Model training method, road generation method, device, medium and equipment

Abstract

The present disclosure relates to a model training method, a road generation method, a device, a medium and equipment, which are beneficial to the generation of missing roads in a region. The method comprises: obtaining sample data for training; for each sample data, determining trajectory feature information corresponding to a sample region according to trajectory information in the sample region, and determining road label information corresponding to the sample region according to road information in the sample region; and training a neural network model by using the trajectory feature information and the road feature information to obtain a road generation model, wherein the neural network model comprises a semantic segmentation network and a language model construction network, the semantic segmentation network is used to output a first result for the trajectory feature information of the input sample data, the first result is used to determine whether each sub-region in the sample region of the input sample data is a road, and the language model construction network is used to output a predicted path for the trajectory feature information of the input sample data.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, in particular, to a model training method, a road generation method, an apparatus, a medium and an equipment. BACKGROUND

[0002] In instant delivery scenarios (for example, food delivery), navigation functions can provide routes for delivery personnel, help delivery personnel find delivery destinations, and improve delivery efficiency. However, with the development of the economy, the scale of areas such as residential areas, communities, commercial areas, and industrial parks gradually expands, and the distribution of buildings within the area becomes more and more complex, resulting in more and more complex road planning within the area, and the problem of missing road data within most areas further increases the complexity of road planning within the area. In this way, due to the lack of road information within the area, the navigation function for assisting delivery personnel can only help delivery personnel reach the entrance of such an area, and delivery usually needs to be face-to-face delivered to the user, so the delivery personnel need to spend a lot of time from entering the park to finding the location of the user, and the delivery efficiency is very low. SUMMARY

[0003] The purpose of the present disclosure is to provide a model training method, a road generation method, an apparatus, a medium and an equipment to obtain a road generation model with the ability to mine roads within an area, and then for an area with missing internal road data, the road generation model can be used to generate road data within the area, which is beneficial to improve the navigation function within the park.

[0004] In order to achieve the above-mentioned purpose, according to the first aspect of the present disclosure, a road generation model training method is provided, the method comprises:

[0005] Obtaining sample data for training, the sample data comprising trajectory information within a sample area and road information within the sample area;

[0006] For each sample data, determining trajectory feature information corresponding to the sample area according to the trajectory information within the sample area, and determining road label information corresponding to the sample area according to the road information within the sample area, wherein the trajectory feature information is used to represent the number characteristics of the trajectories of each sub-area within the sample area, and the road label information is used to represent whether each sub-area within the sample area is a road;

[0007] The trajectory feature information and the road feature information are used to train a neural network model to obtain the road generation model, wherein the neural network model comprises a semantic segmentation network and a language model construction network, the semantic segmentation network is used to output a first result for the trajectory feature information of input sample data, the first result is used to determine whether each sub-region in a sample region of the input sample data is a road, and the language model construction network is used to output a predicted path for the trajectory feature information of the input sample data.

[0008] Optionally, the trajectory feature information and the road feature information are used to train a neural network model to obtain the road generation model, comprising:

[0009] Target trajectory feature information and target road label information of target sample data used for this training are determined.

[0010] The target trajectory feature information is input into the semantic segmentation network of this training to obtain a second result output by the semantic segmentation network of this training, and the second result is used to determine whether each sub-region in a sample region of the target sample data is a road.

[0011] In a case where a training stop condition is not met, a first loss value is determined, wherein the first loss value is determined according to the second result and the target road label information.

[0012] According to the first loss value, the semantic segmentation network of this training is updated, and the updated semantic segmentation network is used for next training.

[0013] In a case where the training stop condition is met, the semantic segmentation network of this training is determined as the road generation model.

[0014] Optionally, the method further comprises:

[0015] The target trajectory feature information is input into the language model construction network of this training to obtain a target path output by the language model construction network of this training for the target trajectory feature information.

[0016] In a case where the training stop condition is not met, a reference path is generated according to the target road label information, and a second loss value is calculated according to the target path and the reference path.

[0017] According to the first loss value, the semantic segmentation network of this training is updated, and the updated semantic segmentation network is used for next training.

[0018] According to the first loss value and the second loss value, a comprehensive loss value is calculated, and the semantic segmentation network of this training is updated according to the comprehensive loss value.

[0019] Optionally, the method further comprises:

[0020] updating the language model construction network of the current training according to the comprehensive loss value, and using the updated language model construction network and the updated semantic segmentation network for the next training.

[0021] Optionally, the trajectory information in the sample region comprises trajectory points in the sample region and position information of the trajectory points.

[0022] The method further comprises:

[0023] dividing the sample region into a plurality of sub-regions of a preset size;

[0024] determining the number of trajectory points contained in each sub-region in the sample region according to the position information of the trajectory points in the sample region;

[0025] obtaining the trajectory feature information of the sample region according to the number of trajectory points contained in each sub-region in the sample region.

[0026] Optionally, the road information in the sample region comprises roads in the sample region and position information of the roads.

[0027] The method further comprises:

[0028] dividing the sample region into a plurality of sub-regions of a preset size;

[0029] determining whether each sub-region in the sample region is a road according to the position information of the roads in the sample region, to obtain the road label information of the sample region.

[0030] The preset size is set according to the size of the roads in the sample region.

[0031] According to a second aspect of the present disclosure, a road generation method is provided, the method comprising:

[0032] obtaining trajectory information in a target region to be processed;

[0033] determining trajectory feature information corresponding to the target region according to the trajectory information in the target region;

[0034] inputting the trajectory feature information corresponding to the target region into the road generation model trained by the method of the first aspect of the present disclosure, to obtain an output result of the road generation model, the output result being used to determine whether each sub-region in the target region is a road.

[0035] According to the output result and position information of each sub-region in the target region, position information of a road contained in the target region is determined.

[0036] According to a third aspect of the present disclosure, a road generation model training apparatus is provided, and the apparatus comprises:

[0037] A first obtaining module is configured to obtain sample data for training, wherein the sample data comprises trajectory information in a sample region and road information in the sample region.

[0038] A first determining module is configured to, for each sample data, determine trajectory feature information corresponding to the sample region according to the trajectory information in the sample region, and determine road label information corresponding to the sample region according to the road information in the sample region, wherein the trajectory feature information is used to represent quantity features of trajectories in each sub-region in the sample region, and the road label information is used to represent whether each sub-region in the sample region is a road.

[0039] A training module is configured to train a neural network model by using the trajectory feature information and the road feature information, so as to obtain the road generation model, wherein the neural network model comprises a semantic segmentation network and a language model construction network, the semantic segmentation network is used to output a first result for the trajectory feature information of input sample data, the first result is used to determine whether each sub-region in the sample region of the input sample data is a road, and the language model construction network is used to output a predicted path for the trajectory feature information of the input sample data.

[0040] According to a fourth aspect of the present disclosure, a road generation apparatus is provided, and the apparatus comprises:

[0041] A second obtaining module is configured to obtain trajectory information in a target region to be processed.

[0042] A second determining module is configured to determine trajectory feature information corresponding to the target region according to the trajectory information in the target region.

[0043] A processing module is configured to input the trajectory feature information corresponding to the target region into a road generation model trained by the road generation model training apparatus according to the third aspect of the present disclosure, so as to obtain an output result of the road generation model, and the output result is used to determine whether each sub-region in the target region is a road.

[0044] A third determining module is configured to determine position information of a road contained in the target region according to the output result and position information of each sub-region in the target region.

[0045] According to a fifth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in the first or second aspect of this disclosure.

[0046] According to a sixth aspect of this disclosure, an electronic device is provided, comprising:

[0047] A memory on which computer programs are stored;

[0048] A processor for executing the computer program in the memory to implement the steps of the method described in the first or second aspect of this disclosure.

[0049] The above technical solution acquires sample data for training. This sample data includes trajectory information and road information within the sample area. For each sample data point, trajectory feature information corresponding to the sample area is determined based on the trajectory information, and road label information is determined based on the road information. The trajectory feature information characterizes the quantity of trajectories in each sub-region of the sample area, while the road label information characterizes whether each sub-region is a road. The neural network model is trained using these trajectory and road feature information to obtain a road generation model. This neural network model includes a semantic segmentation network and a language model building network. The semantic segmentation network outputs a first result based on the trajectory feature information of the input sample data, determining whether each sub-region within the sample area of ​​the input sample data is a road. The language model building network outputs a predicted path based on the trajectory feature information of the input sample data. By training the neural network model based on the trajectory and road features within the area, a road generation model capable of generating road-related information is obtained. This model has the ability to discover roads within the area. Furthermore, for areas lacking internal road data, this model can generate road data within the area, thus improving the navigation function within the park.

[0050] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a road generation model training method according to one embodiment of the present disclosure;

[0053] Figure 2 This is a flowchart of a road generation method provided according to one embodiment of the present disclosure;

[0054] Figure 3 This is a block diagram of a road generation model training apparatus according to one embodiment of the present disclosure;

[0055] Figure 4 This is a block diagram of a road generation apparatus provided according to one embodiment of the present disclosure;

[0056] Figure 5 This is a block diagram illustrating an electronic device according to an exemplary embodiment;

[0057] Figure 6 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0058] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0059] As described in the background section, due to the lack of road information within the area, navigation route planning cannot assist delivery personnel in completing face-to-face delivery to users, resulting in low delivery efficiency. In related technologies, route discovery within a region is typically achieved based on trajectory point clustering. For example, routes are generated through KDE (Kernel Density Estimation) clustering. This primarily involves filtering out roads with fewer than two trajectories by tracing them together, and then calculating the tracing effect to filter out roads with poor tracing quality. However, this method of filtering roads based on thresholds lacks versatility, and its effectiveness is significantly reduced when road data is missing.

[0060] To address the aforementioned problems in related technologies, this disclosure provides a model training method, road generation method, apparatus, medium, and equipment to obtain a road generation model capable of excavating roads within a region. Furthermore, for regions lacking internal road data, the road generation model can generate road data within the region, which is beneficial for improving the navigation function within the park.

[0061] Figure 1 This is a flowchart of a road generation model training method provided according to one embodiment of this disclosure. For example... Figure 1 As shown, the method may include steps 11 to 13.

[0062] In step 11, sample data for training is obtained.

[0063] The sample data includes trajectory information and road information within the sample area.

[0064] Sample data can be obtained from a database containing a large amount of raw AOI (Area of ​​Interest) data. AOI data is used to characterize geographic areas; for example, AOI data can be stored in the database as a set of area boundary point locations (this set is used to identify an area). In addition, the database can also store trajectory information and road information. Trajectory information can be obtained by collecting trajectory points from historical delivery processes and can be stored in the database as GPS (Global Positioning System) coordinates. Road information can be obtained from road network information.

[0065] Based on the above content stored in the database, multiple sets of sample data can be obtained. Each set of sample data includes trajectory information within a sample area (i.e., the area identified by the AOI data) and road information within that sample data. The trajectory information within the sample area includes at least one trajectory point located within that sample area, and the road information within the sample area includes at least one road located within that sample area.

[0066] Furthermore, to ensure the quality of the acquired sample data and that the data used for training is high-quality data conducive to model training, the AOI data in the database can be filtered according to certain screening criteria when acquiring sample data. For example, AOI data that meets at least one of the following conditions can be used as sample data:

[0067] The shape of the area is normal;

[0068] There are roads within the area;

[0069] Trajectory points exist within the region;

[0070] The area of ​​the region is within a preset numerical range;

[0071] The area is passable;

[0072] The area is a non-sensitive area.

[0073] Sample data with similar shapes used in training also have similar data formats. Therefore, AOI data can be filtered by region shape and size to ensure that the obtained sample data have similar shapes and sizes, which is beneficial for model training. For determining whether a region shape is abnormal, abnormal shapes (e.g., triangles) can be pre-defined. The similarity between the shape of the region represented by the AOI data and the pre-defined abnormal shapes is used for judgment. For example, if the region represented by the AOI data has the same shape as one of the pre-defined abnormal shapes, then that AOI data will not be used as sample data for training. For determining the region area, a pre-defined numerical range representing the area can be used for filtering. For example, the preset numerical range can be [5000, 500000], in square meters.

[0074] The model training method disclosed herein aims to train a model capable of generating road information based on trajectory information. Therefore, the presence of trajectory points and roads in the training data will be beneficial for model training. The determination of whether roads and trajectory points exist within a region can be achieved by combining the location of the region represented by the AOI data with the locations of roads and trajectory points stored in the database.

[0075] The model training method disclosed herein is used to train a model capable of generating road information. The generated roads will be used for subsequent path planning; therefore, data on passable areas is more relevant for model training. Thus, AOI data can be filtered based on whether an area is passable or not, and whether it is a non-sensitive area. For example, non-sensitive areas could be, for instance, demilitarized zones.

[0076] Based on the above approach, sample data for training can be obtained from the AOI data, trajectory information, and road information already stored in the database and used for model training in subsequent steps.

[0077] In step 12, for each sample data, the trajectory feature information corresponding to the sample area is determined based on the trajectory information within the sample area, and the road label information corresponding to the sample area is determined based on the road information within the sample area.

[0078] Among them, trajectory feature information is used to characterize the quantity characteristics of trajectories in each sub-region within the sample area, and road label information is used to characterize whether each sub-region within the sample area is a road.

[0079] In one possible implementation, the trajectory information within the sample area includes the trajectory points within the sample area and the position information of each trajectory point. Accordingly, step 12, determining the trajectory feature information corresponding to the sample area based on the trajectory information within the sample area, may include the following steps:

[0080] The sample area is divided into multiple sub-regions of preset size;

[0081] Based on the location information of trajectory points within the sample area, determine the number of trajectory points contained in each sub-region within the sample area;

[0082] Based on the number of trajectory points contained in each sub-region within the sample area, the trajectory feature information of the sample area is obtained.

[0083] The division of the sample area can be based on the dimensions of the road itself, that is, the side length of the sub-area is made close to the width of the actual road.

[0084] In this scheme, the sample region can be divided based on the concept of GeoHash. The GeoHash algorithm is an algorithm that encodes latitude and longitude, transforming two-dimensional coordinates into one-dimensional coordinates, and partitioning address locations. For example, the grid size corresponding to GeoHash9 is approximately 5m, which is very close to the width of an actual road. Therefore, the sample region can be divided using the GeoHash9 grid as the unit of division.

[0085] After dividing the sample region into multiple sub-regions of preset size, the number of trajectory points contained in each sub-region can be determined based on the positional information of the trajectory points within the sample region. Then, the trajectory feature information of the sample region can be obtained based on the number of trajectory points contained in each sub-region. For example, the number of trajectory points contained in a sub-region can be used as the feature value of that sub-region. The feature values ​​of multiple sub-regions constitute the trajectory feature information of the sample region. If the sample region is divided into M*N sub-regions (M and N are both positive integers), the trajectory feature information of the sample region can be an image (i.e., a matrix) composed of M*N grids, where each grid (i.e., a matrix value) corresponds to the feature value of the corresponding sub-region.

[0086] Furthermore, considering that models typically require a uniform scale for input data, but in reality, sample data is taken from actual AOI data, often exhibiting variations in size and shape, making scale unification difficult. Also, considering that one GeoHash7 grid is approximately equal to 32 GeoHash9 grids, GeoHash7 grids can be used as the geographic feature representation unit for the AOI dimension. That is, a sample region can be divided into GeoHash7 grids that completely cover it, so each GeoHash7 grid can be considered as an image of 32*32 GeoHash9 grids. Therefore, to ensure data scale uniformity, the sample region can be divided according to the size of a GeoHash7 grid, and each resulting GeoHash7 grid can be used as input data for model training. During actual prediction, based on the model output results of each GeoHash7 grid, the GeoHash7 grids are restored to their original positions to generate roads for the entire sample region.

[0087] In one possible implementation, the road information within the sample area includes the roads within the sample area and their location information. Accordingly, step 12, determining the road label information corresponding to the sample area based on the road information within the sample area, may include the following steps:

[0088] The sample area is divided into multiple sub-regions of preset size;

[0089] Based on the location information of roads within the sample area, determine whether each sub-area within the sample area is a road, so as to obtain the road label information of the sample area.

[0090] The preset size can be set according to the size of the road within the sample area.

[0091] In this embodiment, the method of dividing the sample region is consistent with that in the previous embodiment, and will not be repeated here.

[0092] In step 13, the neural network model is trained using trajectory feature information and road feature information to obtain a road generation model.

[0093] The neural network model can include a semantic segmentation network and a language model building network. The semantic segmentation network is used to output a first result based on the trajectory feature information of the input sample data. The first result is used to determine whether each sub-region within the sample area of ​​the input sample data is a road. The language model building network is used to output a predicted path based on the trajectory feature information of the input sample data.

[0094] In one possible implementation, step 13 may include the following steps:

[0095] Determine the target trajectory feature information and target road label information of the target sample data used in this training;

[0096] The target trajectory feature information is input into the semantic segmentation network trained in this session to obtain the second result output by the semantic segmentation network trained in this session;

[0097] If the training stopping condition is not met, a first loss value is determined, wherein the first loss value is determined based on the second result and the target road label information;

[0098] Based on the first loss value, update the semantic segmentation network trained this time, and use the updated semantic segmentation network for the next training.

[0099] If the training stopping condition is met, the semantic segmentation network trained in this session will be identified as a road generation model.

[0100] The second result is used to determine whether each sub-region within the sample area of ​​the target sample data is a road.

[0101] In a training process, the trajectory feature information (i.e., target trajectory feature information) and road label information (i.e. target road label information) used in this training are first determined, and the target trajectory feature information is input into the semantic segmentation network for this training to obtain the second result.

[0102] Afterwards, it can be determined whether the semantic segmentation network in the neural network model trained this time meets the stopping conditions, so as to determine the subsequent training steps. The training stopping conditions may be, for example, reaching a preset number of training iterations, reaching a preset training duration, the loss value determined based on the second result and the target road label information being lower than a specified loss value, or the F1 score (a comprehensive indicator used to evaluate the model's capabilities) corresponding to the semantic segmentation network trained this time being higher than a specified score, etc.

[0103] If the training stopping condition is not met, the first loss value of the semantic segmentation network in this training can be determined. The first loss value can be determined based on the second result and the target road label information. For example, the cross-entropy can be calculated using the second result and the target road label information to obtain the first loss value.

[0104] For example, a semantic segmentation network can use Unet (fully convolutional neural network). Unet is essentially a neural network with an autoencoder structure. The input image is downsampled and compressed into a low-dimensional vector, then upsampled or deconvolved to restore it to its original dimension. The image is concatenated with structures similar to residual connections, allowing undamaged raw information to participate in the final decision. After transformation by the fully connected matrix, a probability heatmap of the same size as the input image is output, with each grid cell corresponding to the probability of whether it is a segmentation boundary. Based on this information, in this scheme, the semantic segmentation network using Unet actually models the spatial distribution relationship between GeoHash9 grids corresponding to trajectories and GeoHash9 grids corresponding to actual roads. It summarizes the discrete distribution patterns of most trajectory points and learns how to cluster based on discrete trajectory points to pinpoint the GeoHash9 grids where the actual roads are located. Specifically, a convolutional network is used to extract and calculate the target trajectory feature information, and output a matrix of the same size as the target trajectory feature information. Each grid of the matrix is ​​a GeoHash9 grid, and each value of the matrix represents the probability that the actual road falls into the corresponding GeoHash9 grid.

[0105] As described above, the neural network model may further include a language model construction network for outputting a predicted path based on the trajectory feature information of the input sample data. The method provided in this disclosure may further include the following steps:

[0106] The target trajectory feature information is input into the language model construction network trained in this session to obtain the target path output by the language model construction network trained in this session based on the target trajectory feature information.

[0107] If the training stopping condition is not met, a reference path is generated based on the target road label information, and a second loss value is calculated based on the target path and the reference path.

[0108] Therefore, updating the semantic segmentation network trained this time based on the first loss value can include the following steps:

[0109] Based on the first loss value and the second loss value, calculate the comprehensive loss value, and update the semantic segmentation network trained in this session based on the comprehensive loss value.

[0110] Specifically, the loss value can be calculated by comparing each location point in the target path with the corresponding location point in the reference path, and then summed to obtain the second loss value. For example, the cross-entropy can be calculated by comparing the location points in the target path with the corresponding location points in the reference path, and then summed to obtain the second loss value.

[0111] After obtaining the second loss value, the first and second loss values ​​can be harmonized and averaged to obtain the comprehensive loss value.

[0112] For example, a language model building network can use LSTM (Long Short-Term Memory). LSTM selectively forgets some historical information and adds some current input information through three gate structures: input gate, forget gate, and output gate. Finally, it integrates the information into the current state and generates an output state. Thus, the neurons cyclically absorb the feature information of each step in the sequence and integrate the information of all steps in the sequence in the final step. In the model training scenario provided in this disclosure, the starting point (e.g., a GeoHash9 grid as the starting point) and the ending point (e.g., a GeoHash9 grid as the ending point) can be determined based on the target trajectory feature information. The network is then built based on the language model to progressively predict the path. Simultaneously, based on the target road label information and the aforementioned starting and ending points, the road can be traversed through random sampling to obtain the path that can reach the ending point from the starting point (e.g., a GeoHash9 grid sequence). This path is used as the constructed "natural language structure," and the language model is trained based on this structure. In this way, the second loss value of the language model network is also used to update the semantic segmentation network. The gradient information generated by the language model can be used to introduce sequence constraints, thereby ensuring that the path determined by the output of the semantic segmentation network remains continuous rather than discontinuous.

[0113] Optionally, the training method provided in this disclosure may further include the following steps:

[0114] The language model construction network trained this time is updated based on the comprehensive loss value, and the updated language model construction network and the updated semantic segmentation network are used together for the next training.

[0115] In other words, the combined loss value calculated based on the first and second loss values ​​can be used not only for updating the semantic segmentation network, but also for updating the language model building network, thus enabling better training results in the next training iteration.

[0116] The above technical solution acquires sample data for training. This sample data includes trajectory information and road information within the sample area. For each sample data point, trajectory feature information corresponding to the sample area is determined based on the trajectory information, and road label information is determined based on the road information. The trajectory feature information characterizes the quantity of trajectories in each sub-region of the sample area, and the road label information characterizes whether each sub-region is a road. The neural network model is trained using these trajectory and road feature information to obtain a road generation model. This neural network model includes a semantic segmentation network, which outputs a first result based on the trajectory feature information of the input sample data. This first result determines whether each sub-region within the sample area of ​​the input sample data is a road. By training the neural network model based on the trajectory and road features within the area, a road generation model capable of generating road-related information is obtained. This road generation model has the ability to discover roads within the area. Therefore, for areas lacking internal road data, this model can generate road data within the area, which is beneficial for improving the navigation function within the park.

[0117] Figure 2 This is a flowchart of a road generation method provided according to one embodiment of the present disclosure. Figure 2 As shown, the method may include steps 21 to 24.

[0118] In step 21, trajectory information within the target area to be processed is obtained.

[0119] For example, trajectory information of a target area during a specified historical period can be collected.

[0120] In step 22, trajectory feature information corresponding to the target area is determined based on the trajectory information within the target area.

[0121] The method for determining the trajectory feature information within the target area is the same as the method for determining the trajectory feature information within the sample area provided above, and will not be repeated here.

[0122] In step 23, the trajectory feature information corresponding to the target area is input into the road generation model to obtain the output result of the road generation model.

[0123] The road generation model is trained using the road generation model training method provided in any embodiment of this disclosure. The output is used to determine whether each sub-region within the target area is a road. For example, the output can represent the probability that each sub-region of the target area is a road.

[0124] In step 24, the location information of roads within the target area is determined based on the output results and the location information of each sub-region within the target area.

[0125] Based on the output of the road generation model, target sub-regions with a probability higher than the preset probability can be identified, and the location corresponding to the target sub-region can be used as the location information of the roads contained in the target region.

[0126] Through the above technical solution, for areas lacking internal road data, a pre-trained road generation model capable of mining roads within the area can generate road data for that area, which helps improve the navigation function within the park.

[0127] Figure 3 This is a block diagram of a road generation model training apparatus according to one embodiment of the present disclosure. Figure 3 As shown, the device 30 includes:

[0128] The first acquisition module 31 is used to acquire sample data for training, the sample data including trajectory information within the sample area and road information within the sample area;

[0129] The first determining module 32 is used to determine, for each sample data, trajectory feature information corresponding to the sample area based on trajectory information within the sample area, and road label information corresponding to the sample area based on road information within the sample area, wherein the trajectory feature information is used to characterize the quantity characteristics of trajectories in each sub-region within the sample area, and the road label information is used to characterize whether each sub-region within the sample area is a road.

[0130] Training module 33 is used to train a neural network model using the trajectory feature information and the road feature information to obtain the road generation model. The neural network model includes a semantic segmentation network and a language model construction network. The semantic segmentation network is used to output a first result based on the trajectory feature information of the input sample data. The first result is used to determine whether each sub-region within the sample area of ​​the input sample data is a road. The language model construction network is used to output a predicted path based on the trajectory feature information of the input sample data.

[0131] Optionally, the training module 33 includes:

[0132] The first determination submodule is used to determine the target trajectory feature information and target road label information of the target sample data used in this training.

[0133] The first processing submodule is used to input the target trajectory feature information into the semantic segmentation network trained in this session to obtain the second result output by the semantic segmentation network trained in this session. The second result is used to determine whether each sub-region within the sample area of ​​the target sample data is a road.

[0134] The second determining submodule is used to determine a first loss value when the training stopping condition is not met, wherein the first loss value is determined based on the second result and the target road label information;

[0135] The first update submodule is used to update the semantic segmentation network trained in this session based on the first loss value, and to use the updated semantic segmentation network for the next training session.

[0136] The third determining submodule is used to determine the semantic segmentation network trained in this session as a road generation model when the training stopping condition is met.

[0137] Optionally, the training module 33 further includes:

[0138] The second processing submodule is used to input the target trajectory feature information into the language model construction network trained in this session, and obtain the target path output by the language model construction network trained in this session in response to the target trajectory feature information.

[0139] The calculation submodule is used to generate a reference path based on the target road label information when the training stopping condition is not met, and to calculate a second loss value based on the target path and the reference path.

[0140] The first update submodule is used to calculate a comprehensive loss value based on the first loss value and the second loss value, and update the semantic segmentation network trained in this session based on the comprehensive loss value.

[0141] Optionally, the training module 33 further includes:

[0142] The second update submodule is used to update the language model construction network trained this time according to the comprehensive loss value, and to use the updated language model construction network and the updated semantic segmentation network together for the next training.

[0143] Optionally, the trajectory information within the sample area includes the trajectory points within the sample area and the location information of each trajectory point;

[0144] The first determining module 32 includes:

[0145] The first partitioning submodule is used to divide the sample region into multiple sub-regions of preset size;

[0146] The fourth determining submodule is used to determine the number of trajectory points contained in each subregion of the sample area based on the location information of the trajectory points within the sample area.

[0147] The fifth determining submodule is used to obtain the trajectory feature information of the sample area based on the number of trajectory points contained in each sub-region within the sample area.

[0148] Optionally, the road information within the sample area includes the roads within the sample area and their location information;

[0149] The first determining module 32 includes:

[0150] The second partitioning submodule is used to divide the sample region into multiple sub-regions of preset size;

[0151] The sixth determination submodule is used to determine whether each sub-region within the sample area is a road based on the location information of the roads within the sample area, so as to obtain the road label information of the sample area;

[0152] The preset size is set according to the size of the road within the sample area.

[0153] Figure 4 This is a block diagram of a road generation apparatus provided according to one embodiment of the present disclosure. Figure 4 As shown, the device 40 includes:

[0154] The second acquisition module 41 is used to acquire trajectory information within the target area to be processed;

[0155] The second determining module 42 is used to determine trajectory feature information corresponding to the target area based on the trajectory information within the target area;

[0156] Processing module 43 is used to input trajectory feature information corresponding to the target area into a road generation model trained by the road generation model training device provided in any embodiment of this disclosure, and to obtain the output result of the road generation model. The output result is used to determine whether each sub-region within the target area is a road.

[0157] The third determining module 44 is used to determine the location information of roads contained in the target area based on the output results and the location information of each sub-area within the target area.

[0158] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0159] Figure 5This is a block diagram illustrating an electronic device 700 according to an exemplary embodiment. Figure 5 As shown, the electronic device 700 may include a processor 701 and a memory 702. The electronic device 700 may also include one or more of a multimedia component 703, an input / output (I / O) interface 704, and a communication component 705.

[0160] The processor 701 controls the overall operation of the electronic device 700 to complete all or part of the steps in the road generation model training method or road generation method described above. The memory 702 stores various types of data to support the operation of the electronic device 700. This data may include, for example, instructions for any application or method operating on the electronic device 700, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Multimedia component 703 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 702 or transmitted via communication component 705. The audio component also includes at least one speaker for outputting audio signals. I / O interface 704 provides an interface between processor 701 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IoT, eMTC, or other 5G technologies, or combinations thereof, is not limited here. Therefore, the corresponding communication component 705 may include: a Wi-Fi module, a Bluetooth module, an NFC module, etc.

[0161] In an exemplary embodiment, the electronic device 700 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the road generation model training method or road generation method described above.

[0162] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the road generation model training method or road generation method described above. For example, the computer-readable storage medium may be the memory 702 including the program instructions described above, which may be executed by the processor 701 of the electronic device 700 to complete the road generation model training method or road generation method described above.

[0163] Figure 6 This is a block diagram illustrating an electronic device 1900 according to an exemplary embodiment. For example, the electronic device 1900 may be provided as a server. (Refer to...) Figure 6 The electronic device 1900 includes a processor 1922, which may be one or more, and a memory 1932 for storing computer programs executable by the processor 1922. The computer program stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processor 1922 may be configured to execute the computer program to perform the aforementioned road generation model training method or road generation method.

[0164] Additionally, the electronic device 1900 may also include a power supply component 1926 and a communication component 1950. The power supply component 1926 can be configured to perform power management of the electronic device 1900, and the communication component 1950 can be configured to enable communication of the electronic device 1900, such as wired or wireless communication. Furthermore, the electronic device 1900 may also include an input / output (I / O) interface 1958. The electronic device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM Mac OS X TM UnixTM Linux TM etc.

[0165] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the road generation model training method or road generation method described above. For example, the computer-readable storage medium may be the memory 1932 including the program instructions described above, which may be executed by the processor 1922 of the electronic device 1900 to complete the road generation model training method or road generation method described above.

[0166] In another exemplary embodiment, a computer program product is also provided, comprising a computer program executable by a programmable device, the computer program having a code portion for performing the road generation model training method or road generation method described above when executed by the programmable device.

[0167] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0168] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0169] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A method for training a road generation model, characterized in that, The method includes: Acquire sample data for training, the sample data including trajectory information within the sample area and road information within the sample area; For each sample data, trajectory feature information corresponding to the sample area is determined based on the trajectory information within the sample area, and road label information corresponding to the sample area is determined based on the road information within the sample area. The trajectory feature information is used to characterize the quantity characteristics of trajectories in each sub-region within the sample area, and the road label information is used to characterize whether each sub-region within the sample area is a road. The neural network model is trained using the trajectory feature information and the road feature information to obtain the road generation model. The neural network model includes a semantic segmentation network and a language model construction network. The semantic segmentation network is used to output a first result based on the trajectory feature information of the input sample data. The first result is used to determine whether each sub-region within the sample area of ​​the input sample data is a road. The language model construction network is used to output a predicted path based on the trajectory feature information of the input sample data.

2. The method according to claim 1, characterized in that, The step of training a neural network model using the trajectory feature information and the road feature information to obtain the road generation model includes: Determine the target trajectory feature information and target road label information of the target sample data used in this training; The target trajectory feature information is input into the semantic segmentation network trained in this session to obtain the second result output by the semantic segmentation network trained in this session. The second result is used to determine whether each sub-region within the sample area of ​​the target sample data is a road. If the training stopping condition is not met, a first loss value is determined, wherein the first loss value is determined based on the second result and the target road label information; Based on the first loss value, update the semantic segmentation network trained this time, and use the updated semantic segmentation network for the next training. If the training stopping condition is met, the semantic segmentation network trained in this session will be determined as a road generation model.

3. The method according to claim 2, characterized in that, The method further includes: The target trajectory feature information is input into the language model construction network trained in this session to obtain the target path output by the language model construction network trained in this session in response to the target trajectory feature information. If the training stopping condition is not met, a reference path is generated based on the target road label information, and a second loss value is calculated based on the target path and the reference path. The step of updating the semantic segmentation network trained this time based on the first loss value includes: Based on the first loss value and the second loss value, a comprehensive loss value is calculated, and the semantic segmentation network trained in this session is updated based on the comprehensive loss value.

4. The method according to claim 3, characterized in that, The method further includes: The language model construction network trained this time is updated based on the comprehensive loss value, and the updated language model construction network and the updated semantic segmentation network are used together for the next training.

5. The method according to claim 1, characterized in that, The trajectory information within the sample area includes the trajectory points within the sample area and the location information of each trajectory point; The step of determining trajectory feature information corresponding to the sample region based on the trajectory information within the sample region includes: The sample area is divided into multiple sub-regions of preset size; Based on the location information of the trajectory points within the sample area, determine the number of trajectory points contained in each sub-region within the sample area; The trajectory feature information of the sample area is obtained based on the number of trajectory points contained in each sub-region within the sample area.

6. The method according to claim 1, characterized in that, Road information within the sample area includes the roads within that sample area and their location information; The step of determining road label information corresponding to the sample area based on road information within the sample area includes: The sample area is divided into multiple sub-regions of preset size; Based on the location information of roads within the sample area, determine whether each sub-region within the sample area is a road, so as to obtain the road label information of the sample area; The preset size is set according to the size of the road within the sample area.

7. A road generation method, characterized in that, The method includes: Obtain trajectory information within the target area to be processed; Based on the trajectory information within the target area, determine the trajectory feature information corresponding to the target area; The trajectory feature information corresponding to the target area is input into the road generation model trained by any one of claims 1-6 to obtain the output result of the road generation model, and the output result is used to determine whether each sub-region within the target area is a road; Based on the output results and the location information of each sub-region within the target area, the location information of the roads contained within the target area is determined.

8. A road generation model training device, characterized in that, The device includes: The first acquisition module is used to acquire sample data for training, the sample data including trajectory information within the sample area and road information within the sample area; The first determining module is used to determine, for each sample data, trajectory feature information corresponding to the sample area based on trajectory information within the sample area, and road label information corresponding to the sample area based on road information within the sample area, wherein the trajectory feature information is used to characterize the quantity characteristics of trajectories in each sub-region within the sample area, and the road label information is used to characterize whether each sub-region within the sample area is a road. The training module is used to train a neural network model using the trajectory feature information and the road feature information to obtain the road generation model. The neural network model includes a semantic segmentation network and a language model construction network. The semantic segmentation network is used to output a first result based on the trajectory feature information of the input sample data. The first result is used to determine whether each sub-region within the sample area of ​​the input sample data is a road. The language model construction network is used to output a predicted path based on the trajectory feature information of the input sample data.

9. A road generation device, characterized in that, The device includes: The second acquisition module is used to acquire trajectory information within the target area to be processed; The second determining module is used to determine trajectory feature information corresponding to the target area based on the trajectory information within the target area; The processing module is used to input the trajectory feature information corresponding to the target area into the road generation model trained by the road generation model training device of claim 8, and obtain the output result of the road generation model. The output result is used to determine whether each sub-region within the target area is a road. The third determining module is used to determine the location information of roads contained in the target area based on the output results and the location information of each sub-area within the target area.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method according to any one of claims 1-6, or when the program is executed by a processor, it implements the steps of the method according to claim 7.

11. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method of any one of claims 1-6, or to implement the steps of the method of claim 7.

Images