A method for generating a strip mine road network based on a two-path graph convolution network

Abstract

The application discloses a kind of open-pit mine road network generation methods based on two-way graph convolution network, belong to open-pit mine road network generation field.Region sample division is carried out to open-pit mine track area, and then the region sample is grid divided;Multi-track feature representation of each grid is generated using track data in the grid, and region grid track feature map is obtained;A road centerline prediction model comprising a residual encoder network, a two-way graph convolution network and a decoder network is constructed, and a road centerline probability map is generated using the region grid track feature map;The road centerline probability map is converted into a road centerline map, and all predicted road centerline maps are spliced to generate an initial road network for the open-pit mine, and the topology of the initial road network is refined by connecting broken road edges, and finally the open-pit mine road network is obtained.The application can extract rich track features, enhance the description of roads, improve the perception of spatial information and channel information, and improve the integrity and continuity of the road network topology.

Description

Technical Field

[0001] This invention relates to the field of road network generation, and in particular to a method for generating open-pit mine road networks based on a dual-path graph convolutional network. Background Technology

[0002] Road network data is a crucial component and foundation for national basic geographic information construction and national intelligent transportation development. However, in some special geographical environments, such as open-pit mines, road network data is still severely lacking, and due to production and safety considerations in open-pit mines, the road network is particularly important. As the basic geographic information for the open-pit mine transportation system, the open-pit mine road network not only serves as a base map reference for systems such as automated truck dispatching systems and equipment monitoring systems, but also forms the prerequisite for the intelligent construction of the entire mining system, thus possessing significant value for the development of smart mines.

[0003] The data used in road network generation methods, both domestically and internationally, mainly fall into two categories:

[0004] One approach involves using acquired remote sensing images of roads, processing them through image morphology, or training them with deep learning to generate a road network. However, this method is difficult to apply to open-pit mines. First, acquiring remote sensing images is costly and time-consuming, making it difficult to meet the need for timely updates to road networks in open-pit mines. Second, the relatively harsh environment of open-pit mines, such as the presence of large amounts of dust, easily blurs roads in the acquired remote sensing images, making it difficult to extract the road network structure.

[0005] Another type of method uses collected GPS data to generate road networks. Because GPS data is inexpensive, easy to obtain, and can accurately reflect road conditions, many scholars have studied automatic construction methods for urban road networks based on GPS data and achieved certain results. However, there is still very little research on the generation of road networks for open-pit mines. The methods for this type of data are mainly divided into two categories: machine learning-based and model-based road network generation methods.

[0006] In machine learning-based road network generation research, Edelkamp et al. first used K-means to cluster trajectory points to obtain the latest cluster points that represent the road network structure, and finally used trajectory information to connect these key points to form the road network topology. However, this method is very sensitive to the number and location of the initial seeds for clustering. Therefore, Qiu et al. used the DBSCAN algorithm to cluster trajectory points to identify straight lines in the map, and continuously sampled points along the straight lines as seeds, thus ensuring that the k-means clustering is close to the road centerline to generate a more accurate road network structure. Although such methods have good parallel processing capabilities for GPS data and have achieved good results in some simple urban road network construction, they do not consider the rich trajectory features contained in the trajectory information, resulting in the generation of many false and incorrect roads.

[0007] In model-based road network generation research, Lu et al. proposed a road network generation model based on CGCN, using trajectory point density maps and rasterized road maps as inputs. They also introduced a residual network to enhance the model generator's image generation capabilities, gradually generating realistic road images through a continuous "generation-adversarial" game. Feng et al. proposed an improved U-net residual network to predict road centerlines from rasterized GPS trajectory data, and used an improved U-net neural network to better predict road centerlines. These methods leverage the advantage of massive GPS data and extract rich trajectory features, improving the accuracy of road network generation. However, their methods lack effective establishment of global contextual information, making it difficult to generate complete and continuous road networks in the complex road environment of open-pit mines. Summary of the Invention

[0008] The purpose of this invention is to provide a method for generating open-pit mine road networks based on dual-path graph convolutional networks, which solves the problem that existing technologies do not fully utilize rich trajectory features and cannot effectively use deep learning models to model global context information to generate accurate and continuous open-pit mine road networks.

[0009] The objective of this invention is achieved by including the following steps:

[0010] Step 1: Divide the open-pit mine trajectory area into m region samples, and then divide each sample into n×n grids; use the trajectory data within the grid to generate a multi-trajectory feature representation for each grid.

[0011] Step 2: For the trajectory inflow-outflow correlation features in the multi-trajectory features of Step 1, first process them with an embedding unit; then perform feature fusion with other trajectory features in a concatenate manner; obtain the region grid feature map of each region sample; then input it into the feature encoding module to extract the depth features of the region grid feature map.

[0012] Step 3: Input the deep features into the global context information capture module for processing to obtain deep context features.

[0013] Step 4: Design a road centerline prediction module, perform upsampling and classification processing on the deep context features obtained in Step 3, and obtain the predicted road centerline probability map.

[0014] Step 5: Design a composite loss function for training. The optimal model parameters are obtained from the road centerline prediction model built in Steps 2, 3, and 4.

[0015] Step 6: Collect actual trajectory data of open-pit mine trucks, process it in Step 1, and input it into the road centerline prediction model obtained in Step 5 to obtain a probability map of all predicted road centerlines in the open-pit mine trajectory area; then, use the probability map of road centerlines to generate a road centerline map, and stitch together all the road centerline maps to form an initial road network; finally, use the topology connection module to effectively connect the broken road edges in the initial road network to form the final open-pit mine road network.

[0016] Further, in step 1, the multi-trajectory feature representation of the grid consists of trajectory point quantity features, trajectory line quantity features, trajectory velocity features, trajectory direction features, and trajectory inflow / outflow association features. The specific generation method is as follows: The trajectory point quantity feature is obtained by counting the number of all trajectory points in the grid; the trajectory line quantity feature is obtained by counting the number of times each trajectory segment crosses the grid, where a trajectory segment is formed by consecutive trajectory points; the trajectory velocity feature is obtained by calculating the average velocity of all trajectory points in the grid; and the trajectory direction feature is obtained by counting the number of times the movement direction of all trajectory points in the grid appears in the eight divided directions. The eight directions are obtained by dividing 0° to 360° into 45° intervals, where 0° is due north. Within a certain neighborhood of the grid, the trajectory inflow / outflow association features are obtained by recording the movement relationship between grids; the inflow refers to the trajectory flowing from a neighboring grid into the current grid, and the outflow refers to the trajectory flowing from the current grid into other neighborhood grids.

[0017] The fully connected layers, convolutional layers, max pooling layers, and transposed convolutional layers mentioned below are existing technologies.

[0018] Furthermore, in step 2, the embedding unit performs dimensionality reduction on the trajectory inflow-outflow correlation features to obtain the same dimension as other trajectory features; it consists of two fully connected layers, each followed by a ReLU function for nonlinear activation.

[0019] Furthermore, the feature encoding module is implemented by a residual encoder network to extract the depth feature X of the region grid feature map. This network consists of convolutional layers, max pooling layers, and four encoding units. The specific processing flow is as follows: First, the region grid feature map is convolved using a convolutional layer with a stride of 2 and a kernel size of 7×7. Then, the convolution result is downsampled using a max pooling layer with a window size of 3×3 and a stride of 2. Second, the downsampled result is input into four encoding units for processing. Each encoding unit consists of multiple residual convolutional blocks, each consisting of two consecutive convolutional layers with a kernel size of 3×3 and a residual structure. Except for the last encoding unit, each of the other encoding units is followed by a max pooling layer with a window size of 3×3 and a stride of 2 for downsampling. Finally, the depth feature X∈R of the region grid feature map is obtained after processing by the last encoding unit.H×W×D H, W, and D represent the height, width, and number of channels of the depth feature X, respectively.

[0020] Furthermore, in step 3, the global context information capture module is implemented by a dual-path graph convolutional network. This network uses the depth feature X obtained in step 2 to calculate its spatial adjacency matrix and channel adjacency matrix, respectively, and then uses them to perform spatially aware graph convolutional inference and channel-aware graph convolutional inference to obtain the corresponding features. and This is used to fuse with the deep feature X to obtain the deep context feature M. D Specifically, it includes the following sub-steps:

[0021] Step 3A: Spatial Adjacency Matrix A S ∈R L×L The calculation method is as follows:

[0022]

[0023] Represents the Softmax normalization function, and RP(·) means reshaping the input tensor into a matrix. It is a matrix multiplication operation, where T represents the transpose of the matrix.

[0024] Step 3B: The specific steps of spatially perceptual map convolutional inference are as follows: First, the depth feature X is processed continuously through convolutional layers, reshape transformation, and transpose transformation to obtain... After that, with A S Perform matrix multiplication, then multiply the result by a trainable weight matrix. Performing the operation from the hidden layer to the output layer yields the output features of the spatial awareness map convolution. The calculation method is as follows:

[0025]

[0026] Among them, W S The weighting operation is implemented through a 1-dimensional convolutional layer with a kernel size of 1×1.

[0027] Step 3C: The channel adjacency matrix is ​​generated as follows: First, a linear transformation function is used... Reduce the dimensionality of the input deep features X, and for The matrix is ​​obtained after performing the reshape transformation. To project X into the channel interaction space C, a linear transformation function is then used. Then, the X matrix is ​​processed using the reshape operation to obtain the projection matrix. Combine it with the matrix Multiplication yields a new feature V that projects X onto the channel interaction space C. C The calculation method is as follows:

[0028]

[0029] Where, θ(·) and The functions are all implemented using convolutional layers with a kernel size of 1×1; the generated new features This can be represented as having D1 nodes, each with dimension D2. It can be accessed via V. C Construct a connected graph to obtain the channel adjacency matrix.

[0030] Step 3D: The specific steps of channel-aware graph convolutional inference are as follows: Based on the channel adjacency matrix A obtained in step 3C... C Design it as an IA C This is used to perform Laplacian smoothing aggregation of neighboring node features. Then, it is combined with V... C and W C Matrix multiplication is used to obtain output features The calculation method is as follows:

[0031]

[0032] in, Let I represent the weight matrix of trainable edges in a specific layer, and let I represent the identity matrix.

[0033] Step 3E: Convolve the spatial awareness map to output features Channel-aware map convolution output features The deep context feature M is obtained by fusing it with the deep feature X. D ∈R H×W×D The calculation method is as follows:

[0034]

[0035]

[0036] Where ξ(·) and φ(·) are functions that restore the channel dimension of the input features to D dimensions. This indicates a point-by-point summation operation.

[0037] Furthermore, in step 4, the road centerline prediction module is implemented by a decoder network, used to process the deep context features M obtained in step 3. D The network performs upsampling and classification processing; it consists of decoding units, transposed convolutional layers, and convolutional layers. The specific processing flow is as follows: First, four sets of decoding units are used to process M... DUpsampling is performed; each decoding unit has the same structure, consisting of a 1×1 convolutional layer, a 3×3 transposed convolutional layer with a stride of 2, and another 1×1 convolutional layer connected in sequence. The output of each decoding unit is skipped to the output of the corresponding encoding unit in the residual encoder, performing addition-style feature fusion. Then, the output of the last decoding unit is processed sequentially through a 4×4 transposed convolutional layer with an upsampling stride of 2 and a 3×3 convolutional layer with one filter. The processed result is then classified using the sigmoid function to obtain the predicted road centerline probability map Y∈R. n×n .

[0038] Furthermore, in step 5, the composite loss function L Total It is determined by the binary cross-entropy loss function L BCE and Dice loss function L Dice Composition. The calculation method is as follows:

[0039]

[0040]

[0041] L Total =L BCE +L Dice

[0042] Among them, Y * ∈R n×n It is the true value of the road centerline relative to Y. ε is a local minimum value used to ensure that the numerator and denominator are not zero.

[0043] Further, in step 6, the method for generating the road centerline map is as follows: the grids with a probability value greater than 0.5 in the road centerline probability map are regarded as road center points, and the other road center points in the eight-neighborhood of all road center points are connected to form the road centerline, and the final map is the road centerline map.

[0044] The topology connection module effectively connects the broken road edges in the initial road network to obtain the final open-pit mine road network. The specific steps are as follows: First, by determining whether the extension of a road edge within a certain radius R intersects with other edges, all possible connections between road edges are generated. If a perpendicular extension from the end point o of edge e intersects with another edge e2, a new intersection point v1 is generated. Connecting the end point o of edge e with intersection point v1 generates a new road connection. If there is no intersection point, consider connecting the ends of two edges within a radius R. If the resulting angle is less than 90 degrees, a new road connection is generated; otherwise, no new road connection is generated.

[0045] Due to the adoption of the above solution, the beneficial effects of the present invention are as follows:

[0046] (1) Rich trajectory features were extracted and fused, which enhanced the description of road features and reduced the generation of erroneous and false open-pit mine roads.

[0047] (2) Using coding units composed of residual convolutional blocks to extract deep features effectively avoids model degradation and enhances the capture of high-level semantic information.

[0048] (3) Spatial-aware map convolution was used to construct global contextual information of deep features in the spatial dimension, enhancing the long-distance dependencies between road segments. Channel-aware map convolution was used to construct global contextual information of deep features in the channel dimension, enhancing the expressive power of road features. These two processes enhanced the model's prediction of continuous and complete open-pit mine roads, effectively addressing the complex road environment of open-pit mines.

[0049] (4) By generating all possible connections between road edges under certain conditions, the generated open-pit mine road network has better topology.

[0050] In summary, this invention solves the problem that existing technologies fail to fully utilize rich trajectory features and cannot effectively use deep learning models to model global context information, thus failing to generate accurate and continuous open-pit mine road networks. Therefore, the objective of this invention is achieved. Attached image description:

[0051] Figure 1 This is a framework diagram of a method for generating open-pit mine road networks based on a dual-path graph convolutional network.

[0052] Figure 2 This is a schematic diagram of a dual-path graph convolutional network.

[0053] Figure 3 This is a schematic diagram of the road edge connection. Detailed Implementation

[0054] Example 1: The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious. All inventions utilizing the concept of the present invention are protected.

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0056] like Figure 1As shown, a method for generating open-pit mine road networks based on dual-path graph convolutional networks includes the following steps:

[0057] Step 1: Divide the open-pit mine trajectory area into m region samples, and then divide each sample into n×n grids; use the trajectory data within the grid to generate a multi-trajectory feature representation for each grid.

[0058] Step 2: For the trajectory inflow-outflow correlation features in the multi-trajectory features of Step 1, first process them with an embedding unit; then perform feature fusion with other trajectory features in a concatenate manner; obtain the region grid feature map of each region sample; then input it into the feature encoding module to extract the depth features of the region grid feature map.

[0059] Step 3: Input the deep features into the global context information capture module for processing to obtain deep context features.

[0060] Step 4: Design a road centerline prediction module, perform upsampling and classification processing on the deep context features obtained in Step 3, and obtain the predicted road centerline probability map.

[0061] Step 5: Design a composite loss function for training. The optimal model parameters are obtained from the road centerline prediction model built in Steps 2, 3, and 4.

[0062] Step 6: Collect actual trajectory data of open-pit mine trucks, process it in Step 1, and input it into the road centerline prediction model obtained in Step 5 to obtain a probability map of all predicted road centerlines in the open-pit mine trajectory area; then, use the probability map of road centerlines to generate a road centerline map, and stitch together all the road centerline maps to form an initial road network; finally, use the topology connection module to effectively connect the broken road edges in the initial road network to form the final open-pit mine road network.

[0063] Further, in step 1, the multi-trajectory feature representation of the grid consists of trajectory point quantity features, trajectory line quantity features, trajectory velocity features, trajectory direction features, and trajectory inflow / outflow association features. The specific generation method is as follows: The trajectory point quantity feature is obtained by counting the number of all trajectory points in the grid; the trajectory line quantity feature is obtained by counting the number of times each trajectory segment crosses the grid, where a trajectory segment is formed by consecutive trajectory points; the trajectory velocity feature is obtained by calculating the average velocity of all trajectory points in the grid; the trajectory direction feature is obtained by counting the number of times the movement direction of all trajectory points in the grid appears in eight divided directions. The eight directions are obtained by dividing 0° to 360° into 45° intervals, where 0° is due north; within a certain neighborhood of the grid, the trajectory inflow / outflow association features are obtained by recording the movement relationship between grids; inflow refers to the trajectory flowing from a neighboring grid into the current grid, and outflow refers to the trajectory flowing from the current grid into other neighboring grids. This movement relationship is recorded using a binary matrix with values ​​of 0 and 1. When a trajectory flows from a neighboring grid into the current grid, the binary matrix between them is set to 1, otherwise it is set to 0. Similarly, when a trajectory flows from the current grid to another neighboring grid, the binary matrix between them is set to 1, otherwise it is set to 0.

[0064] Furthermore, in step 2, the embedding unit performs dimensionality reduction on the trajectory inflow and outflow correlation features to obtain the same dimension as other trajectory features. It consists of two fully connected layers, each followed by a ReLU function for non-linear activation.

[0065] Furthermore, the feature encoding module is implemented by a residual encoder network and is used to extract the depth features X of the region grid feature map; such as Figure 1 The residual encoder network consists of convolutional layers, max-pooling layers, and four encoding units. The specific processing flow is as follows: First, a convolutional layer with a stride of 2 and a kernel size of 7×7 is used to convolve the region grid feature map. Then, the convolution result is downsampled using a max-pooling layer with a window size of 3×3 and a stride of 2. Next, the downsampled result is input into four encoding units for processing. Each encoding unit consists of multiple residual convolutional blocks, each composed of two consecutive convolutional layers with a kernel size of 3×3 and a residual structure. Except for the last encoding unit, each of the other encoding units is followed by a max-pooling layer with a window size of 3×3 and a stride of 2 for downsampling. Finally, after processing by the last encoding unit, the depth feature X∈R of the region grid feature map is obtained. H×W×D H, W, and D represent the height, width, and number of channels of the depth feature X, respectively.

[0066] Furthermore, in step 3, the global context information capture module is implemented using a dual-path graph convolutional network. For example... Figure 2As shown, the network uses the deep features X obtained in step 2 to calculate its spatial adjacency matrix and channel adjacency matrix, respectively. These are then used to perform spatially aware map convolutional inference and channel-aware map convolutional inference, respectively, to obtain the corresponding features. and This is used to fuse with the deep feature X to obtain the deep context feature M. D Specifically, it includes the following sub-steps:

[0067] Step 3A: As Figure 2 As shown in the convolutional part of the spatially perceptual graph, the spatial adjacency matrix A S ∈R L×L The calculation method is as follows:

[0068]

[0069] Represents the Softmax normalization function, and RP(·) means reshaping the input tensor into a matrix. It is a matrix multiplication operation, where T represents the transpose of the matrix.

[0070] Step 3B: The specific steps of spatially perceptual map convolutional inference are as follows: First, the depth feature X is sequentially processed through a convolutional layer with a kernel size of 1×1, followed by a reshape transformation and a transpose transformation to obtain... After that, with A S Perform matrix multiplication, then multiply the result by a trainable weight matrix. Perform the operation from the hidden layer to the output layer. Obtain the output features of the spatial awareness map convolution. The calculation method is as follows:

[0071]

[0072] Among them, W S The weighting operation is implemented through a 1-dimensional convolutional layer with a kernel size of 1×1.

[0073] Step 3C: As Figure 2 As shown in the convolution part of the channel-sensing map, the specific method for generating the channel adjacency matrix is ​​as follows: First, through a linear transformation function... Reduce the dimensionality of the input deep features X, and for The matrix is ​​obtained after performing the reshape transformation. To project X into the channel interaction space C, a linear transformation function is then used. Then, the X matrix is ​​processed using the reshape operation to obtain the projection matrix. Combine it with the matrix Multiplication yields a new feature V that projects X into the channel interaction space G. CThe calculation method is as follows:

[0074]

[0075] Where, θ(·) and The functions are all implemented using convolutional layers with a kernel size of 1×1. New features are generated. This indicates there are D1 nodes, and each node has a dimension of D2. This can be achieved through V... C Construct a connected graph whose adjacency matrix is

[0076] Step 3D: The specific steps of channel-aware graph convolutional inference are as follows: Based on the channel adjacency matrix A obtained in step 3C... C Design it as an IA C This is used to perform Laplacian smoothing aggregation of neighboring node features. Then, it is combined with V... C and W C Matrix multiplication is used to obtain output features The calculation method is as follows:

[0077]

[0078] in, represents the trainable edge weight matrix for a specific layer, implemented using a 1D convolutional layer with a kernel size of 1×1. I represents the identity matrix, implemented using skip connections.

[0079] Step 3E: Convolve the spatial awareness map to output features Channel-aware map convolution output features The deep context features are obtained by fusing them with the deep feature X. The calculation method is as follows:

[0080]

[0081]

[0082] Here, ξ(·) and φ(·) are functions that restore the channel dimension of the input features to D-dimensionality, which is achieved by a convolutional layer with a kernel size of 1×1. This indicates a point-by-point summation operation.

[0083] Further, in step 4, the road centerline prediction module is implemented by a decoder network, used to perform upsampling and classification processing on the deep context features (MD) obtained in step 3. For example... Figure 1The decoder network, as shown, consists of decoding units, transposed convolutional layers, and convolutional layers. The specific processing flow is as follows: First, four sets of decoding units perform upsampling on the MD (Modal Matrix). Each decoding unit has the same structure, consisting of a 1×1 convolutional layer, a 3×3 transposed convolutional layer with a stride of 2, and another 1×1 convolutional layer connected sequentially. The output of each decoding unit is then skipped to the output of the corresponding encoding unit in the residual encoder, performing addition-style feature fusion. Next, the output of the last decoding unit is processed sequentially through a 4×4 transposed convolutional layer with an upsampling stride of 2 and a 3×3 convolutional layer with one filter. Finally, the processing result is classified using the sigmoid function to obtain the predicted road centerline probability map Y∈R. n×n .

[0084] Furthermore, in step 5, the composite loss function L Total It is determined by the binary cross-entropy loss function L BCE and Dice loss function L Dice Composition. The calculation method is as follows:

[0085]

[0086]

[0087] L Total =L BCE +L Dice

[0088] Among them, Y * ∈R n×n It is the true value of the road centerline relative to Y. ε is a local minimum value used to ensure that the numerator and denominator are not zero.

[0089] Further, in step 6, the open-pit mine truck trajectory data is represented as T = {tr1, tr2, ... tr i ...,tr |T|}. tr i It is the i-th trajectory segment in T, which consists of a series of continuous trajectory points, denoted as... i |tr| Let |tr| represent the number of trajectory points of the i-th trajectory. This represents the j-th trajectory point of the i-th trajectory segment. Contains five elements (lat) ij ,lon ij ,t ij ,t ij ,dr ij(), representing latitude, longitude, time, vehicle speed, and direction of travel, respectively. The trajectory points within the trajectory segment are arranged in chronological order, i.e.,

[0090] The method for generating the road centerline map is as follows: grids with a probability value greater than 0.5 in the road centerline probability map are regarded as road center points, and all other road center points within the eight-neighborhood of all road center points are connected to form the road centerline.

[0091] The topology connection module effectively connects the broken road edges in the initial road network to obtain the final open-pit mine road network. The specific steps are as follows: Figure 3 As shown, by determining whether the extension of a road edge within a certain radius R intersects with other edges, all possible connections between road edges are generated:

[0092] (1) If the extension from edge e can intersect with another edge e2, a new intersection point v1 is generated. Connect the end point o of edge e with the intersection point v1 to generate a new road connection.

[0093] (2) If there is no intersection point, consider connecting the ends of the two sides within a radius R. If the resulting angle is less than 90 degrees, a new road connection is generated; otherwise, it is not. Figure 3 As shown, the end point o of edge e is connected to the end point of another edge e1, and the included angle is less than 90 degrees. Then a new road connection is generated between o and the intersection point v2.

[0094] In summary, this invention addresses the problem that existing technologies fail to fully utilize rich trajectory features and cannot effectively use deep learning models to model global context information, thus hindering the generation of accurate and continuous open-pit mine road networks. A method for generating open-pit mine road networks based on a dual-path graph convolutional network is proposed.

Claims

1. A method for generating open-pit mine road networks based on dual-path graph convolutional networks, characterized in that, Includes the following steps: Step 1: Divide the open-pit mine trajectory area into Each sample is divided into several regions, and each sample is further divided into... A grid of varying sizes; generating multi-trajectory feature representations for each grid using trajectory data within the grid; Step 2: The trajectory inflow and outflow correlation features in the multi-trajectory features of Step 1 are first processed with an embedding unit; then, they are fused with other trajectory features using a concatenation method to obtain the region grid feature map of each region sample; finally, they are input into the feature encoding module to extract the depth features of the region grid feature map. Step 3: Input the deep features into the global context information capture module for processing to obtain deep context features; Step 4: Design a road centerline prediction module, perform upsampling and classification processing on the deep context features obtained in Step 3, and obtain the predicted road centerline probability map; Step 5: Design a composite loss function for training. Based on the road centerline prediction model built in Steps 2, 3, and 4, obtain the optimal model parameters. Step 6: Collect actual trajectory data of open-pit mine trucks, process it in Step 1, and input it into the road centerline prediction model obtained in Step 5 to obtain a probability map of all predicted road centerlines in the open-pit mine trajectory area; then, use the road centerline probability map to generate a road centerline map, and stitch together all the road centerline maps to form an initial road network; finally, use the topology connection module to effectively connect the broken road edges in the initial road network to form the final open-pit mine road network; The global context information capture module is implemented by a dual-path graph convolutional network, which utilizes the depth features obtained in step 2. The spatial adjacency matrix and channel adjacency matrix are calculated separately, and then used to perform spatial perceptual map convolutional inference and channel perceptual map convolutional inference, respectively, to obtain the corresponding features. and Used for depth features By fusing the features, deep contextual features are obtained. ; Specifically, it includes the following sub-steps: Step 3A: Spatial Adjacency Matrix Calculation method: ； in, , Given two linear transformation functions, This represents the Softmax normalization function. This means reshaping the input tensor into a matrix. It is a matrix multiplication operation, where T represents the transpose of the matrix; Step 3B: The specific steps of spatially perceptual map convolutional inference are as follows: First, for deep features... After continuous processing through convolutional layers, reshape transformation, and transpose transformation, the result is... ; after that and Perform matrix multiplication, then multiply the result by a trainable weight matrix. Performing the operation from the hidden layer to the output layer yields the output features of the spatial awareness map convolution. The calculation method is as follows: ； Step 3C: The channel adjacency matrix is ​​generated as follows: First, a linear transformation function is used... Reduce the dimensionality of the input deep features X, and for The matrix is ​​obtained after performing the reshape transformation. In order to project X into the channel interaction space In the middle; then use a linear transformation function. Then, the X matrix is ​​processed using the reshape operation to obtain the projection matrix. ; and combine it with the matrix Multiplying the results gives the projection of X onto the channel interaction space. New features The calculation method is as follows: ； New features generated Indicated as having There are nodes, and the dimension of each node is . ; can be passed Construct a connected graph to obtain the channel adjacency matrix. ; Step 3D: The specific steps of channel-aware graph convolutional inference are as follows: Based on the channel adjacency matrix obtained in Step 3C... Design it as To perform Laplacian smoothing aggregation of neighboring node features; then respectively with and Matrix multiplication is used to obtain output features The calculation method is as follows: ； in, This represents the weight matrix of trainable edges in a specific layer. Represents the identity matrix; Step 3E: Convolve the spatial awareness map to output features Channel-aware map convolution output features and depth features Fusion is performed to obtain deep contextual features The calculation method is as follows: ； ； ； in, and It is a function that restores the channel dimension of the input features to D dimensions. This indicates a point-by-point summation operation; Representing depth features Height, width, and number of channels.

2. The method for generating open-pit mine road networks based on dual-path graph convolutional networks according to claim 1, characterized in that, In step 1, the multi-trajectory feature representation of the grid consists of trajectory point quantity features, trajectory line quantity features, trajectory velocity features, trajectory direction features, and trajectory inflow / outflow correlation features; The specific generation method is as follows: the number of all trajectory points in the grid is counted to obtain the trajectory point quantity feature; the number of times each trajectory segment crosses the grid is counted to obtain the trajectory line quantity feature, where the trajectory segment is formed by continuous trajectory points; the average velocity of all trajectory points in the grid is calculated to obtain the trajectory velocity feature; The trajectory direction characteristics are obtained by counting the number of times the movement direction of all trajectory points in the grid appears in the eight divided directions; The eight directions are... to by The interval is divided to obtain, where, It is due north; within a certain neighborhood of the grid, the trajectory inflow and outflow correlation features are obtained by recording the movement relationship between the grids; Inflow refers to a trajectory flowing from a neighboring grid into the current grid, while outflow refers to a trajectory flowing from the current grid into other neighboring grids.

3. The method for generating open-pit mine road networks based on dual-path graph convolutional networks according to claim 1, characterized in that, In step 2, the embedding unit performs dimensionality reduction on the trajectory inflow-outflow correlation features to obtain the same dimension as other trajectory features; It consists of two fully connected layers, each followed by a ReLU function for non-linear activation; The feature encoding module is implemented by a residual encoder network and is used to extract the depth features of the region grid feature map. The residual encoder network consists of convolutional layers, max pooling layers, and four sets of encoding units. The specific processing flow is as follows: First, it passes through a convolutional layer with a stride of 2 and a kernel size of... The convolutional layer performs convolution processing on the region grid feature map; then, the convolution result is processed using a window. Downsampling is performed on the max pooling layer with a step size of 2; Secondly, the downsampling results are input into four encoding units for processing; each encoding unit consists of multiple sets of residual convolutional blocks, and each residual convolutional block consists of two consecutive convolutional kernels of size [missing information]. It consists of convolutional layers with residual structures; except for the last coding unit, each coding unit is followed by a window. A max-pooling layer with a stride of 2 is used for downsampling; finally, the depth features of the region grid feature map are obtained after processing by the last encoding unit. ,in Representing depth features Height, width, and number of channels.

4. The method for generating open-pit mine road networks based on dual-path graph convolutional networks according to claim 1, characterized in that, In step 4, the road centerline prediction module is implemented by a decoder network and is used to process the deep context features obtained in step 3. Upsampling and classification processing are performed; the decoder network consists of decoding units, transposed convolutional layers, and convolutional layers; the specific processing flow is as follows: first, four sets of decoding units are used to perform upsampling and classification processing. Upsampling is performed; each decoding unit has the same structure, consisting of sequentially connected convolutional kernels of size [size missing]. The convolutional layer and kernel size are A transposed convolutional layer with a stride of 2, and another convolutional kernel with a size of... The system consists of convolutional layers; and the output of each decoding unit is skipped to the output of the corresponding encoding unit in the residual encoder to perform addition-style feature fusion; then, the output of the last decoding unit is passed through a convolutional kernel of size [size missing]. A transposed convolutional layer with an upsampling stride of 2 and a series of... After processing by a convolutional layer with one filter, the result is classified using the sigmoid function to obtain the predicted road centerline probability map. .

5. The method for generating open-pit mine road networks based on dual-path graph convolutional networks according to claim 1, characterized in that, In step 6, the method for generating the road centerline map is as follows: grids with a probability value greater than 0.5 in the road centerline probability map are regarded as road center points, and all other road center points within the eight-neighborhood of all road center points are connected to form the road centerline. The final map obtained is the road centerline map. The topology connection module effectively connects the broken road edges in the initial road network to obtain the final open-pit mine road network. The specific steps are: by determining whether the extension of a road edge within a certain radius R intersects with other edges, all possible connections between road edges are generated; if from the edge... End point Vertical extension can connect with another edge If they intersect, a new intersection point is generated. Connecting edges End point Intersection If there is no intersection, consider connecting the ends of the two sides within a radius R; if the angle generated is less than 90 degrees, a new road connection is generated, otherwise no new road connection is generated.

Images