realize the road network sample with road network disconnection and multi-lane adhesion, so as to learn the multi-lane separation mode and the road connectivity mode. The fourth sample grid after erosion and multi-lane expansion is subjected to feature extraction to obtain the changed trajectory feature of the fourth sample grid. The changed trajectory feature of the fourth sample grid can also be normalized. The changed trajectory feature of the fourth sample grid is input into the road network generation model trained above to obtain the changed road segmentation information of the fourth sample grid output by the road network generation model.

[0113] In step D2, the third initial model is fourth trained using the fourth sample data.

[0114] For each fourth sample grid, the trajectory contained in the fourth sample grid is subjected to feature extraction to obtain the trajectory feature of the fourth sample grid. The trajectory feature of the fourth sample grid can also be normalized. For each fourth sample grid, the actual road segmentation information of the fourth sample grid can be obtained according to the initial road network data corresponding to the fourth sample area. The fourth sample grid is set with a fourth label indicating the actual road segmentation information of the fourth sample grid.

[0115] The fourth sample data includes the changed road segmentation information of the fourth sample grid and the trajectory feature of the fourth sample grid, and the fourth label of the fourth sample grid.

[0116] The structure of the third initial model can adopt any applicable image segmentation network structure, which can include but is not limited to FCN, FPN, UNet, etc.

[0117] In the process of the fourth training of the third initial model by using the fourth sample data, the changed road segmentation information (single channel) of the fourth sample grid and the track features of the fourth sample grid are superimposed in the channel to be used as the input of the third initial model, and the prediction road segmentation information of the fourth sample grid output by the third initial model is obtained. Based on the prediction road segmentation information of the fourth sample grid and the second label of the fourth sample grid, the loss function of the fourth training can be calculated, and when the loss function of the fourth training reaches the convergence condition, the fourth training is determined to be completed. Optionally, the loss function of the fourth training can be selected from any suitable loss function. For example, the Dice loss function, the cross-entropy loss function, etc.

[0118] Optionally, the changed track features of the fourth sample grid can include the above-mentioned multi-channel track features, or include the above-mentioned filtered track features of the target channel. The changed road segmentation information of the fourth sample grid can include the changed road surface segmentation information and the changed road center line segmentation information of the fourth sample grid. The track features of the fourth sample grid can include the above-mentioned multi-channel track features, or include the above-mentioned filtered track features of the target channel. The actual road segmentation information of the fourth sample grid can include the actual road surface segmentation information and the actual road center line segmentation information of the fourth sample grid. The prediction road segmentation information of the fourth sample grid can include the prediction road surface segmentation information and the prediction road center line segmentation information of the fourth sample grid.

[0119] Step D3, after the fourth training is completed, the model of the fourth training is completed is used as the road network optimization model.

[0120] Optionally, some sample data can be used to further test the road network optimization model, so as to further improve the connectivity and multi-lane resolution of the road network.

[0121] Based on the trained road network optimization model, for the road segmentation information corresponding to the grid contained in the to-be-processed region obtained in step 204, the road segmentation information corresponding to the grid and the track features of the grid are input into the pre-trained road network optimization model, and the optimized road segmentation information corresponding to the grid output by the road network optimization model is obtained. Then, based on the optimized road segmentation information, the road network data corresponding to the to-be-processed region is generated.

[0122] Step 206, vectorization.

[0123] The optimized road segmentation information obtained in step 205 is subjected to skeleton extraction, symbol edge optimization, vectorization and other processing. The specific process will not be discussed in detail here.

[0124] Step 207, generating a road network.

[0125] The vectorized results from step 206 are then subjected to Douglas thinning, short road cleaning, and other processes to generate the final vectorized road network map. The specific process will not be discussed in detail here.

[0126] In this embodiment, to address issues such as feature redundancy, poor road connectivity, and low parallel road resolution in trajectory-generated road networks, an attention mechanism is used to filter redundant features, expand the model's field of view to improve road connectivity, and transfer learning is used to fine-tune the network by labeling specific patterns. Post-processing of the original trajectory further optimizes parallel road separation and road connectivity. This deep learning-based trajectory generation method eliminates the need for manually setting empirical parameters and leverages prior knowledge of existing road networks. Compared to traditional machine learning methods, it is more flexible, has better generalization performance, and compresses the number of input feature channels, saving feature generation time and storage costs. Improving road network connectivity and multi-lane resolution from both the segmentation and post-processing models has significant application value in industry. Iterating the road network by updating trajectory data automates production and facilitates maintenance.

[0127] Reference Figure 3 The diagram shows a structural block diagram of a road network generation device according to an embodiment of the present disclosure.

[0128] like Figure 3 As shown, the road network generation device may include the following modules:

[0129] The extraction module 301 is used to extract features from the trajectories contained in the grid within the processing area to obtain the trajectory features of the grid.

[0130] The prediction module 302 is used to input the trajectory features of the grid into a pre-trained road network generation model, and sequentially perform encoding, multi-scale dilated convolution and decoding operations on the trajectory features to obtain the road segmentation information corresponding to the grid output by the road network generation model.

[0131] The generation module 303 is used to generate road network data corresponding to the area to be processed based on the road segmentation information.

[0132] Optionally, the trajectory features include the trajectory features of a target channel pre-selected from a preset multi-channel, wherein the cumulative weight ratio of the target channel in the multi-channel is greater than a preset threshold.

[0133] Optionally, the target channel is screened through the following modules: a first training module, configured to perform first training on a first initial model containing an attention network by using first sample data; the first sample data includes multi-channel trajectory features of a first sample grid in a first sample region and a first label of the first sample grid, the first label indicating whether the first sample grid actually contains road network data; the attention network assigns weights to the multi-channels respectively; and a first determination module, configured to, after the first training is completed, count weight proportions of the channels in the attention network, and select a channel with a cumulative weight proportion greater than a preset threshold as the target channel.

[0134] Optionally, the road segmentation information includes road surface segmentation information and road centerline segmentation information.

[0135] Optionally, the road network generation model is trained through the following modules: a first acquisition module, configured to cluster trajectories in a second sample region to obtain clustering clusters, select representative trajectories from the clustering clusters, and select a grid in the second sample region as a second sample grid if the grid is successfully bound to the representative trajectories for road matching; a second training module, configured to perform second training on a second initial model containing an encoding network, a multi-scale hollow convolution network and a decoding network connected in sequence by using second sample data; the second sample data includes trajectory features of the second sample grid and a second label of the second sample grid, the second label indicating actual road segmentation information of the second sample grid; and a second determination module, configured to, after the second training is completed, take a model trained by the second training as the road network generation model.

[0136] Optionally, the device further includes: a second acquisition module, configured to predict a grid in a third sample region by using the model trained by the second training, and acquire a grid predicted to contain a road but actually not containing a road as a third sample grid; a third training module, configured to perform third training on the model trained by the second training by using third sample data; the third sample data includes trajectory features of the third sample grid and a third label of the third sample grid, the third label indicating actual road segmentation information of the third sample grid; and the second determination module, specifically configured to, after the third training is completed, take a model trained by the third training as the road network generation model.

[0137] Optionally, the device further includes: an optimization module, configured to input road segmentation information corresponding to the grid and trajectory features of the grid into a pre-trained road network optimization model to obtain optimized road segmentation information of the grid output by the road network optimization model; and the generation module 303, specifically configured to generate road network data corresponding to the to-be-processed region based on the optimized road segmentation information.

[0138] Optionally, the road network optimization model is trained by a third obtaining module, configured to obtain a fourth sample grid, perform erosion and multi-lane dilation on road network data in the fourth sample grid, and obtain change road segmentation information of the fourth sample grid; a fourth training module, configured to perform fourth training on a third initial model by using fourth sample data; the fourth sample data includes the change road segmentation information of the fourth sample grid and trajectory features of the fourth sample grid, and a fourth label of the fourth sample grid, the fourth label indicating actual road segmentation information of the fourth sample grid; and a third determining module, configured to take a model after the fourth training is completed as the road network optimization model.

[0139] In the embodiments of the present disclosure, the encoding operation, the multi-scale hollow convolution operation and the decoding operation are sequentially performed on the trajectory features in the road network generation model. The field of view can be expanded to the entire input feature by the hollow convolution operation, and the multi-scale features are integrated in the center part, so as to adapt to the narrowness, complexity, connectivity and large span of the road network, solve the problems of poor connectivity of road network data and difficulty in separating multi-lane, and improve the accuracy of road network data.

[0140] For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0141] In the embodiments of the present disclosure, an electronic device is also provided. The electronic device can include one or more processors, and one or more computer readable storage media having stored thereon instructions, e.g., application programs. When the instructions are executed by the one or more processors, the processors perform the road network generation method of any one of the above embodiments.

[0142] In the embodiments of the present disclosure, a non-transitory computer readable storage medium is also provided, having stored thereon a computer program executable by a processor of an electronic device. When the computer program is executed by the processor, the processor performs the road network generation method of any one of the above embodiments.

[0143] The aforementioned processor can be a general processor, and can include, but is not limited to, a Central Processing Unit (CPU), a Network Processor (NP), a Digital Signal Processing (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component, and the like.

[0144] The aforementioned computer readable storage medium can include, but is not limited to, a Read Only Memory (ROM), a Random Access Memory (RAM), a Compact Disc Read Only Memory (CD-ROM), an Electronic Erasable Programmable Read Only Memory (EEPROM), a hard disk, a floppy disk, a flash memory, and the like.

[0145] The algorithms and displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems can be used with these teachings, based upon the description as provided herein. Those skilled in the art will recognize that the structures described above can be combined or divided in a manner not specifically described above, and that the present disclosure can be implemented using a variety of programming languages. The specific language chosen to describe the present disclosure is not intended to limit the scope of the present disclosure.

[0146] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the present disclosure can be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been described in detail in order not to obscure the understanding of this description.

[0147] Similarly, it is to be understood that the embodiments of the disclosure can sometimes alternately be termed as a system, a method, or an apparatus, depending on the context in which the embodiments of the disclosure are used. Similarly, it is to be understood that, for the purpose of brevity and clarity, in the detailed description of the exemplary embodiments of the disclosure above, various features of the embodiments of the disclosure are sometimes grouped together in a single embodiment, figure, or description of related features. However, this method of disclosure should not be construed to reflect an intention that the embodiments of the disclosure require more features than are explicitly recited in each claim. Rather, inventive aspects lie in less than all features of a single disclosed embodiment. Thus, the following claims are hereby expressly incorporated into this detailed description, with each claim acting as a separate embodiment of the embodiments of the disclosure. The claims, as set forth below, are intended to encompass the embodiments of the disclosure in their various forms. The following claims are intended to be construed to include one or more embodiments of the disclosure. The following claims are intended to include one or more embodiments of the disclosure alone or in combination with one or more other claims. The following claims are intended to include one or more embodiments of the disclosure alone or in combination with one or more other claims. The claims, as set forth below, are intended to encompass the embodiments of the disclosure in their various forms.

[0148] Those skilled in the art will appreciate that the modules in the apparatus of the embodiments can be adapted and arranged in one or more apparatuses other than the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and further can be divided into multiple sub-modules or sub-units or sub-components. Except that at least some of such features and / or processes or units are mutually exclusive, all combinations of all features disclosed in the specification (including the accompanying claims, abstract and drawings), and all processes or units of any methods or apparatuses so disclosed can be used in combination with one another. Unless explicitly stated otherwise, each feature disclosed in the specification (including the accompanying claims, abstract and drawings) can be replaced by alternative features providing the same, equivalent or similar functionality.

[0149] The various component embodiments of the embodiments of the disclosure can be implemented in hardware, or as software modules running in one or more processors, or in combinations thereof. Those skilled in the art will appreciate that a microprocessor or a digital signal processor (DSP) can be used in practice to implement some or all of the functions of some or all of the components in the dynamic picture generation apparatus according to the embodiments of the disclosure. The embodiments of the disclosure can also be implemented as a program for executing part or all of the methods described herein on a device or an apparatus. Such a program implementing the embodiments of the disclosure can be stored on a computer readable medium, or can be in the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or in any other form.

[0150] It should be noted that the above-mentioned embodiments illustrate rather than limit the embodiments of the present disclosure, and that one skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word 'comprising' does not exclude the presence of elements or steps not listed in a claim. The word 'a' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The embodiments of the present disclosure can be implemented by means of both hardware and software, and the embodiments of the present disclosure can be implemented by means of a suitably programmed computer. In a unit claim, several devices can be listed with a comma. The use of 'a' or 'an' preceding the citation of an element does not exclude the presence of a plurality of such elements, nor does it imply that a single element is preferred. Other expressions of the sole or single element also cover the implementation of a plurality of elements notwithstanding the variable used to describe the quantity. The word 'first','second', 'third', etc. do not imply any order or sequence unless otherwise specified. These words are to be construed as names.

[0151] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0152] The above describes only specific implementations of the embodiments of the present disclosure, but the protection scope of the embodiments of the present disclosure is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the embodiments of the present disclosure, which should be covered within the protection scope of the embodiments of the present disclosure.

Claims

1. A road network generation method characterized by comprising: The method comprises the following steps: feature extraction is performed on trajectories contained in a grid in a to-be-processed region to obtain trajectory features of the grid; the trajectory features of the grid are input into a pre-trained road network generation model, and the trajectory features are sequentially subjected to an encoding operation, a multi-scale hollow convolution operation and a decoding operation to obtain road segmentation information corresponding to the grid output by the road network generation model; road network data corresponding to the to-be-processed region is generated based on the road segmentation information; the road network generation model is trained in the following manner: trajectories in a second sample region are clustered to obtain clustering clusters, and representative trajectories are selected from each of the clustering clusters, and grids in the second sample region that are successfully bound to the representative trajectories by road matching are taken as second sample grids; a second initial model comprising an encoding network, a multi-scale hollow convolution network and a decoding network connected in sequence is subjected to second training by using second sample data; the second sample data comprises trajectory features of the second sample grids and second labels of the second sample grids, and the second labels indicate actual road segmentation information of the second sample grids; after the second training is completed, the model after the second training is completed is taken as the road network generation model; before the model after the second training is completed is taken as the road network generation model, the method further comprises the following steps: the model after the second training is completed is used to predict grids in a third sample region to obtain grids that are predicted to contain roads but actually do not contain roads as third sample grids; the model after the second training is completed is subjected to third training by using third sample data; the third sample data comprises trajectory features of the third sample grids and third labels of the third sample grids, and the third labels indicate actual road segmentation information of the third sample grids; after the third training is completed, the model after the third training is completed is taken as the road network generation model.

2. The method of claim 1, wherein, The trajectory features comprise trajectory features of a target channel preselected from a plurality of preset channels, and an accumulated weight proportion of the target channel in the plurality of channels is greater than a preset threshold.

3. The method of claim 2, wherein, The target channel is selected in the following manner: a first initial model comprising an attention network is subjected to first training by using first sample data; the first sample data comprises multi-channel trajectory features of first sample grids in a first sample region and first labels of the first sample grids, and the first labels indicate whether the first sample grids actually contain road network data; each channel in the attention network is assigned a weight; after the first training is completed, weight proportions of each channel in the attention network are counted, and a channel with an accumulated weight proportion greater than a preset threshold is selected as the target channel.

4. The method of claim 1, wherein, The road segmentation information comprises road surface segmentation information and road center line segmentation information.

5. The method according to any one of claims 1 to 4, characterized in that, Before generating the road network data corresponding to the to-be-processed region based on the road segmentation information, the method further includes: inputting the road segmentation information corresponding to the grid and the track feature of the grid into a pre-trained road network optimization model to obtain the optimized road segmentation information of the grid output by the road network optimization model; The road network data corresponding to the to-be-processed region is generated based on the road segmentation information, including: generating the road network data corresponding to the to-be-processed region based on the optimized road segmentation information.

6. The method of claim 5, wherein, The road network optimization model is trained in the following manner: A fourth sample grid is obtained, and the road network data in the fourth sample grid is eroded and expanded in multiple lanes to obtain changed road segmentation information of the fourth sample grid; The third initial model is trained using fourth sample data; the fourth sample data includes the changed road segmentation information of the fourth sample grid and the track feature of the fourth sample grid, and a fourth label of the fourth sample grid, the fourth label indicating actual road segmentation information of the fourth sample grid; After the fourth training is completed, the model after the fourth training is completed is used as the road network optimization model.

7. A road network generation apparatus characterized by comprising: It includes: The extraction module is configured to extract a track feature of a track in a grid in a to-be-processed region. The prediction module is configured to input the track feature of the grid into a pre-trained road network generation model, and sequentially perform an encoding operation, a multi-scale hollow convolution operation and a decoding operation on the track feature to obtain road segmentation information of the grid output by the road network generation model. The generation module is configured to generate road network data corresponding to the to-be-processed region based on the road segmentation information. The road network generation model is trained in the following manner: The tracks in the second sample region are clustered to obtain clustering clusters, and a representative track is selected from each of the clustering clusters, and a grid in the second sample region that is successfully bound to the representative track by road matching is used as a second sample grid. The second initial model including a sequentially connected encoding network, a multi-scale hollow convolution network and a decoding network is trained using second sample data; the second sample data includes a track feature of the second sample grid and a second label of the second sample grid, the second label indicating actual road segmentation information of the second sample grid. After the second training is completed, the model after the second training is completed is used as the road network generation model. Before the model after the second training is completed is used as the road network generation model, the method further includes: The grid in the third sample region is predicted using the model after the second training is completed, and a grid that is predicted to contain a road but actually does not contain a road is obtained as a third sample grid. The second trained model is trained using third sample data; the third sample data includes a track feature of the third sample grid and a third label of the third sample grid, the third label indicating actual road segmentation information of the third sample grid. The second trained model is taken as the road network generation model, comprising: taking the third trained model as the road network generation model after the third training is completed.

8. An electronic device, comprising: Comprising: one or more processors; and one or more computer readable storage media having stored thereon instructions; When the instructions are executed by the one or more processors, the processor executes the road network generation method as claimed in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program is stored on the one or more computer readable storage media, and when the computer program is executed by the processor, the processor executes the road network generation method as claimed in any one of claims 1 to 6.

Images