Self-adaptive incremental updating method and device for strip mine area map and medium

Abstract

The invention provides a self-adaptive incremental updating method and device for a strip mine area map and a medium, relates to the technical field of automatic driving, and particularly relates to a method for acquiring road and operation area data by deploying multiple sensors on a mine vehicle and adopting a self-adaptive strategy. And carrying out hierarchical feature extraction on the preprocessed point cloud and image data, constructing a feature pyramid, and realizing multi-modal feature fusion by using an attention mechanism. And through comparing new and old data with semantic prediction, identifying a change area, implementing vehicle secondary updating on an operation area, implementing threshold triggering type updating on a main road, and finally outputting a new version map through global correction. Accurate updating of the mining area map is achieved, and calculation and transmission loads are reduced. And the recognition precision and robustness of the environment change are improved through hierarchical perception and multi-modal fusion. The self-adaptive updating strategy gives consideration to the timeliness of a dynamic area and the stability of a static area, and effectively supports the safe and stable operation of the automatic driving system in the mining area.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an adaptive incremental update method, device, and medium for open-pit mine maps. Background Technology

[0002] With the rise of smart mines and unmanned operations, autonomous driving technology in open-pit mines is developing rapidly. Compared to traditional manual driving, autonomous mining vehicles offer higher scheduling efficiency, lower operating costs, and better safety, and have been gradually implemented in coal mines, non-ferrous metal mines, and sand and gravel yards.

[0003] In related technologies, incremental maps generated through difference calculations lack consideration for the irregular shapes of mining pits and mounds, resulting in missing details and distorted outlines. This makes it impossible to accurately reflect actual terrain changes and affects map update accuracy. Identifying changed areas by comparing old and new data, without considering the functional differences and change predictions of mining areas and main roads, can easily misjudge noise as change or miss slowly expanding mound areas, leading to updates that are either too large or too small.

[0004] When the relevant technologies are used to update the entire mining area, they use incremental updates within a certain range. This cannot adapt to the difference between the frequent changes in the number of vehicles in the work area and the low-frequency changes in the main road, resulting in low update efficiency and high resource consumption. Summary of the Invention

[0005] This invention provides an adaptive incremental update method for open-pit mine maps. By using a deep learning feedback mechanism to prioritize updates for changed areas and dynamically adjust the update strategy, it achieves high-precision, low-latency, and adaptive incremental updates of mine maps, ensuring the efficiency and safety of navigation and positioning for autonomous vehicles in the mine area.

[0006] The methods include: S101: Install data acquisition equipment on mining vehicles, adopt an adaptive acquisition strategy to collect data from road areas and work areas, and perform time synchronization, spatial calibration and normalization preprocessing of point cloud and image data. S102: Perform hierarchical feature extraction on the preprocessed data, process the point cloud and image data of the near, medium and far distance layers respectively, and construct a feature fusion pyramid; S103: Based on the feature fusion pyramid, features are weighted through channel attention and spatial attention mechanisms to achieve the fusion of image and point cloud features and obtain multimodal feature representations; S104: Utilize generative adversarial networks to optimize multimodal features and generate high-quality incremental map images; S105: Compare the newly collected data with the existing map, and combine the semantic information and prediction data generated by the deep learning model to identify the areas of change in the mining area; S106: Based on the dynamic changes in area and positioning accuracy, and combined with update frequency weight and accuracy weight, the quality of the map is evaluated. S107: Based on the results of change area identification and map quality assessment, an adaptive incremental update strategy is adopted to update the work area at the vehicle level and the main road at the threshold level. S108: Perform global correction on the map after multiple incremental updates to obtain a new version of the global map.

[0007] It should be further explained that S102 specifically includes the following methods: To process near-field layer data, PointNet++ network is used to extract local features from high-resolution point clouds, and HRNet is used to extract high-resolution detail features from images. The two networks are then stitched together to obtain near-field layer fusion features. To process mid-range layer data, the point cloud is voxelized and features are extracted using VoxelNet. The image is then processed using HRNet to extract features, and the mid-range layer fusion features are obtained by stitching them together. To process the data of the distant layer, the sparse point cloud is projected onto spherical coordinates to generate a 2D image, and features are extracted using a 2D CNN to obtain the fused features of the distant layer. Upsampling and downsampling operations are performed on the fusion features of the near-range, mid-range, and far-range layers respectively, and the features are spliced ​​together to construct a preliminary feature fusion pyramid; The initial feature fusion pyramid is fused using alternating top-down and bottom-up approaches to obtain multi-scale features.

[0008] It should be further noted that S103 specifically includes the following methods: Global average pooling is performed on the multi-scale feature maps of the point cloud in the feature fusion pyramid to extract global information for each channel. Global information is input into the fully connected layer and activated by the Sigmoid function to generate channel attention weights. Global average pooling and max pooling are performed on the multi-scale feature maps of the point cloud in the feature fusion pyramid, and the spatial description features are obtained by concatenating them. Spatial description features are convolved and then activated by Sigmoid to generate spatial attention weights. Image and point cloud features are weighted using channel and spatial attention weights respectively, and then the weighted features are added and fused to obtain a multimodal feature representation.

[0009] It should be further explained that S104 specifically includes the following methods: S1041: Construct a generator network, input multimodal features, and generate incremental map maps through a multi-layer network structure; Construct a discriminator network, take real map data or incremental map output by the generator as input, and output the discrimination result of the authenticity of the input data; Set the generator loss function to use the cross-entropy loss function so that the generated map increment map is close to the real map; Set the discriminator loss function to maximize the distinction between the real map and the generated map increment; The generator and discriminator are trained alternately, and their parameters are iteratively optimized to obtain a high-quality incremental map.

[0010] It should be further noted that S105 specifically includes the following methods: The newly acquired point cloud, image, and pose data are compared with existing map data in multiple dimensions to identify areas of initial change. Using a deep learning model, semantic classification is performed on the initial changed regions to determine the region type; Input the current mining area's characteristic data and historical data into the prediction model, and predict the future changes in the area by calculating the change Pchange as follows; Prediction formula:

[0011] in, It is the predicted change. This is the current characteristic data of the mining area. It is historical data; By combining semantic classification results and predicted changes, the update priority and range of change for each changed region are determined. Integrate the priority, scope, and type information of all changed areas to clarify the final changed mining area.

[0012] It should be further noted that S106 specifically includes the following methods: Determine the target area for map quality assessment, covering both identified dynamic and static areas, and count the total number of assessment areas N; Collect dynamic area changes of each assessment area and changes in positioning accuracy parameter; The update frequency weight is adaptively adjusted based on the degree of dynamic region change. The accuracy weights are adaptively adjusted according to the regional accuracy requirements. ; Substitute the parameters

[0013] Calculate the individual evaluation values ​​for each region, and then average them to obtain the overall map quality evaluation value. ; This represents the dynamic change in the area of ​​the region. To represent the variation in positioning accuracy, N is the number of areas being evaluated; Set the differential quality assessment threshold Q threshold ,Will The individual assessment values ​​for each region are compared with the corresponding thresholds to determine whether the update conditions are met.

[0014] It should be further noted that S107 specifically includes the following methods: Based on the types of changed areas identified in S105 and the quality assessment results in S106, the areas to be updated are divided into two categories: work areas and main roads. For the work area, based on the type of work, train frequency, and S106... The work progress adjustment coefficient is dynamically determined. ; S105 and certain Substitute into the following formula to calculate the area of ​​the work area update zone. Based on the incremental map of S104, perform vehicle-level incremental updates;

[0015] in, It updates the area. It is an adjustment coefficient updated based on the progress of the work; For main roads, based on their historical change data and S106 Set adaptive update threshold The formula is determined by updating the threshold:

[0016] Determine if the update conditions are met; if so, perform an incremental update. Record update information for two types of areas, including the coordinate range of the updated area, the update time, the map increment map identifier used, and the update result verification data.

[0017] It should be further explained that S108 specifically includes the following methods: Collect all spatial data and update records of local incremental updates; Detect and handle overlaps and conflicts between adjacent update regions; Map data fusion based on the principle of spatial consistency; Apply a global optimization algorithm to correct map geometric distortions; Generate a new version of global map data with consistent topology.

[0018] According to another embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the adaptive incremental update method for open-pit mine maps.

[0019] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the adaptive incremental update method for the open-pit mine map.

[0020] As can be seen from the above technical solutions, the present invention has the following advantages: The adaptive incremental update method for open-pit mine maps provided by this invention employs an adaptive acquisition strategy that differentiates acquisition based on the frequency of regional changes, avoiding resource waste or data loss and improving acquisition efficiency. Layered processing by distance adapts to the characteristics of data at different resolutions, and a feature fusion pyramid integrates multi-scale features, achieving high precision at close range and comprehensive feature representation at medium and long ranges. Channel and spatial attention mechanisms precisely focus on key features and suppress redundant information, while multimodal fusion integrates the advantages of image texture and point cloud geometry, enhancing feature robustness and discriminative power. Generative adversarial networks optimize through game theory, generating incremental maps with rich details, providing reliable data for updates. Combining data comparison, semantic classification, and prediction data, precise location of changed areas is achieved. An evaluation system is constructed using quantitative indicators and adaptive weights to replace subjective judgment. Differentiated incremental strategies are designed for work areas and main roads to avoid computational overload caused by uniform updates across the entire area, while ensuring real-time updates for high-frequency changing areas and on-demand updates for static areas. Global correction eliminates the cumulative errors and regional connection conflicts of local updates, and standardized processing allows the map to be directly adapted to the autonomous driving system. The output map can accurately reflect the real-time terrain of the mining area, providing reliable support for autonomous driving path planning, positioning and navigation, and ensuring the safe and efficient operation of mining operations. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 Flowchart of an embodiment of the adaptive incremental update method for open-pit mine maps; Figure 2 A schematic diagram of sensor configuration and data acquisition; Figure 3This is a schematic diagram of hierarchical feature extraction; Figure 4 A schematic diagram of the pyramid of fusion features; Figure 5 This is a schematic diagram of feature fusion for the attention mechanism. Figure 6 Flowchart of an adaptive incremental update method for open-pit mine maps; Figure 7 This is a schematic diagram of an electronic device. Detailed Implementation

[0023] Combination Figures 1 to 5 In summary, the adaptive incremental update method for open-pit mine maps provided by this invention addresses the problems of frequent dynamic changes in open-pit mines, poor adaptability of existing map update technologies, missing blind spot data, and heavy manual intervention. This method integrates multi-source sensor data, extracts key features using a deep convolutional neural network, and identifies changed areas through semantic segmentation and dynamic change detection. It also uses generative adversarial networks to predict and complete uncollected map data, ensuring the integrity and accuracy of the map data. Furthermore, it prioritizes updating changed areas through a deep learning feedback mechanism and dynamically adjusts the update strategy. This achieves high-precision, low-latency, and adaptive incremental updates of mine maps, ensuring the efficiency and safety of navigation and positioning for autonomous vehicles in mines.

[0024] The adaptive incremental update method for open-pit mine maps according to this application will be described in detail below. Specific details such as particular system structures and technologies are presented for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0025] It should be understood that, when used in this specification, terms include indicating the presence of a described feature, integral, step, operation, element, and / or component, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms include, encompass, have, and variations thereof mean including but not limited to, unless otherwise specifically emphasized.

[0026] The statements such as "one embodiment" or "some embodiments" described in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the statements such as "in one embodiment," "in some embodiments," "in other embodiments," and "in still other embodiments" in this application do not necessarily refer to the same embodiment, but rather mean one or more, but not all, embodiments, unless otherwise specifically emphasized.

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 6 The diagram shows a flowchart of an adaptive incremental update method for an open-pit mine map in a specific embodiment. The method includes: S101: Install data acquisition equipment on mining vehicles, adopt an adaptive acquisition strategy to collect data from road areas and work areas, and perform time synchronization, spatial calibration, and normalization preprocessing of point cloud and image data.

[0029] In some embodiments, the data acquisition equipment installed on mining vehicles may include lidar, integrated navigation GNSS / IMU sensors, and cameras, employing different acquisition strategies for different areas. For main road areas with low change frequencies, point cloud, image, and vehicle pose data acquisition begins upon receiving a signal from the cloud, with lidar, cameras, and GNSS / IMUs having relatively low acquisition frequencies. Conversely, for high-change operation areas, such as loading / unloading areas and spoil heaps, data acquisition begins when the vehicle leaves the loading or unloading position after completing loading / unloading operations and stops after traveling a certain distance; the sensor acquisition frequency is even higher in these areas.

[0030] It should be noted that the adaptive acquisition strategy in this embodiment can be based on the data acquisition target area, including road areas and work areas. Different acquisition strategies can be adopted according to the frequency and importance of area changes, which fully improves resource utilization and data acquisition efficiency.

[0031] As the main road area experiences relatively low-frequency changes, theoretically every significant change can be predicted. After receiving signals from the cloud, the vehicle begins collecting point cloud, image, and vehicle pose data for map updates. The LiDAR collection frequency can be set to 10Hz, the camera frame rate to 10Hz, and the GNSS / IMU frequency to 100Hz. Data collection begins when the vehicle enters the area and continues until it leaves.

[0032] In high-frequency operational areas, especially loading / unloading areas and spoil heaps, data acquisition begins when mining vehicles are about to leave their loading or unloading positions after a loading / unloading operation. Data acquisition ceases once the vehicle has traveled a certain distance. The lidar acquisition frequency can be set to 20Hz, the camera frame rate to 30Hz, and the GNSS / IMU frequency to 100Hz. This high-frequency acquisition of the operational area helps capture dynamic environmental changes during operations, and is particularly suitable for handling transient changes that may occur during mining operations.

[0033] This embodiment also involves time synchronization, which is to ensure the accurate alignment of data from multiple sensors and to minimize the time synchronization error of all sensors.

[0034] Hardware-level time synchronization is achieved using GNSS-PTP (Precision Time Protocol) to ensure that all sensors maintain a consistent clock. The timestamp error of all sensors is guaranteed to be less than 1ms.

[0035] Formula (1) This embodiment also performs spatial calibration. Spatial calibration is to ensure that data from different sensors can be accurately aligned, especially when performing multi-sensor fusion and data projection, where calibration accuracy is crucial.

[0036] The extrinsic parameters of the lidar and the integrated navigation GNSS / IMU are calibrated separately, as are the lidar and the camera. The relative positions and attitudes of each sensor are calculated.

[0037] Optimize the rotation matrix R and translation vector T using the Kabsch algorithm: Formula (2) The calibration accuracy of this embodiment is: angle error < 0.1°, position error < 2cm.

[0038] This embodiment also performs normalization processing in the following manner.

[0039] The input data for normalization processing is based on autonomous driving in open-pit mines. The input data mainly comes from LiDAR point clouds and visual images (Camera). The point cloud data contains spatial coordinates. and reflection intensity Visual images provide texture and color information. For each input data, preprocessing is required, including standardization, noise reduction, and conversion into a form suitable for network input.

[0040] Point cloud data preprocessing: Before being input into the neural network, the point cloud data needs to be normalized to make the data range more suitable for the needs of the neural network and avoid the influence of features of different scales on network training. This involves normalizing the spatial coordinates of the point cloud. Normalization can be performed as follows: Formula (3) in, These are the means for each axis. These are the standard deviations of each axis. These are the normalized point cloud coordinates.

[0041] The Statistical Outlier Removal (SAC) algorithm is used to remove noisy points and ensure the quality of point cloud data.

[0042] Image data preprocessing: Images are normalized and resized to ensure each image has the same resolution. The normalization formula is: Formula (4) Where I is the image pixel value, and These are the mean and standard deviation of the image, respectively. This is the normalized image data.

[0043] The image is resized to a fixed size. The resolution of the image data can be adjusted according to the needs of the neural network. A common input size is 224x224 for network models like VGG16; for larger networks (such as ResNet), a size of 256x256 can be chosen.

[0044] S102: Perform hierarchical feature extraction on the preprocessed data, process the point cloud and image data of the near, medium and far distance layers respectively, and construct a feature fusion pyramid.

[0045] S102 specifically includes the following methods: S1021: Process near-range layer data. Use PointNet++ network to extract local features from high-resolution point clouds and HRNet to extract high-resolution detail features from images. Then, stitch together the data to obtain near-range layer fusion features.

[0046] In some embodiments, the inputs are: preprocessed high-resolution point cloud (>100 points / m²) and image data.

[0047] Point cloud processing: The PointNet++ network is used to directly process the raw point cloud, which processes local features in the point cloud through a hierarchical structure. Formula (5) MLP stands for Multilayer Perceptron. It is a max pooling method that downsamples the feature map output by the convolutional layer, retains the most significant features and reduces the dimensionality of the data, thus avoiding overfitting. Point cloud coordinates and reflection intensity.

[0048] Image processing: High-resolution feature extraction is maintained using HRNet (High-Resolution Network), which utilizes parallel network layers of multiple resolutions to preserve image details.

[0049] Formula (6) Feature fusion: Formula (7) Output: High-precision geometric features (error < 2cm).

[0050] Processing near-field layer data can accurately extract geometric and texture details of the near-field environment, ensuring high accuracy of near-field features and supporting fine-grained updates of near-field area maps.

[0051] S1022: Process mid-range layer data, voxelize the point cloud and extract features using VoxelNet, extract features from the image using HRNet, and stitch together to obtain mid-range layer fusion features.

[0052] In some embodiments, the input for processing mid-range layer data is: a medium-resolution point cloud (20-50 points / m²) and an image.

[0053] The principle of point cloud processing is: point cloud voxelization, converting the point cloud into a 3D mesh. Formula (8) in, This refers to the voxel size, which is set to 10cm here.

[0054] VoxelNet processes the data as follows: Formula (9) The image processing method is as follows: similar to that used for close-up shots, a high-resolution image is employed to obtain... .

[0055] The feature fusion method is as follows: Formula (10) Output: Object contour features (accuracy < 5cm) This embodiment balances the 3D structure of mid-range point clouds with computational efficiency, effectively extracting volumetric contour features and meeting the needs of mid-range regional map updates.

[0056] S1023: Process the data of the distant layer, project the sparse point cloud into spherical coordinates to generate a 2D image, extract features using a 2D CNN, and obtain the fused features of the distant layer.

[0057] In some embodiments, the processing method for long-distance layers (50-200m) is as follows: Input: Sparse point cloud (<10 points / ㎡) Spherical coordinate projection: Transformation formula: Formula (11) Formula (12) Formula (13) Projection diagram generation: Generate a 2D image from a distant point cloud using spherical coordinates: Formula (14) After processing with a 2D CNN, sparse feature maps are obtained. .

[0058] This embodiment solves the problem of feature extraction from sparse point clouds, effectively capturing features in distant areas and covering a wide range of mining environments.

[0059] S1024: Upsampling and downsampling operations are performed on the fusion features of the near-range, mid-range, and far-range layers respectively, and the features are spliced ​​together to construct a preliminary feature fusion pyramid.

[0060] In some embodiments, the feature fusion pyramid is a hierarchical feature representation method that captures information at different scales by constructing feature maps at multiple scales from low to high resolution. Point cloud features and image features at each scale have already been fused. The fused feature maps are then upsampled and downsampled to form a pyramid. Feature map fusion is performed alternately using top-down and bottom-up methods to obtain multi-scale features. Figure 4 As shown.

[0061] Fusion principle: Formula (15).

[0062] This embodiment integrates features from different distance layers to form a multi-scale feature representation, adapting to environmental changes at different scales in the mining area.

[0063] S1025: The initial feature fusion pyramid is fused using alternating top-down and bottom-up methods to obtain multi-scale features.

[0064] In some embodiments, preprocessed data undergoes hierarchical feature extraction at near (0-10m), mid (10-50m), and far (50-200m) distances. For near-distance processing, the point cloud is processed using a PointNet++ network to obtain local features, and the image is processed using HRNet to preserve high-resolution features and retain image details. Then, the point cloud and image features are fused. For mid-distance processing, the point cloud is voxelized into a 3D mesh and processed using VoxelNet. The image also undergoes high-resolution image processing, and then the point cloud and image features are fused. For far-distance processing, sparse point clouds are projected into 2D images using spherical coordinates, and then processed using a 2D CNN to obtain sparse feature maps. Feature map fusion is performed alternately using top-down and bottom-up approaches to obtain multi-scale features that capture information at different scales.

[0065] S103: Based on the feature fusion pyramid, features are weighted through channel attention and spatial attention mechanisms to achieve the fusion of image and point cloud features, resulting in a multimodal feature representation.

[0066] In some embodiments, after the feature fusion pyramid is constructed, the features are weighted using channel attention and spatial attention mechanisms. The channel attention mechanism assigns different importance weights to each channel, extracts global information for each channel through global average pooling, and then generates attention weights for each channel via fully connected layers. The spatial attention mechanism assigns importance weights to each location, pools and concatenates the spatial dimensions of each channel, and then calculates spatial attention using convolution. Finally, the calculated attention weights are used to weight and fuse the image and point cloud features to obtain the final multimodal feature representation.

[0067] S104: Utilize generative adversarial networks to optimize multimodal features and generate high-quality incremental map images.

[0068] In some embodiments, generative adversarial networks are used to optimize multimodal features. The generator and discriminator learn through game theory. The generator imitates real terrain changes to generate new map features, and the discriminator judges the difference between the generated map and the actual terrain, continuously optimizing the generator to ultimately generate a high-quality incremental map.

[0069] S105: Compare the newly collected data with the existing map, and combine the semantic information and prediction data generated by the deep learning model to identify the areas of change in the mining area.

[0070] In some embodiments, newly acquired data is compared with existing maps, and semantic information and predicted data generated by a deep learning model are combined to identify areas of change in the mining area. By comparing newly acquired data with the existing ground, areas of change, including variations in height and area, are identified. The deep learning model predicts possible changes based on current mining area feature data and historical data, thereby guiding the update strategy and selecting areas with significant changes.

[0071] S106: Based on the dynamic changes in area and positioning accuracy, and combined with update frequency weight and accuracy weight, the quality of the map is evaluated.

[0072] In some embodiments, map quality is assessed based on changes in dynamic area and positioning accuracy, combined with update frequency weighting and accuracy weighting. Considering the characteristics of dynamic areas with high change frequency, such as mining areas or earthwork areas, and areas with frequent vehicle traffic requiring high accuracy, as well as static areas such as main roads with low change frequency, map quality assessment values ​​are calculated using appropriate weighting coefficients to determine whether a map update is necessary.

[0073] Dynamic area = work area, which is also the mining and loading area and spoil heap.

[0074] This embodiment evaluates map quality based on the different characteristics of dynamic and static areas, and comprehensively considers update frequency and accuracy weights to reasonably determine whether the map needs to be updated, thus improving the rationality of updates.

[0075] S107: Based on the results of change area identification and map quality assessment, an adaptive incremental update strategy is adopted to update the work area at the vehicle-time level and the main road at the threshold level.

[0076] In some embodiments, an adaptive incremental update strategy is adopted based on the results of change area identification and map quality assessment. For high-frequency change areas such as work zones, real-time map updates are performed at the vehicle frequency level, updating only areas with significant changes, with the updated area area correlated with the vehicle frequency change. For static areas such as main roads, a change threshold is set to determine whether to update the map; if the area change exceeds the threshold, incremental updates are performed, reducing unnecessary calculations for static areas.

[0077] The adaptive incremental update strategy in this embodiment adopts different update methods for different regions. It updates high-frequency changing regions in real time and updates static regions according to a threshold, thereby reducing the computation and transmission burden and improving update efficiency.

[0078] S108: Perform global correction on the map after multiple incremental updates to obtain a new version of the global map.

[0079] In some embodiments, a global correction is performed on the map after multiple incremental updates, integrating all local updates to ensure the consistency and accuracy of the map, resulting in a new version of the global map.

[0080] This embodiment performs global correction on the map after multiple incremental updates to ensure map consistency and accuracy, guarantee map quality, and provide reliable map data for applications such as autonomous driving.

[0081] In one embodiment of the present invention, based on step S103, the following is a possible embodiment and its specific implementation is described in a non-limiting manner. S103 specifically includes the following methods: S1031: Perform global average pooling on the multi-scale feature maps of the point cloud in the feature fusion pyramid to extract global information for each channel.

[0082] In some embodiments, after the feature pyramid is constructed, an attention fusion mechanism is proposed to weight the features through channel and spatial attention mechanisms. The weighted feature fusion yields the final multimodal feature representation, thereby improving the model's performance in complex dynamic environments.

[0083] The channel attention mechanism in this embodiment assigns different importance weights to each channel, enabling the model to focus on which channels are most critical to the current task. For multi-scale features of point cloud data, global average pooling can be used to extract global information for each channel, and then a fully connected layer is used to generate attention weights for each channel.

[0084] Channel attention calculation formula: For point cloud feature maps Its channel attention weight The calculation process is as follows: 1. Global Average Pooling: First, perform global average pooling on each channel to calculate the description of each channel: Formula (16).

[0085] S1032: Input global information into the fully connected layer, and generate channel attention weights by activating the Sigmoid function.

[0086] In some embodiments, the fully connected layer calculates the weights: the weight coefficient for each channel is obtained through the fully connected layer. : Formula (17) in, This is the weight matrix, and b is the bias term. It is the activation function of Sigmoid.

[0087] Spatial attention mechanisms are used to assign importance weights to each location, typically focusing on changes in spatial location within point cloud data. Similar to channel attention, spatial attention dynamically focuses on more important spatial regions by calculating the weight of each spatial location. This embodiment generates channel attention weights, allowing the model to prioritize channels critical to map updates, thus improving feature targeting.

[0088] S1033: Perform global average pooling and max pooling on the multi-scale feature maps of the point cloud in the feature fusion pyramid, and then stitch them together to obtain spatial description features.

[0089] In some embodiments, for spatial attention calculation methods, for point cloud feature maps Its channel attention weight The calculation process is as follows: Global average pooling and max pooling: Pooling is performed on the spatial dimension of each channel, resulting in two descriptions:

[0090] Formula (18) The concatenation method is as follows: the results of average pooling and max pooling are concatenated. Formula (19).

[0091] This embodiment captures key information about spatial dimensions through two pooling methods, providing a comprehensive basis for spatial weight calculation.

[0092] S1034: Perform convolution on the spatial description features and generate spatial attention weights by activating them with Sigmoid.

[0093] In some embodiments, the spatial attention of convolution is calculated as follows: Formula (20) in, It is a convolutional kernel. It generates spatial attention weights, enabling the model to focus on key spatial regions and enhance its ability to represent features.

[0094] S1035: Image and point cloud features are weighted using channel and spatial attention weights respectively, and then the weighted features are added and fused to obtain a multimodal feature representation.

[0095] In some embodiments, the feature weighting and fusion method is as follows: the calculated attention weights are used to weight the image and point cloud features, as shown in the following formula: Formula (21) in, This represents element-wise multiplication. and These are the attention weights for image and point cloud features, respectively.

[0096] The feature fusion method is as follows: the weighted image and point cloud features can be fused using an additive approach. Assuming the image and point cloud features have already been weighted, they can be fused in the following way: Formula (22).

[0097] S1035 integrates the advantages of image texture and point cloud geometry by weighted fusion of multimodal features, resulting in more robust features and ensuring the quality of subsequent map generation.

[0098] As can be seen, by assigning importance weights to different feature channels through channel attention and to different spatial locations through spatial attention, key features are dynamically highlighted. The weighted image is then fused with point cloud features to obtain more discriminative multimodal features, supporting subsequent map updates.

[0099] In one embodiment of the present invention, based on step S104, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S104 specifically includes the following: S1041: Construct a generator network, input multimodal features, and generate incremental map maps through a multi-layer network structure.

[0100] In some embodiments, Generative Adversarial Networks (GANs) are used to optimize map reconstruction, with GANs optimizing the map update process. GANs generate high-quality incremental map images through game-theoretic learning between a generator and a discriminator. The generator produces new map features by mimicking real-world terrain changes, while the discriminator continuously optimizes the generator by judging the differences between the generated map and the actual terrain.

[0101] In this embodiment, the generator takes multimodal features as input and uses network structures such as multilayer perceptrons and convolutions to generate incremental map maps that mimic real terrain changes.

[0102] S1042: Construct a discriminator network, take real map data or incremental map output by the generator as input, and output the discrimination result of the authenticity of the input data.

[0103] In some embodiments, the discriminator receives real map data and outputs a discrimination probability through a network structure such as convolution to determine whether the input is a real map or a generated map. In this way, constructing the discriminator provides a basis for optimizing the generator, ensuring that the authenticity verification of the generated map increment has a reliable evaluation subject.

[0104] S1043: Set the generator loss function, using the cross-entropy loss function, to make the generated map increment map closer to the real map.

[0105] In some embodiments, the generator's loss function is as follows: The goal of the generator is to minimize the difference from the real map, and the cross-entropy loss function is typically used to measure the difference between the generated map and the real map.

[0106] Formula (23) in, D is the loss of the generator, and D is the loss of the discriminator. It is the generated incremental map. is the input noise, and N is the number of samples.

[0107] In this embodiment, the generator loss function is set to clarify the optimization objective of the generator, so that the generated incremental map is highly similar to the real map at the pixel level or feature level.

[0108] S1044: Set the discriminator loss function to maximize the distinction between the real map and the generated map increment.

[0109] In some embodiments, the discriminator objective in the discriminator's loss function is to maximize the discriminator's ability to distinguish between the real and generated graphs:

[0110] Formula (24) in, It is the loss of the discriminator. It is the discriminator's judgment of the real data.

[0111] By maximizing this loss, the discriminator can accurately distinguish between the real map and the generated map increment.

[0112] This embodiment sets a discriminator loss function to clarify the discriminator's optimization goal, enabling it to accurately distinguish between real and generated maps, and in turn, drive the generator to improve the generation quality.

[0113] S1045: Alternately train the generator and discriminator, and obtain a high-quality incremental map by iteratively optimizing the parameters of the generator and discriminator.

[0114] In some embodiments, a discriminator is first trained to effectively distinguish between real and generated data; then, the discriminator is fixed to train the generator so that the generated map increment is closer to the real data; the training is repeated alternately until the generator and discriminator reach a game equilibrium and generate a high-quality map increment.

[0115] It can be seen that the alternating training mechanism allows the generator and discriminator to promote each other and iteratively optimize, resulting in a high-quality incremental map that accurately reflects the terrain changes in the mining area and provides reliable incremental data for map updates.

[0116] In this way, the generator and the discriminator learn through game theory. The generator continuously generates incremental map maps that closely resemble the real terrain, while the discriminator continuously improves its ability to distinguish between the real and generated maps. Ultimately, the generator can output high-quality incremental map maps for map updates.

[0117] In one embodiment of the present invention, based on step S105, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S105 specifically includes the following: S1051: Compare the newly acquired point cloud, image, and pose data with existing map data in multiple dimensions to identify areas of initial change.

[0118] In some embodiments, the newly acquired data includes preprocessed point cloud spatial coordinates (X,Y,Z), reflection intensity, image and vehicle pose data, which are compared with the corresponding data of the existing map point by point and pixel by pixel in dimensions such as height and area, and the areas with differences are marked.

[0119] S1052: Use a deep learning model to perform semantic classification on the preliminary changed regions and determine the region type.

[0120] In some embodiments, the region type can be a mining area, spoil heap, main road, etc. Based on the multimodal features obtained in step S103, the deep learning model performs semantic segmentation on the initially changed region and outputs the semantic category of the region.

[0121] S1053: Input the current mining area characteristic data and historical data into the prediction model, calculate the change Pchange using formula 25, and predict the future changes in the area.

[0122] In some embodiments, unlike traditional map updating methods, the incremental updating in this embodiment does not solely rely on difference data, but rather combines semantic information generated by deep learning with predicted data to achieve dynamic map updates. Traditional updating methods identify and update changed parts of the map based on differences in point clouds or images, while this embodiment, by introducing a deep learning model, can more intelligently select which areas need to be updated first based on a deep understanding and prediction of the mining area environment, thereby improving update efficiency and accuracy.

[0123] For example, when large-scale changes occur in the mining area, such as the excavation progress of the mining area or the expansion of the soil dumping area, the system can prioritize updating the changed areas based on the feedback mechanism of the deep learning model. After each new data collection, the system first compares it with the existing map to identify the changed areas and then updates them according to the following steps: The data comparison and change area identification in this embodiment identifies changed areas by comparing newly acquired data with the existing ground. These changed areas include not only changes in height but also changes in area.

[0124] Semantic information and predictive data, generated based on deep learning models (such as potential changes in the mining area over the next few hours), can further guide update strategies. For example, when a deep learning model predicts that the soil dumping area in a certain region may expand, the system can prioritize incremental updates for that region to ensure that the updated map accurately reflects the actual operational progress of the mining area.

[0125] Prediction formula:

[0126] Formula (25) in, It is the predicted change. This is the current characteristic data of the mining area. It is historical data. Predicting changes allows for forward-looking assessments of regional changes, enabling advance planning of update timing and scope.

[0127] S1054: Combine semantic classification results and predicted changes to determine the update priority and range of change for each changed region.

[0128] In some embodiments, based on semantic categories, such as the mining area being a high-frequency change area and the main road being a low-frequency change area, and the size of Pchange, an update priority is assigned to each change area, and the specific change range is determined by combining the data comparison results, such as the change range of the excavation area of ​​a certain mining area.

[0129] S1055: Integrate the priority, scope, and type information of all changed areas to clarify the final changed mining area.

[0130] In some embodiments, the priority, scope of change, and semantic type of all changed areas are summarized to form a list of changed areas, clarifying which areas need to be updated first and the extent of the update. Preliminary differences are discovered through multi-dimensional data comparison, regional attributes are clarified by combining deep learning semantic classification, and change trends are predicted using historical and current data, ultimately integrating precise mining area change regions.

[0131] In one embodiment of the present invention, based on step S106, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S106 specifically includes the following methods: S1061: Determine the target area for map quality assessment, covering both identified dynamic and static areas, and count the total number of assessment areas N.

[0132] In some embodiments, the target area for assessment is the mining area change area identified in S105, including dynamic areas such as mining and loading areas and spoil heaps, and static areas such as main roads. The total number of these areas is counted and recorded as N to ensure that the assessment scope completely matches the actual change area.

[0133] S1062: Collect dynamic area changes of each assessment area and changes in positioning accuracy parameter.

[0134] In some embodiments, The calculation is performed by comparing the spatial extent of the newly collected point cloud data with the corresponding dynamic area on the existing map, reflecting the increase or decrease in the area. By comparing the vehicle pose data collected by the current GNSS / IMU sensors with existing map positioning data, the change in positioning error is calculated to reflect the fluctuation of map positioning accuracy.

[0135] S1063: Adaptively adjust update frequency weights based on the degree of dynamic region change. The accuracy weights are adaptively adjusted according to the regional accuracy requirements. .

[0136] In some embodiments, Based on the dynamic area change rate The ratio to the total area of ​​the region is adjusted, and the change rate automatically increases when it exceeds a preset ratio. Take a value, and vice versa. Adjustments are made based on the frequency of train traffic and the importance of operations in the region. Higher values ​​are used for densely trafficked operation areas and important roads in mining areas, while lower values ​​are used for static areas such as main roads.

[0137] S1064: Substitute the parameters

[0138] Calculate the individual evaluation values ​​for each region, and then average them to obtain the overall map quality evaluation value. .

[0139] In some embodiments, prioritizing updates to areas with significant changes is based on deep learning feedback. The system prioritizes updating areas that change drastically, especially dynamically changing areas such as earthwork areas and mining areas. For these areas, the system selects higher-frequency updates to ensure that the autonomous driving system uses the latest map data during real-time operations.

[0140] Map quality assessment determines whether a map update is necessary. The assessment criteria primarily include: Update frequency: Mining areas or soil dumping areas have a higher frequency of change and therefore a higher update frequency, while static areas such as main roads have a lower frequency of change and therefore a lower update frequency.

[0141] Update accuracy: For areas with frequent train traffic, higher accuracy map updates are needed to ensure the accuracy of route planning and positioning.

[0142] The quality assessment formula is as follows:

[0143] Formula (26) in, It is a map quality assessment value. This represents the dynamic change in the area of ​​the region. To represent the variation in positioning accuracy, N is the number of areas being evaluated.

[0144] The weighting coefficient is used to represent the degree of impact of dynamic area changes within the mining area on map updates. Dynamic areas typically change frequently and have high update requirements; therefore, the system assigns higher weights to these areas to ensure timely updates. It will automatically adjust according to the degree of change. When the change is large, the weight coefficient will increase and the update frequency will increase.

[0145] The map update accuracy is used as a weighting factor to measure the impact of map update accuracy on the overall system operation. For areas with high accuracy requirements, such as important roads in work areas or mining areas, the system needs to provide higher map update accuracy.

[0146] The evaluation value is calculated quantitatively using formula (26), eliminating subjective judgment and ensuring the objectivity and consistency of the map quality evaluation results.

[0147] S1065: Set the threshold Q for differentiated quality assessment threshold ,Will The individual assessment values ​​for each region are compared with the corresponding thresholds to determine whether the update conditions are met.

[0148] In some embodiments, a predefined difference is established between dynamic and static regions, with the dynamic region having a lower threshold than the static region; the individual evaluation value of each region is compared with the corresponding region threshold. Compared with the global threshold, if the value is higher than the threshold, it is determined that the update condition is met; if the value is lower than the threshold, it is determined that the condition is not met.

[0149] This embodiment takes the changed area identified by S105 as the evaluation object, extracts the core parameters of area change and positioning accuracy change, matches the update needs and accuracy requirements of different areas through adaptive weights, quantifies and calculates the map quality evaluation value, and then judges it through differentiated thresholds to provide an objective decision basis for adaptive incremental updates.

[0150] In one embodiment of the present invention, based on step S107, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S107 specifically includes the following methods: S1071: Based on the types of changed areas identified in S105 and the quality assessment results in S106, the areas to be updated are divided into two categories: work areas and main roads.

[0151] In some embodiments, based on the semantic classification results of S105, such as mining and loading areas, spoil heaps, main roads, and individual evaluation values ​​of S106, areas that meet the update conditions are selected, and dynamic work areas and static main roads are clearly delineated to ensure that the update objects match the characteristics of the areas.

[0152] S1072: For the work area, based on the work type, train frequency, and S106. The work progress adjustment coefficient is dynamically determined. .

[0153] In some embodiments, The value ranges from 0.9 to 1.3, but is taken as 1.2 to 1.3 after loading operations due to more significant regional changes, and as 1.0 to 1.1 after unloading operations. For every 5 vehicles / hour increase in vehicle frequency, Increase by 0.1, while combining with S106 ,when If the value is greater than 0.7, adjust the value by an additional 0.1 to ensure... Adapt to work dynamics.

[0154] S1073: S105 And determined by S1072 Substitute into Formula 27 to calculate the area of ​​the updated work area. Based on the map incremental map of S104, perform vehicle-level incremental updates.

[0155] In some embodiments, based on the area changes and map quality assessment of the output dynamic region of the deep learning model, the system intelligently selects an incremental update strategy according to the needs of different environments, without needing to fully reconstruct the entire mining area each time. The main adaptive update strategies include: Work area vehicle-level updates: For frequently changing areas such as mining areas or soil dumping areas, real-time map updates are performed at the vehicle-level. These areas typically change frequently with the progress of operations and equipment movement. The update process adopts incremental updates, updating only the areas with significant changes to reduce the computational and transmission burden.

[0156] The updated area is related to changes in train frequency and is dynamically adjusted based on the work progress and train frequency.

[0157] Formula (27) in, It updates the area. It is an updated adjustment coefficient based on the work progress to adapt to frequent changes at the train number level.

[0158] In this embodiment, the corresponding area map incremental map generated in S104 is called, and only the changed parts in the existing map that match the incremental map are replaced, without changing the data of the unchanged areas.

[0159] S1074: For main roads, based on their historical change data and S106 Set adaptive update threshold The formula is determined by updating the threshold:

[0160] Determine if the update conditions are met; if so, perform an incremental update.

[0161] In some embodiments, for main roads and other static areas, the system determines whether to update the map by setting an adaptively adjustable change threshold. If the area change of the main road is below the threshold, no update is performed; only when the change exceeds the threshold is an incremental update performed. This method reduces unnecessary computation of static areas.

[0162] Update the threshold determination formula:

[0163] Formula (28) in, It refers to the change in the area of ​​the main road region. This is the set update threshold. The adaptive threshold is adapted to the stability and accuracy requirements of the main road, reducing invalid update operations in static areas while ensuring timely updates for necessary changes.

[0164] The adaptive update strategy is based on the system intelligently determining which areas need updating and which can be skipped, thus significantly reducing the computational burden, especially for large-area mining scenarios. The specific process is as follows: Incremental updates for dynamic areas involve using incremental updates for mining and loading areas and soil stockpiling areas, updating only the parts with significant area changes.

[0165] For static areas, the system updates the main road. For relatively stable areas such as the main road, the system determines whether to update based on a threshold. If the area changes little, no update is performed; if the change is significant, an incremental update is performed.

[0166] Global map correction is based on the fact that after multiple incremental updates are completed, the system will integrate all local updates through a global map correction algorithm to ensure the consistency and accuracy of the map.

[0167] It should be noted that, The initial value is set at 5㎡, and dynamically adjusted based on the historical 30-day average change area of ​​the main road. When the average change area is >3㎡, the threshold is lowered to 4㎡; when it is <1㎡, it is raised to 6㎡. This is combined with S106... ,when If the threshold is greater than 0.8, the threshold is lowered by another 10%; this is determined using formula 28. The change in the area of ​​the main road identified by S105 and Relationship, update flag Incremental update is performed when =1.

[0168] S1075: Record update information for two types of areas, including the coordinate range of the updated area, update time, map increment map identifier used, and update result verification data.

[0169] In some embodiments, the update information log includes a unique regional ID, the three-dimensional coordinate boundary of the update range, the update trigger time, the map increment map number used, the positioning accuracy data before and after the update, and the area change verification value, providing a traceability basis for global correction in S108.

[0170] Based on the regional classification in S105 and the quality assessment results in S106, this embodiment adopts a differentiated incremental strategy of dynamically adjusting the update range at the vehicle level and judging the update based on adaptive thresholds for high-frequency changing work areas and low-frequency stable main roads, respectively. By using the map incremental map in S104 to update only the changed parts, accurate and efficient local map updates are achieved.

[0171] In one embodiment of the present invention, based on step S108, the following will provide a possible embodiment and its specific implementation will be described in a non-limiting manner. S108 specifically includes the following methods: S1081: Collect all spatial data and update records of local incremental updates.

[0172] In some embodiments, when collecting partial update data, all update logs recorded in step S107 are integrated, including those calculated using formula (27). The determined work area update range and the update determination by formula (28) flag The main road update region is triggered. Simultaneously, corresponding multimodal feature data is collected to provide a complete input dataset for global correction.

[0173] S1082: Detect and handle overlaps and conflicts between adjacent update regions.

[0174] In some embodiments, when detecting overlapping conflicts, a unified coordinate system is established using the sensor extrinsic parameters calibrated by formula (2) to identify the boundary overlap of adjacent update areas. For overlapping areas, the multimodal features obtained by formula (22) in different update batches are compared, and the data with higher feature consistency is selected as the benchmark. By detecting and processing regional overlapping conflicts, data contradictions that may arise from multiple updates are eliminated, ensuring the logical consistency of the map.

[0175] S1083: Map data fusion based on the principle of spatial consistency.

[0176] In some embodiments, the feature pyramid construction principle is applied during the data fusion process to hierarchically fuse locally updated data at different scales, ensuring spatial consistency of multi-resolution features. This hierarchical data fusion method maintains the spatial coherence of multi-scale features and improves the overall quality of map data.

[0177] S1084: Apply a global optimization algorithm to correct map geometric deformation.

[0178] In some embodiments, in geometric deformation correction, the optimization method of formula (2) is used to perform nonlinear optimization on the global map to correct the cumulative geometric error that may be caused by multiple local updates and maintain the overall geometric accuracy of the map.

[0179] S1085: Generate a new version of global map data with consistent topology.

[0180] In some embodiments, when generating a new version of the map, all corrected local update data are integrated to establish a topologically consistent map data structure. The consistency of data in each region of the map after normalization by formulas (3) and (4) is verified to ensure that the new version of the global map meets the requirements for use by the autonomous driving system.

[0181] This embodiment systematically integrates data generated from multiple local incremental updates, detects and resolves conflicts between regions, performs data fusion and geometric correction based on the principle of spatial consistency, and finally generates a new version of the global map with topological consistency and geometric accuracy, ensuring the integrity and reliability of the map data.

[0182] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0183] like Figure 7 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, a communication module 104, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of an adaptive incremental update method for open-pit mine maps.

[0184] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0185] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0186] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0187] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0188] The communication module 104 transmits radio signals to and / or receives radio signals from at least one of a base station, an external terminal, and a server. Such radio signals may include voice call signals, video call signals, or various types of data sent and / or received according to text and / or multimedia messages.

[0189] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the adaptive incremental update method for the open-pit mine map.

[0190] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0191] The storage medium stores a program product capable of implementing the methods described above in this specification. In some possible implementations, various aspects of this disclosure can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the exemplary methods section of this specification according to various exemplary embodiments of this disclosure.

[0192] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for adaptive incremental updating of an open pit mine site map, characterized by, The method comprises: S101: installing a collection device on a mining vehicle, collecting data of a road area and a working area by using an adaptive collection strategy, performing time synchronization, space calibration, and normalization preprocessing of point cloud and image data; S102: performing hierarchical feature extraction on the preprocessed data, processing point cloud and image data of near, medium, and long distance layers respectively, and constructing a feature fusion pyramid; S103: on the basis of the feature fusion pyramid, weighting the features by using a channel attention mechanism and a spatial attention mechanism, fusing image and point cloud features to obtain multi-modal feature representation; S104: optimizing the multi-modal features by using a generative adversarial network to generate a high-quality map increment map; S105: comparing newly collected data with an existing map, combining semantic information and predicted data generated by a deep learning model, and identifying a changed area in a mining area; S106: according to changes in the area of a dynamic area and the positioning accuracy, combining an update frequency weight and an accuracy weight, and performing quality assessment on the map; S107: based on the change area identification result and the map quality assessment result, using an adaptive incremental update strategy, performing a vehicle-level update on the working area and a threshold update on the main road; S108: performing global correction on the map after multiple incremental updates to obtain a new version of the global map.

2. The method of claim 1, wherein, S102 specifically comprises the following methods: processing near distance layer data, extracting local features from high-resolution point cloud using PointNet++ network, extracting high-resolution detail features from images using HRNet, and splicing to obtain near distance layer fusion features; processing middle distance layer data, voxelizing point cloud and extracting features using VoxelNet, extracting features from images using HRNet, and splicing to obtain middle distance layer fusion features; processing long distance layer data, projecting sparse point cloud into 2D image using spherical coordinate projection, extracting features using 2D CNN, and obtaining long distance layer fusion features; performing upsampling and downsampling operations on the fusion features of the near, middle, and long distance layers, and splicing to construct a preliminary feature fusion pyramid; alternately performing feature map fusion on the preliminary feature fusion pyramid in a top-down and bottom-up manner to obtain multi-scale features.

3. The method of claim 1, wherein, S103 specifically comprises the following methods: performing global average pooling on the point cloud multi-scale feature maps in the feature fusion pyramid to extract global information of each channel; inputting the global information into a fully connected layer, generating channel attention weights through Sigmoid activation; performing global average pooling and maximum pooling on the point cloud multi-scale feature maps in the feature fusion pyramid, and splicing to obtain spatial description features; performing convolution operation on the spatial description features, and generating spatial attention weights through Sigmoid activation; weighting image and point cloud features using channel and spatial attention weights respectively, and adding and fusing the weighted features to obtain multi-modal feature representation.

4. The method of claim 1, wherein, S104 specifically comprises the following methods: S1041: constructing a generator network, inputting multi-modal features, and generating a map increment map through a multi-layer network structure; The discriminator network is constructed to input real map data or the map delta generated by the generator, and output a discrimination result of the input data; The generator loss function is set to use a cross-entropy loss function to make the generated map delta close to the real map; The discriminator loss function is set to maximize the discrimination between the real map and the generated map delta; The generator and the discriminator are alternately trained to optimize the parameters of the generator and the discriminator through iteration, and a high-quality map delta is obtained.

5. The method of claim 1, wherein, S105 specifically includes the following modes: The newly collected point cloud, image and pose data are compared with the existing map data in multiple dimensions to identify the preliminary change area; The deep learning model is used to perform semantic classification on the preliminary change area to determine the region type; Input the current mine area characteristic data and historical data to the prediction model, and calculate the change amount by the following formula to predict the future change of the area. The prediction formula is: wherein, is the predicted change amount, is the current mine site characteristic data, is historical data; Combined with the semantic classification result and the predicted change, the update priority and change range of each change area are determined; The priority, range and type information of all change areas are integrated to determine the final mine area change area.

6. The method of claim 1, wherein, S106 specifically includes the following modes: Determine the target area for map quality evaluation, which includes the identified dynamic area and static area, and count the total number N of evaluation areas; Collecting dynamic area size changes of each evaluation area and positioning accuracy changes parameters; Adaptively adjusting update frequency weight according to dynamic region change degree Adaptively adjusting precision weight according to region precision requirement level ; Substitute the parameters into The individual regional evaluation values are calculated and averaged to obtain the global map quality evaluation value ; for dynamic area size variation, for positioning accuracy variation, N is the number of areas evaluated; Setting differentiated quality evaluation threshold Q threshold , if and each regional single evaluation value is compared with the corresponding threshold value, and it is determined whether the update condition is met.

7. The method of claim 1, wherein, S107 specifically includes the following modes: According to the change area type identified by S105 and the quality evaluation result of S106, the area to be updated is divided into two categories: work area and main road; For the work area, according to the work type, the train density and the S106 work, determine the work progress adjustment coefficient ; The S105 is and the determined Substitute the following formula to calculate the area of the update area of the operation area Perform train sub-incremental update based on the map increment map of S104; wherein, is an update area, is an update adjustment coefficient according to the progress of the work; For the main road, according to its historical change data and S106 , the adaptive update threshold is set , and the update threshold judgment formula is determined Determine whether the update condition is met, and if so, perform incremental update; Record the update information of the two types of areas, including the update area coordinate range, update time, map delta identifier used and update result verification data.

8. The method of claim 1, wherein, S108 specifically includes the following modes: Collect all local incremental update spatial data and update records; Detect and handle the overlap and conflict between adjacent update areas; Based on the principle of spatial consistency, the map data is fused; Apply a global optimization algorithm to correct the geometric deformation of the map; Generate a new version of topologically consistent global map data.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the open-pit mine map adaptive incremental update method according to any one of claims 1 to 8.

10. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the open-pit mine map adaptive incremental update method according to any one of claims 1 to 8.

Images