CN-115204304-B - Strip mine road network generation method based on two-way graph convolution network

CN115204304BCN 115204304 BCN115204304 BCN 115204304BCN-115204304-B

Abstract

A strip mine road network generation method based on a double-path graph convolution network belongs to the field of strip mine road network generation. Dividing a strip mine track area into area samples, dividing the area samples into grids, generating multi-track feature representation of each grid by utilizing track data in the grids to obtain an area grid track feature map, constructing a road center line prediction model comprising a residual encoder network, a double-path graph convolution network and a decoder network, generating a road center line probability map by utilizing the area grid track feature map, converting the road center line probability map into a road center line map, splicing all predicted road center line maps to generate an initial strip mine road network, and refining the topological structure of the initial road network by connecting broken road edges to finally obtain the strip mine road network. The method has the advantages that rich track features can be extracted, the description of the road is enhanced, the relation perception of space information and channel information is improved, and the integrity and the continuity of a road network topological structure are improved.

Inventors

ZHANG LEI
CHEN FENGHUA
Liu bailong
LIANG ZHIZHEN

Assignees

中国矿业大学

Dates

Publication Date: 20260512
Application Date: 20220725

Claims (5)

1. A strip mine road network generation method based on a two-way graph convolution network is characterized by comprising the following steps: step 1, dividing the track area of the strip mine into Samples of each region, each sample being subdivided into Generating a multi-track characteristic representation of each grid by utilizing track data in the grids; Step 2, processing the track in-out flow associated features in the multi-track features in the step 1 by using an embedding unit, then carrying out feature fusion in Concatenate modes with other track features to obtain a region grid feature map of each region sample, and finally inputting the region grid feature map to a feature coding module to extract depth features of the region grid feature map; step 3, inputting the depth characteristics into a global context information capturing module for processing to obtain the depth context characteristics; Step 4, designing a road center line prediction module, and performing up-sampling processing and classification processing on the depth context characteristics obtained in the step 3 to obtain a predicted road center line probability map; step 5, designing a composite loss function training, and obtaining optimal model parameters by the road center line prediction model constructed in the step 2, the step 3 and the step 4; Step 6, collecting actual track data of the strip mine trucks, inputting the actual track data into the road center line prediction model obtained based on the step 5 after the actual track data are processed in the step 1, and obtaining all predicted road center line probability maps of the strip mine track areas; The global context information capturing module is realized by a two-way graph convolution network, and the two-way graph convolution network utilizes the depth characteristics obtained in the step 2 Respectively calculating a space adjacent matrix and a channel adjacent matrix, and then respectively performing space perception graph convolution reasoning and channel perception graph convolution reasoning to obtain corresponding characteristics And For and depth features Fusion is carried out to obtain depth context characteristics Specifically comprises the following substeps: Step 3A spatial adjacency matrix Is calculated by the following steps: ; Wherein, the , For two of the linear transformation functions, Representing the Softmax normalization function, Representing the input tensor reshape as a matrix, Is a matrix multiplication operation, T represents the transpose of the matrix; Step 3B, the specific step of the convolution reasoning of the space perception graph is that firstly, depth characteristics are subjected to Is obtained after the continuous processing of the convolution layer, reshape transformation and transposition transformation After and after Performing matrix multiplication operation, and multiplying the matrix multiplication result by a trainable weight matrix Performing operation from the hidden layer to the output layer to obtain output characteristics of the convolution of the space perception map The calculation method is as follows: ; Step 3C, generating the channel adjacency matrix by a linear transformation function Reducing the dimension of the input depth feature X and The matrix is obtained after reshape transformation For projecting X into the channel interaction space In, re-using a linear transformation function And processing X by combining reshape operations to obtain a projection matrix And combining it with matrix Multiplication results in projection of X into channel interaction space Novel features of (a) The calculation method is as follows: ; generated new features Is shown as having Each node has the dimension of Can pass through Constructing a connected graph to obtain a channel adjacency matrix ; The step 3D is a specific step of channel perception graph convolution reasoning, which is based on the channel adjacency matrix obtained in the step 3C Designed as To perform Laplace smoothing to aggregate the features of neighboring nodes, and then to separate them And Matrix multiplication to obtain output characteristics The calculation method is as follows: ; Wherein, the A weight matrix representing trainable edges for a particular layer, Representing the identity matrix; step 3E, convolving the space perception image with the output characteristics Channel aware graph convolution output features And depth features Fusion to obtain depth context features The calculation method is as follows: ; ; ; Wherein, the And Is a function of restoring the channel dimension of the input feature to the D dimension, Representing a point-by-point summation operation; Respectively represent depth features Is a high, wide and channel number.
2. The strip mine road network generating method based on the two-way graph rolling network according to claim 1, wherein in the step 1, the multi-track feature representation of the grid consists of track point number features, track line number features, track speed features, track direction features and track in-out flow correlation features; counting the number of all track points in a grid to obtain track point number characteristics, counting the times of all track segments passing through the grid to obtain track line number characteristics, wherein the track segments are formed by continuous track points, calculating the average speed of all track points in the grid to obtain track speed characteristics, counting the times of all track point moving directions in the grid in eight divided directions to obtain track direction characteristics; The eight directions are to To the point of To be used for The division is made for intervals which, among other things, Obtaining track in-out flow correlation characteristics by recording the movement relation of tracks among grids in a certain neighborhood range of the grids; inflow refers to the trajectory flowing from the neighborhood grid into the current grid, and outflow refers to the trajectory flowing from the current grid into the other neighborhood grid.
3. The strip mine road network generating method based on the two-way graph rolling network is characterized in that in the step 2, the embedded unit performs dimension reduction processing on the trace in-out flow correlation characteristics to obtain the same dimension as other trace characteristics, and the strip mine road network generating method based on the two-way graph rolling network is composed of two full-connection layers, wherein a ReLU function is followed by each full-connection layer for nonlinear activation; the feature coding module is realized by a residual coder network and is used for extracting depth features of the regional grid feature map The residual encoder network consists of a convolution layer, a max pooling layer and four groups of encoding units, and the specific processing flow is that firstly, the convolution kernel size is as follows by a step length of 2 The convolution layer of the (2) carries out convolution processing on the regional grid characteristic diagram, and then uses a window as the convolution result The maximum pooling layer with the step length of 2 is processed by downsampling, then the downsampling result is input into 4 coding units for processing, wherein the coding units are composed of a plurality of groups of residual convolution blocks, and the residual convolution blocks are formed by two continuous convolution kernels with the size of A convolution layer with residual structure, wherein the other coding units are followed by windows except the last coding unit And finally, processing the largest pooling layer with the step length of 2 by the last coding unit to obtain the depth characteristic of the regional grid characteristic map Wherein Respectively represent depth features Is a high, wide and channel number.
4. The strip mine road network generating method based on two-way graph rolling network as set forth in claim 1, wherein in step 4, the road center line prediction module is implemented by a decoder network for characterizing the depth context obtained in step 3 The decoder network comprises decoding units, a transposed convolution layer and a convolution layer, and comprises four decoding units Performing up-sampling processing in which each decoding unit has the same structure and is formed by sequentially connecting convolution kernels of the size Is of the convolution layer, convolution kernel size of A transposed convolution layer of step size 2, and another convolution kernel of size The output of each decoding unit is connected with the output of the corresponding coding unit in the residual coder in a jump way to execute the characteristic fusion of the Addition mode, and then the output result of the last decoding unit is sequentially processed by a convolution kernel with the size of Transposed convolutional layer with up-sampling step length of 2 and one by one After the convolution layer processing with the number of the filters being 1, classifying and outputting the processing result by utilizing a sigmoid function to obtain a predicted road center line probability map 。
5. The strip mine road network generating method based on the two-way graph rolling network of claim 1, wherein in the step 6, the road center line graph generating method comprises the steps of regarding grids with probability values larger than 0.5 in a road center line probability graph as road center points, connecting all road center points with other road center points in eight adjacent areas to form a road center line, and finally obtaining a graph which is a road center line graph; the topology connection module is used for effectively connecting broken road edges in an initial road network to obtain a final strip mine road network, and specifically comprises the steps of generating all possible connection between the road edges by judging whether an extension line of a road side within a certain radius R has an intersection with other edges or not, and if so, generating all possible connection between the road edges from the edges End point Can extend vertically to the other side Intersection, then generate new intersection point Connecting edge End point of (2) With the crossing point Generating a new road connection, if there is no intersection point, considering that the ends of the two sides are connected within the radius R, if the generated included angle is smaller than 90 degrees, generating a new road connection, otherwise, not generating a new road connection.

Description

Strip mine road network generation method based on two-way graph convolution network Technical Field The invention relates to the field of road network generation, in particular to a strip mine road network generation method based on a two-way graph convolution network. Background Road network data is an important component and foundation for national basic geographic information construction and national intelligent traffic construction, but in some special geographic environment areas, such as open-pit mines, road network data is very lacking, and road network is particularly important for open-pit mine production, safety and other reasons. The strip mine road network is used as basic geographic information of a strip mine transportation system, can be used as a base map reference of a strip mine truck automatic scheduling system, a strip mine equipment monitoring system and other systems, is also a precondition foundation for intelligent construction of the whole mine system, and has important value for intelligent mine construction. The data used in the domestic and foreign road network generation method mainly comprises two types: the road network is finally generated by training the collected road remote sensing image through image morphological processing or using deep learning. However, the method is difficult to apply to the strip mine, firstly, the acquisition cost of the remote sensing image is high, the acquisition period is long, and the method is difficult to adapt to the requirement of timely updating of the strip mine road network. Secondly, the relatively harsh environment of the strip mine, such as the existence of a large amount of sand dust, makes the road in the collected remote sensing image easily blurred, and makes it difficult to extract the road network structure. The other type of road network is generated by utilizing collected GPS data, because the GPS data has low cost, is easy to obtain and can truly reflect road conditions, a plurality of students research an automatic construction method of an urban road network on the basis of the GPS data and obtain a certain result at present, but the research on the generation of the road network of an open pit mine is very little, and the methods of the data are mainly divided into two types, namely a road network generation method based on machine learning and a road network based on a model. In a road network generation study based on machine learning, edelkamp et al first use K-means to cluster the trace points to obtain the latest cluster points that can represent the road network structure, and finally connect these key points with the trace information to form the topology structure of the road network. However, this method is very sensitive to the number of initial seeds and their locations, so Qiu et al identified straight lines in the map by using the DBSCAN algorithm to cluster the trace points and continuously sampled points along the straight lines as seeds to ensure that k-means clusters are close to the road centerline to generate a more accurate road network structure. Although such methods have good parallel processing capability for GPS data and achieve good results on some simple urban road network constructions, they do not consider the rich track features contained in the track information, resulting in the generation of many false and erroneous roads. In model-based road network generation research, land et al propose a CGCN-based road network generation model, taking a track point density map and a rasterized road map as inputs, simultaneously introducing the image generation capability of a residual network enhancement model generator, and gradually generating a real road image through continuous game of 'generation-antagonism'. Feng et al propose an improved U-net residual network to predict the road centerline of the rasterized GPS trajectory data and use the improved U-net neural network to better predict the road centerline of the road network. The method utilizes the advantage of massive GPS data, extracts rich track characteristics and improves the accuracy of road network generation. Their method does not effectively build global context information, which makes it difficult to generate a complete continuous road network in a complex road environment of a strip mine. Disclosure of Invention The invention aims to provide a strip mine road network generation method based on a two-way graph convolution network, which solves the problems that the prior art does not fully utilize rich track characteristics, and an effective deep learning model cannot be used for modeling global context information so as to generate an accurate and continuous strip mine road network. The object of the invention is achieved in that it comprises the following steps: Dividing the track area of the strip mine into m area samples, dividing each sample into grids with the size of n multiplied by n, and generating multi-track characteri