Underground three-dimensional high-precision map construction method based on semantic segmentation and neural radiance field

Abstract

The application discloses a downhole three-dimensional high-precision map construction method based on semantic segmentation and neural radiation field, comprising the following steps: obtaining downhole multi-view images and sparse depth data, and carrying out denoising, defogging and pose estimation processing; performing pixel-by-pixel semantic classification on the processed images through a semantic segmentation network, and assigning a semantic label to each pixel; performing dynamic and static separation by using a semantic mask, and performing degradation correction on the static area; constructing a neural radiation field model based on hierarchical voxel decomposition, and performing joint training by using a scattering-aware multi-task loss function; and rendering a three-dimensional semantic high-precision map by using a dynamic object gradual fusion strategy and a local incremental update mechanism. Through the collaborative mechanisms of semantic-guided scattering correction and dynamic and static separation, hierarchical voxel decomposition, scattering-aware multi-task joint loss, dynamic object gradual fusion and local incremental update, the application realizes a high-fidelity, lightweight and continuously updated downhole three-dimensional semantic map.

Description

Technical Field

[0001] This invention relates to the field of intelligent mine mapping and 3D vision technology, specifically to a method for constructing a high-precision 3D underground map based on semantic segmentation and neural radiation field. Background Technology

[0002] High-precision 3D underground maps are the foundation of smart mines, playing an irreplaceable role in tasks such as unmanned transportation, equipment positioning and navigation, post-disaster rescue, and mine digital twins. Traditional underground mapping methods mainly rely on simultaneous localization and mapping (SLAM) with lidar or multi-view vision reconstruction. While lidar offers accurate ranging, it is costly, consumes a lot of power, and suffers from severe point cloud sparsity in dusty environments. Traditional vision methods are highly dependent on lighting and texture, making stable operation difficult in low-light, repetitive texture tunnels. In recent years, Neural Radiation Field (NeRF) has emerged as a new paradigm for outdoor and indoor scene reconstruction due to its continuous implicit scene representation and superior viewpoint synthesis capabilities, demonstrating enormous application potential in the field of 3D reconstruction.

[0003] However, the application of standard NeRF in complex underground environments still faces the following prominent problems. Dust and water mist are prevalent in underground environments, leading to decreased image contrast, color shifts, and blurred details. Standard NeRF optimizes for photometric consistency based on clear input images; degraded images directly cause reconstruction errors in the density and color fields, resulting in distorted tunnel contours and lost textures. Simultaneously, dynamic objects (such as moving mine carts, workers, and mining equipment) frequently appear within tunnels. The static world assumption implicit in NeRF cannot handle these non-rigid movements, resulting in ghosting or trailing images in the rendering results, and dynamic objects are incorrectly fixed as permanent structures in the map. Furthermore, mining electrical equipment must meet intrinsically safe explosion-proof requirements, and the computing and storage resources of edge computing platforms are severely limited. Ordinary NeRF models often have hundreds of megabytes of parameters, requiring hours of training and only a few frames per second for rendering, which is simply insufficient to meet the needs of online mapping and real-time updates in underground environments. Existing methods lack deep integration of semantic information and radiation field. The reconstructed map only contains appearance and geometry and cannot directly distinguish semantic categories such as alley structure, pipelines, and dynamic objects, making it difficult to support map-based intelligent decision-making and path planning.

[0004] Based on the above analysis, the main challenges in achieving this task are: 1) How to effectively resist visual degradation caused by underground dust and water mist, so that NeRF can still accurately reconstruct the tunnel structure under degradation conditions; 2) How to separate and manage dynamic objects to avoid them interfering with the construction of static maps; 3) How to achieve lightweight deployment and real-time updates of the model under the condition of strictly limited edge device resources; 4) How to integrate semantic information in the radiation field to enable the map to have scene understanding and intelligent decision support capabilities. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field. Through a series of collaborative mechanisms, such as semantically guided scattering correction and dynamic-static separation, hierarchical voxel decomposition, scattering perception multi-task joint loss, dynamic object gradual fusion, and local incremental update, a high-fidelity, lightweight, and continuously updated 3D downhole semantic map is achieved.

[0006] The present invention adopts the following technical solution: a method for constructing a three-dimensional high-precision map of underground mine based on semantic segmentation and neural radiation field, including the following steps: S1, acquiring underground multi-view images and sparse depth data, and preprocessing the images; constructing a semantic segmentation network, performing pixel-by-pixel semantic classification on the preprocessed images, and assigning semantic labels to each pixel; after processing by the semantic segmentation network, obtaining dynamic object masks, static tunnel masks and scattering medium parameters.

[0007] S2. Based on the dynamic object mask and the static alleyway mask, the scene is separated into dynamic and static states, and the static area image is degraded by combining the scattering medium parameters.

[0008] S3. Construct a hierarchical voxel decomposition neural radiation field model based on semantic importance and pre-train it until convergence.

[0009] The hierarchical voxel decomposition divides the three-dimensional space into non-uniform voxel partitions based on semantic importance. The features encoded by each voxel are mapped to volume density and color values ​​through a network to characterize the high-precision radiation field of the downhole scene.

[0010] The neural radiation field model is used to encode a three-dimensional scene into a continuous volume density field and color field, and synthesize images from any new perspective through volume rendering to perform high-precision three-dimensional reconstruction of the underground tunnel structure.

[0011] S4. Construct a multi-task loss function for scattering perception, including photometric loss, semantic loss, scattering depth loss, and temporal consistency loss; perform weighted joint training on the four losses to enable the 3D reconstruction network to synchronously learn the radiation field representation, semantic structure, spatial geometry, and dynamic temporal variation patterns of the scene in the downhole environment.

[0012] S5. A dynamic object gradual fusion strategy is adopted to gradually integrate dynamic objects that have been stationary for a long time into the static map, while keeping objects that move briefly hidden. The local incremental update mechanism and elastic weights are used to consolidate the constraints, and local fine-tuning is performed on the changed areas to render and generate a high-precision 3D semantic map.

[0013] Furthermore, the preprocessing in step S1 includes: defogging: using a dark channel prior algorithm to perform defogging to eliminate the contrast reduction caused by dust and water mist.

[0014] Denoising: Denoising is performed using guided filtering.

[0015] Calculate camera pose parameters; calculate the camera position coordinates for each image using COLMAP software. and observation direction ;in, Indicates the direction of observation and The angle between the axes, Indicates the direction of observation and The angle between axes.

[0016] Furthermore, the semantic segmentation network described in step S1 employs U... 2 -Net or DeepLabV3+ backbone architecture, including a shared encoder, classification head and regression head.

[0017] Specifically, the network extracts common image features through a shared encoder, and then sets up two processing paths: one is a classification path, which predicts the semantic category of each pixel and generates dynamic object masks and static alleyway masks accordingly; the other is a regression path, which simultaneously estimates the scattering medium parameters corresponding to each pixel, including the transmittance map and atmospheric light value.

[0018] The input image is semantically classified pixel by pixel using a classification head, and semantic labels are output, including categories of dynamic objects such as mine cars, personnel, and mining equipment, as well as categories of static structures such as tunnel walls, support structures, pipelines, and floor slabs. Dynamic object masks and static tunnel masks are derived from these labels.

[0019] The scattering medium parameter map, transmittance distribution map, and global atmospheric light value are output synchronously through the regression head.

[0020] Furthermore, the degradation correction described in step S2 involves processing the image of the static tunnel using the scattering medium parameters obtained in S1, and then analyzing the transmittance map. and atmospheric light value The atmospheric scattering model is solved pixel-by-pixel for static region images to obtain clear images after degradation correction to remove the influence of downhole scattering media. After this correction process, clear images from multiple perspectives are output. This ensures that the color of the same physical point is consistent in radiometrics from different perspectives, thus satisfying the radiometric consistency requirement for NeRF training.

[0021] Furthermore, the neural radiation field model of hierarchical voxel decomposition described in step S3 includes the following processing: S3.1, performing non-uniform voxel partitioning of the three-dimensional space according to semantic importance: activating dense voxels in the textured alleyway wall and support structure regions at a high resolution level, while preserving fine geometry.

[0022] By sparsely distributing voxels in uniform tunnel cross-sections and open areas with single textures at low resolution levels, computational overhead is reduced.

[0023] S3.2. Using the feature vector encoded by each voxel position as input, it is mapped to the volume density σ and color value c in the observation direction through a multi-layer fully connected network. A new perspective is synthesized through volume rendering, and the three-dimensional scene is encoded into a continuous volume density field and color field, so as to realize high-precision three-dimensional reconstruction of the synthesized image from any new perspective.

[0024] Further, the multi-task loss function described in step S4 is expressed as: : ;in, Photometric loss measures the radiometric consistency error between network-rendered colors and real-world images. The semantic loss is used to measure the consistency between the volume-rendered semantic graph and the two-dimensional semantic segmentation labels. To compensate for the scattering depth loss, the deviation between the depth estimate and the true sparse depth is measured, and a transmittance-based scattering compensation term is embedded to correct the depth estimation bias caused by dust. The temporal consistency loss measures the smoothness of rendering results between adjacent frames; , , , These are the weighting coefficients for photometric loss, semantic loss, scattering depth loss, and temporal consistency loss, used to balance the dimensional differences and contribution ratios of each loss term.

[0025] Furthermore, the dynamic object fading and blending strategy described in step S5 sets a timer for each dynamic object. When the duration of a certain object's continuous stillness exceeds a threshold At that time, it is removed from the dynamic mask and incorporated into the static voxel update.

[0026] The three-dimensional semantic high-precision map integrates semantic information and geometric structure, eliminates interference from moving objects through dynamic object management, and continuously maintains high map accuracy and real-time performance through local incremental update mechanism and adaptive fine-tuning for new scenes.

[0027] Furthermore, the elastic weight consolidation constraint described in step S5 calculates the Fisher information matrix for historical parameters during incremental updates, applying a larger penalty factor to parameters with higher importance to prevent catastrophic forgetting. The constraint term is: ;in, This is the i-th diagonal element of the Fisher information matrix. The current parameter value. These are the historically optimal parameters; To balance the weighting coefficients, The loss term is consolidated with elastic weights.

[0028] The present invention also provides: an electronic device, comprising: one or more processors; a storage device storing one or more programs thereon; when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for constructing a three-dimensional high-precision map of a well based on semantic segmentation and neural radiation fields.

[0029] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the aforementioned method for constructing a three-dimensional high-precision map of a well based on semantic segmentation and neural radiation fields.

[0030] Compared with existing technologies, the present invention, employing the above technical solutions, has the following technical effects: The method of the present invention introduces semantically guided hierarchical voxel decomposition into the neural radiation field. The network can adaptively allocate computing resources according to the semantic complexity of the scene, significantly reducing the number of model parameters and computational overhead while ensuring reconstruction accuracy, and reducing inference latency to a level usable at the edge. Through a dynamic object fading fusion strategy and elastic weight consolidation constraints, the map can continuously manage dynamic interference during long-term operation, while avoiding catastrophic forgetting of learned knowledge. The final generated downhole 3D semantic high-precision map has geometric accuracy, semantic understanding ability, and real-time update capability, which can effectively support applications such as downhole autonomous driving, remote control, emergency rescue, and digital twins. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall process of constructing a three-dimensional high-precision map of a mine shaft according to the present invention.

[0032] Figure 2 This is a flowchart of the semantically guided scattering degradation correction and dynamic-static separation process of the present invention.

[0033] Figure 3 This is a flowchart of the hierarchical voxel decomposition and multi-task training structure of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0035] This invention provides a method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field. The main process is as follows: Figure 1 As shown, the input is multi-view images and sparse depth data from the well. The processing includes: S1, image preprocessing and semantic segmentation to obtain dynamic and static masks and medium parameters: the input image is preprocessed, including denoising, dehazing and calculating camera pose parameters, including camera position and observation direction; a semantic segmentation network is constructed to perform pixel-by-pixel semantic classification on the preprocessed image to obtain dynamic object masks, static tunnel masks and scattering medium parameters.

[0036] S2. Scene dynamic separation and degradation correction: The scene is separated into dynamic and static parts based on the dynamic object mask and the static alleyway mask, and the static area image is degraded by combining the scattering medium parameters.

[0037] S3. Pre-training the hierarchical voxel decomposition neural radiation field model: Construct a hierarchical voxel decomposition neural radiation field model based on semantic importance, and pre-train it until convergence.

[0038] S4. Consider multi-task loss for training the 3D reconstruction network: Design a scattering-aware multi-task loss function to train the 3D reconstruction network. The multi-task loss function includes photometric loss, semantic loss, scattering depth loss and temporal consistency loss.

[0039] S5. Dynamic Fusion and Incremental Update: A dynamic object gradual fusion strategy is adopted to gradually integrate dynamic objects that have been stationary for a long time into the static map, while keeping objects that move briefly hidden. The local incremental update mechanism and elastic weights are used to consolidate constraints and make local fine adjustments to the changed areas to render and generate a high-precision 3D semantic map.

[0040] In one embodiment of the present invention, a main underground transport roadway in a coal mine is used as the modeling object. The roadway has a semi-circular arched cross-section, a net width of 5.2 m, a net height of 3.8 m, and cables, compressed air and drainage pipes are hung on both sides, with double rails laid in the middle. The mobile acquisition platform is equipped with an explosion-proof RGB camera and a LiDAR, and the edge computing unit is an NVIDIA Jetson Orin AGX (15W power consumption mode).

[0041] In this embodiment, the semantically guided scattering degradation correction and dynamic-static separation process is as follows: Figure 2 As shown, step S1 specifically includes: the vehicle traveling at a speed of 1.5 m / s continuously acquiring video streams. A dark channel prior dehazing algorithm is applied to each frame of the image, and a reverse solution is performed based on the atmospheric scattering model. Where I(x) is the pixel value of the original observed image, A is the atmospheric light value, t(x) is the transmittance, and t0 is the lower limit of transmittance, set to 0.1 to prevent division by zero.

[0042] After dehazing, histogram equalization is performed simultaneously to enhance details in dark areas. Then, COLMAP is used to extract feature points and IMU data is fused to estimate the pose of each frame. The LiDAR point cloud is projected onto the image plane to obtain a sparse depth map.

[0043] Furthermore, the semantic segmentation network mentioned in step S1 is specifically a MobileNetV3-DeepLabV3+ network with channel compression to 1 / 4 and 8-bit quantization, with the structure as follows: Figure 2 As shown, single-frame inference takes approximately 12 ms; the network outputs a dynamic mask M. dyn (Mining car, personnel), static mask M sta (Rock wall, steel frame, track), and simultaneously predict transmittance correction t(x) and atmospheric light value A.

[0044] Furthermore, the dynamic-static separation and degradation correction described in step S2 are performed as follows: For each sampling point on a camera ray r, if its projected pixel belongs to the dynamic mask M... dyn If the weight w=0, it will not participate in the optimization; if it belongs to the static mask M sta Then, inverse correction is performed according to the atmospheric scattering model: ; where J(x) is the corrected sharp pixel value. This process completely masks dynamic objects, restores the tunnel texture, and calculates the photometric loss using the corrected image J(x).

[0045] Furthermore, the neural radiation field model of hierarchical voxel decomposition described in step S3 is as follows: Figure 3 As shown, the specific structure is as follows: the scene is divided into three layers based on the semantic mask: M sta The voxel side length is 0.08 m (L1), the voxel side length of other static areas is 0.25 m (L2), and the voxel side length of dynamic and open areas is 0.8 m or directly pruned (L3).

[0046] Vector-matrix low-rank tensor decomposition is applied to L1 and L2 voxel meshes: Where r is the rank index, It is a globally shared vector. , , These are the factor vectors on three orthogonal planes. This indicates the outer product operation.

[0047] As a preferred embodiment, in this case, the rank is taken. After training, the final model size is only 18MB, and the rendering speed is 155FPS.

[0048] Furthermore, the multi-task loss function described in step S4 is implemented as follows: the total training loss of the network is defined as... : ;in, , , , These are the weighting coefficients for photometric loss, semantic loss, scattering depth loss, and temporal consistency loss, used to balance the dimensional differences and contribution ratios of each loss term. and Determined by grid search on the validation set, , In engineering, there are usually fixed values ​​that can be taken.

[0049] As a preferred embodiment, the weights are respectively set as follows: , , , .

[0050] Only the photometric error after scattering correction in the static region is calculated; KL divergence is used to make the volume rendering semantic graph approximate two-dimensional segmentation labels; Constrain the normals of adjacent frames to be consistent.

[0051] Includes scattering depth compensation term: ;in, For compensation coefficient, For the depth predicted by the network, For true depth.

[0052] As a preferred option, the scattering depth compensation term The calculation is as follows: Where t(s) is the transmittance along the light path s, and this integral compensates for the deviation in depth estimation caused by scattering.

[0053] Furthermore, the dynamic object fading fusion strategy described in step S5 specifically operates as follows: recording the number of consecutive moving frames for each dynamic instance. and static frame rate Define fusion weights : Where σ is the Sigmoid function with parameter τ = 0.8. When At that time, the voxel features of the instance are slowly integrated into the static map. In this embodiment, the temporarily parked mining car is naturally absorbed after about 35 seconds.

[0054] Furthermore, the local incremental update and elastic weight consolidation constraint described in step S5 are performed as follows: When a new keyframe arrives, a sliding window is constructed using 5 adjacent frames, optimizing only the pose and changing voxels within the window, while freezing the remaining parameters. An elastic weight consolidation loss is introduced: ;in, For the elastic weighting consolidation loss term, The current parameter value. To train the converged historical optimal parameter values, F i For the corresponding diagonal elements of the Fisher information matrix, These are the balancing weighting coefficients.

[0055] Furthermore, for any new viewpoint, color is calculated separately through volume rendering. and semantics : ; ;in, The cumulative transmittance along the light ray is... For bulk density, The color value of a spatial point in the direction of observation. This represents the semantic value corresponding to a spatial point.

[0056] This embodiment implements the above method using PyTorch and conducts experiments on the NVIDIA Jetson Orin AGX platform. The network is trained using the Adam optimizer with an initial learning rate of 5×10⁻⁶. -4 A total of 200,000 iterations were performed. The evaluation metrics used included: root mean square error (RMSE) for geometric accuracy and mean intersection-union ratio (mIoU) for semantic accuracy.

[0057] Experimental results show that in this embodiment, the geometric accuracy RMSE is 7.2 mm, the semantic accuracy mIoU is 83.1%, and the entire pipeline operates in real time.

[0058] The above embodiments fully verify the effectiveness and practicality of the present invention in different restricted scenarios in mines, and can provide a high-precision, lightweight and continuously updated three-dimensional semantic map construction solution for smart mines.

[0059] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field, characterized in that, Includes the following steps: S1. Acquire multi-view images and sparse depth data from the well, and preprocess the images; construct a semantic segmentation network to perform pixel-by-pixel semantic classification on the preprocessed images and assign semantic labels to each pixel; After processing by the semantic segmentation network, dynamic object masks, static tunnel masks, and scattering medium parameters are obtained. S2. Based on the dynamic object mask and the static alleyway mask, the scene is separated into dynamic and static states, and the static area image is degraded by combining the scattering medium parameters. S3. Construct a hierarchical voxel decomposition neural radiation field model based on semantic importance and pre-train it until convergence. The hierarchical voxel decomposition divides the three-dimensional space into non-uniform voxel partitions based on semantic importance. The features encoded by each voxel are mapped to volume density and color values ​​through a network to characterize the high-precision radiation field of the downhole scene. The neural radiation field model is used to encode a three-dimensional scene into a continuous volume density field and color field, and synthesize images from any new perspective through volume rendering to perform high-precision three-dimensional reconstruction of the underground tunnel structure. S4. Construct a multi-task loss function for scattering perception, including photometric loss, semantic loss, scattering depth loss, and temporal consistency loss; perform weighted joint training on the four losses to enable the 3D reconstruction network to synchronously learn the radiation field representation, semantic structure, spatial geometry, and dynamic temporal variation rules of the scene in the downhole environment. S5. A dynamic object gradual fusion strategy is adopted to gradually integrate dynamic objects that have been stationary for a long time into the static map, while keeping objects that move briefly hidden. The local incremental update mechanism and elastic weights are used to consolidate the constraints, and local fine-tuning is performed on the changed areas to render and generate a high-precision 3D semantic map.

2. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 1, characterized in that, The preprocessing in step S1 includes: Defogging: A dark channel prior algorithm is used for defogging to eliminate the contrast reduction caused by dust and water mist; Denoising: Denoising is performed using guided filtering; Calculate camera pose parameters; calculate the camera position coordinates for each image using COLMAP software. and observation direction ;in, Indicates the direction of observation and The angle between the axes, Indicates the direction of observation and The angle between axes.

3. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 1, characterized in that, The semantic segmentation network described in step S1 uses U 2 -Net or DeepLabV3+ architecture, including a shared encoder, classification head and regression head; The classification head performs pixel-by-pixel semantic classification on the input image and outputs semantic labels, including dynamic object categories related to mine cars, personnel, and mining equipment, as well as static structure categories related to roadway walls, support structures, pipelines, and floor slabs, and derives dynamic object masks and static roadway masks. The regression head synchronously outputs a scattering medium parameter map, which includes a transmittance map and atmospheric light values.

4. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 3, characterized in that, The degradation correction in step S2 involves processing the image of the static tunnel using the scattering medium parameters obtained in S1, and then analyzing the transmittance map. and atmospheric light value By solving the atmospheric scattering model pixel by pixel in the static region image, a clear image is obtained after degradation correction to remove the influence of the downhole scattering medium. : ； in, For observation images, corrected multi-view clear images The consistency constraint is satisfied in the radiometric measurement.

5. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 1, characterized in that, The neural radiation field model of hierarchical voxel decomposition described in step S3 includes the following processing: S3.

1. Based on semantic importance, the three-dimensional space is divided into non-uniform voxels. The textured tunnel walls and support structures are activated with dense voxels at a high resolution level to preserve fine geometry. The textured uniform tunnel cross sections and open areas are sparsely distributed with voxels at a low resolution level to reduce computational overhead. S3.

2. Using the feature vector encoded by each voxel position as input, the feature vector is mapped to volume density and color value through a multi-layer fully connected network. The 3D scene is encoded into a continuous volume density field and color field through volume rendering, realizing high-precision 3D reconstruction of the synthesized image from any new perspective.

6. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 1, characterized in that, The multi-task loss function described in step S4 is expressed as follows: : ； in, Photometric loss measures the radiometric consistency error between network-rendered colors and real-world images. The semantic loss is used to measure the consistency between the volume-rendered semantic graph and the two-dimensional semantic segmentation labels. To compensate for the scattering depth loss, the deviation between the depth estimate and the true sparse depth is measured, and a transmittance-based scattering compensation term is embedded to correct the depth estimation bias caused by dust. The temporal consistency loss measures the smoothness of rendering results between adjacent frames; , , , These are the weighting coefficients for photometric loss, semantic loss, scattering depth loss, and temporal consistency loss, used to balance the dimensional differences and contribution ratios of each loss term.

7. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 1, characterized in that, The dynamic object fading and blending strategy described in step S5 involves setting a timer for each dynamic object. When the duration of a certain object's continuous stillness exceeds a threshold When this happens, remove it from the dynamic mask and incorporate it into the static voxel update; The three-dimensional semantic high-precision map integrates semantic information and geometric structure, eliminates interference from moving objects through dynamic object management, and continuously maintains high map accuracy and real-time performance through local incremental update mechanism and adaptive fine-tuning for new scenes.

8. The method for constructing a high-precision 3D downhole map based on semantic segmentation and neural radiation field according to claim 7, characterized in that, The elastic weight consolidation constraint described in step S5 calculates the Fisher information matrix for historical parameters during incremental updates, applying a larger penalty factor to parameters with higher importance to prevent catastrophic forgetting. The constraint term is as follows: ； in, This is the i-th diagonal element of the Fisher information matrix. The current parameter value. These are the historically optimal parameters; To balance the weighting coefficients, The loss term is consolidated with elastic weights.

9. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing a downhole 3D high-precision map based on semantic segmentation and neural radiation field as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps in the method for constructing a downhole 3D high-precision map based on semantic segmentation and neural radiation field as described in any one of claims 1 to 8.

Images