An uncertainty-driven robust mapping method and system for an indoor scene, a terminal and a storage medium

Abstract

The present application relates to the technical field of three-dimensional scene reconstruction and unmanned aerial vehicle multi-sensor fusion mapping, and discloses an uncertainty-driven robust mapping method and system for indoor scenes, a terminal and a storage medium, wherein the present application takes an image sequence, pose information and depth or point cloud observation for establishing geometric prior as input, constructs a dual-representation joint mapping architecture composed of an explicit Gaussian surface representation and an implicit signed distance field representation, the explicit branch is used for undertaking an efficient rendering task, and the implicit branch is used for providing continuous geometric manifold, surface distance and local confidence representation, so that appearance expression capability and geometric availability are considered in the same system. The present application can improve rendering fidelity, boundary structure stability, detail recovery capability and training convergence robustness under an indoor unseen view angle.

Description

Technical Field

[0001] This invention relates to the field of 3D scene reconstruction and UAV multi-sensor fusion mapping technology, and in particular to an uncertainty-driven robust mapping method, system, terminal and computer-readable storage medium for indoor scenes. Background Technology

[0002] Existing high-fidelity 3D indoor mapping solutions typically follow three approaches: visual reconstruction, depth or point cloud geometric modeling, and explicit-implicit hybrid representation. Visual reconstruction can restore relatively good texture appearance, but it lacks sufficient geometric constraints under conditions such as large areas of weak texture planes, specular reflection, exposure variations, and local occlusion, easily leading to boundary blurring, local stretching, and texture drift from unseen viewpoints. Depth or point cloud geometric modeling offers strong structural observability, but it has limitations in photorealistic rendering capabilities, detailed texture restoration, and adaptability to complex lighting. The hybrid mapping approach, which uses explicit Gaussian representation combined with signed distance fields to construct the scene, can balance appearance representation with continuous geometric modeling.

[0003] In existing hybrid representation indoor mapping schemes, a scene representation in a unified coordinate system is typically established first, then fixed-scale sampling is performed along the rays near the estimated surface, and a geometric coupling constraint of uniform strength is applied between the explicit representation and the implicit surface to obtain a renderable and queryable 3D map. Although this type of scheme combines explicit rendering efficiency and implicit geometric continuity, it still has at least the following problems in indoor scenes: (1) The observation uncertainty in indoor scenes is highly heterogeneous in space. Existing solutions often use a uniform sampling perturbation intensity, which cannot dynamically adjust the sampling range according to local differences such as weak texture planes, occluded edges, mirror or glass reflection areas and small furniture structures. This results in insufficient exploration of high uncertainty areas, while high certainty areas have invalid sampling overhead.

[0004] (2) Existing schemes usually apply a globally fixed explicit-implicit geometric coupling strength to all sampling points. In low confidence regions, incorrect signed distance field estimates, noisy observations or insufficient sampling will be continuously amplified by the unified strong constraint, which can easily cause the Gaussian elements to be pulled to the wrong position, thereby causing local structural collapse, surface tearing or training oscillation.

[0005] (3) Existing schemes separate the sampling strategy from the loss weight design and have not yet formed a closed-loop optimization mechanism that drives the sampling distribution adjustment and constraint strength adjustment simultaneously with a unified local confidence signal. This makes it difficult for the system to automatically switch between "strengthening exploration and moderately reducing penalties in high uncertainty areas" and "concentrating sampling in high certainty areas and maintaining strong constraints".

[0006] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0007] The main objective of this invention is to provide an uncertainty-driven robust mapping method, system, terminal, and computer-readable storage medium for indoor scenes. This invention aims to address the problem that existing hybrid representation methods for indoor 3D mapping lack adaptive sampling and constraint strength adjustment mechanisms guided by local confidence, resulting in insufficient structural recovery in high-uncertainty regions and continuous amplification of errors.

[0008] To achieve the above objectives, this invention provides an uncertainty-driven robust mapping method for indoor scenes, which includes the following steps: Input image frames, pose information, and depth or point cloud observations. Perform multimodal time alignment on all input data and unify different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information. Output pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system. Explicit surface parameters are maintained in space, and a continuous distance field network is established to process pixel observations, spatial points, ray origins, ray directions, pose relationships and validity information under a unified coordinate system. The output is the explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query position, gradient and local geometric reliability scale term under the current iteration. Query the NSDF parameters at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, query the NSDF parameters at each sampling point location to obtain the sampling point-level local geometric reliability scale term. Input the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. Calculate the adaptive sampling standard deviation of each ray based on the ray-level local geometric reliability scale term. Increase the sampling perturbation range of the surface neighborhood for high uncertainty regions and decrease the perturbation range for high certainty regions. Form a stable and controllable heteroscedastic sampling distribution within the clipping boundary and output the set of sampling points generated around the current estimated surface position. The input sample point is the SDF value, the original geometric constraint weights, and the sample point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, the sample point-level local geometric reliability scale term is uniquely mapped to a confidence adjustment factor. The original geometric constraint weights are rewritten as reweighted geometric constraint weights, and the reweighted geometric coupling loss term is output. Input image reconstruction error, geometric error, geometric coupling loss term and sampling point set, perform explicit forward rendering and implicit distance field query respectively in the same training iteration, and update explicit and implicit parameters after comprehensive loss calculation; Output the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map generated based on the explicit Gaussian parameters and NSDF parameters for rendering from a new perspective and the continuous geometric results for geometric querying, surface extraction and structural analysis.

[0009] Optionally, the aforementioned uncertainty-driven robust mapping method for indoor scenes, wherein performing multimodal time alignment on all input data and unifying different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information specifically includes: Match image frames, pose information, and depth or point cloud observations by timestamp; When the timestamps of the various sensors are not completely consistent, the pose information is interpolated. The sensor coordinate system is transformed to the camera coordinate system using sensor extrinsic parameters, and the camera pose is transformed to the world coordinate system. Rays are generated based on camera intrinsic parameters, and consistency checks are performed on depth backprojection points.

[0010] Optionally, in the aforementioned uncertainty-driven robust mapping method for indoor scenes, the explicit surface parameters are used to handle differentiable rasterization and appearance or depth rendering, and the continuous distance field network is used to handle continuous distance and gradient queries.

[0011] Optionally, in the aforementioned uncertainty-driven robust mapping method for indoor scenes, the SDF value is used to constrain the Gaussian surface to approach the zero level set, and the local geometric reliability scale term is used to provide a basis for the adaptive sampling standard deviation and confidence adjustment factor.

[0012] Optionally, the uncertainty-driven robust mapping method for indoor scenes, wherein the input includes the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term; the adaptive sampling standard deviation of each ray is calculated based on the ray-level local geometric reliability scale term; the sampling perturbation range of the surface neighborhood is increased for high uncertainty regions, and the perturbation range is decreased for high certainty regions; a stable and controllable heteroscedastic sampling distribution is formed within the clipping boundary; and the output is a set of sampling points generated around the current estimated surface position, specifically including: Under a fixed sampling strategy, sampling points are generated around the current surface position along the ray direction: ; in, The first random perturbation applied along the ray depth direction under a fixed sampling strategy is used to generate neighborhood sampling points near the current surface estimate. The standard deviation of the system's fundamental perturbation is used to describe the default sampling scale under a fixed sampling strategy; The location of the perturbed ray parameters under a fixed sampling strategy is used to determine the depth of the sampling point in the ray direction; The current estimated surface location depth of the ray under a fixed sampling strategy; These are three-dimensional sampling points obtained by mapping the positions of ray parameters under a fixed sampling strategy. Ray origin under a fixed sampling strategy; The ray direction vector under a fixed sampling strategy; Based on ray-level local geometric reliability metric r Construct the adaptive sampling standard deviation for each ray : ; For adaptive sampling standard deviation Perform top and bottom bounds clipping: ; Here, clip(·) is a clipping operation used to clip... Constraints on [ , Within this range, ensure that the sampling range is stable and controllable; This is the lower bound of the adaptive sampling standard deviation; This is the upper bound of the adaptive sampling standard deviation; Under heteroscedastic sampling conditions, surface neighborhood sampling points are generated around the currently estimated surface position for each ray: ; in, For the adaptive sampling strategy, the first Ray 1 The second random perturbation applied to each sampling point along the ray depth direction; For the first adaptive sampling strategy Ray 1 The location of the disturbed ray parameter corresponding to each sampling point; For the first The current estimated surface location depth of the ray is used as the center for adaptive sampling; For the first adaptive sampling strategy Ray 1 Each sampling point corresponds to a three-dimensional sampling point; For the first The starting point of each ray is determined by the current camera center or projection center; For the first The ray direction vector of a ray; The number of surface neighborhood sampling points generated for each ray.

[0013] Optionally, the uncertainty-driven robust mapping method for indoor scenes, wherein the SDF value at the input sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term are uniquely mapped to a confidence adjustment factor through a monotonically increasing confidence mapping function with saturation characteristics, and the original geometric constraint weights are rewritten as reweighted geometric constraint weights, outputting the reweighted geometric coupling loss term, specifically including: Local geometric reliability scale term at the sampling point level m Mapped to confidence adjustment factor: ; in, For the first The confidence adjustment factor corresponding to each sampling point is used to adjust the strength of the geometric constraint; τ is the scale parameter for the transition from the low confidence region to the high confidence region, used to control the rate of weight change. This is the lower limit of the confidence adjustment factor; The original geometric constraint weights are reweighted using a confidence adjustment factor: ; in, These are the reweighted geometric constraint weights after local confidence modulation; For the first The original geometric constraint weights of each sampling point; Introduce reweighted weights into the explicit-implicit geometric coupling loss: ; in, The weighted explicit-implicit geometric coupling loss; For the first The signed distance field values ​​are taken at each sampling point.

[0014] Optionally, in the aforementioned uncertainty-driven robust mapping method for indoor scenes, the reweighted geometric constraint weights are used to maintain flexible constraints in low-confidence regions and strong constraints in high-confidence regions.

[0015] Furthermore, to achieve the above objectives, the present invention also provides an uncertainty-driven robust mapping system for indoor scenes, wherein the uncertainty-driven robust mapping system for indoor scenes includes: The multimodal input and coordinate unification module is used to input image frames, pose information, and depth or point cloud observations. It performs multimodal time alignment on all input data and unifies different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information. The output is pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system. The dual-representation mapping module is used to maintain explicit surface parameters in space and establish a continuous distance field network. It processes pixel observations, spatial points, ray origins, ray directions, pose relationships and validity information under a unified coordinate system, and outputs explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query location, gradient and local geometric reliability scale term under the current iteration. The local confidence assessment module is used to query the NSDF parameters at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, it queries the NSDF parameters at each sampling point location to obtain the sampling point-level local geometric reliability scale term. The adaptive sampling module is used to take into account the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. Based on the ray-level local geometric reliability scale term, it calculates the adaptive sampling standard deviation of each ray, increases the sampling perturbation range of the surface neighborhood for high uncertainty regions, decreases the perturbation range for high certainty regions, and forms a stable and controllable heteroscedastic sampling distribution within the clipping boundary. It outputs a set of sampling points generated around the current estimated surface position. The loss reweighting module is used to input the SDF value at the sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, it uniquely maps the sampling point-level local geometric reliability scale term to a confidence adjustment factor, rewrites the original geometric constraint weights into reweighted geometric constraint weights, and outputs the reweighted geometric coupling loss term. The joint optimization module is used to take the input image reconstruction error, geometric error, geometric coupling loss term and sampling point set as input. In the same training iteration, it performs explicit forward rendering and implicit distance field query respectively, and updates the explicit and implicit parameters after comprehensively calculating the loss. The map output module is used to output the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map generated based on the explicit Gaussian parameters and NSDF parameters for rendering from a new perspective and the continuous geometric results for geometric query, surface extraction and structural analysis.

[0016] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an uncertainty-driven robust mapping program for indoor scenes stored in the memory and executable on the processor, wherein when the uncertainty-driven robust mapping program for indoor scenes is executed by the processor, it implements the steps of the uncertainty-driven robust mapping method for indoor scenes as described above.

[0017] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an uncertainty-driven robust mapping program for indoor scenes, and the uncertainty-driven robust mapping program for indoor scenes, when executed by a processor, implements the steps of the uncertainty-driven robust mapping method for indoor scenes as described above.

[0018] In this invention, input image frames, pose information, and depth or point cloud observations are processed. Multimodal temporal alignment is performed on all input data, and different modal observations are unified to the world coordinate system based on sensor extrinsic parameters and pose information. The output is pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system. Explicit surface parameters are maintained in space, and a continuous distance field network is established to process the pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system. The output is the explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query location, gradient, and local geometric reliability scale term under the current iteration. The NSDF parameters are queried at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, the NSDF parameters are queried at each sampling point location to obtain the sampling point-level local geometric reliability scale term. The current estimated surface location of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term are input. Adaptive sampling is calculated for each ray based on the ray-level local geometric reliability scale term. The standard deviation of the sample is used to increase the sampling perturbation range of the surface neighborhood in high uncertainty areas and decrease the perturbation range in high certainty areas, forming a stable and controllable heteroscedastic sampling distribution within the clipping boundary. The output is a set of sampling points generated around the current estimated surface position. The input is the SDF value at the sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, the sampling point-level local geometric reliability scale term is uniquely mapped to a confidence adjustment factor, and the original geometric constraint weights are rewritten as reweighted geometric constraint weights. The output is the reweighted geometric coupling loss term. The input is the image reconstruction error, geometric error, geometric coupling loss term, and the sampling point set. In the same training iteration, explicit forward rendering and implicit distance field query are performed respectively. After comprehensively calculating the loss, the explicit and implicit parameters are updated. The output is the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map for new perspective rendering and the continuous geometric results for geometric query, surface extraction, and structural analysis generated based on the explicit Gaussian parameters and the NSDF parameters. This invention can improve rendering fidelity, boundary structure stability, detail recovery capability, and training convergence robustness in indoor unseen viewpoints. Attached Figure Description

[0019] Figure 1 This is a flowchart of a preferred embodiment of the uncertainty-driven robust mapping method for indoor scenes of the present invention; Figure 2 This is a schematic diagram illustrating the entire mapping process in a preferred embodiment of the uncertainty-driven robust mapping method for indoor scenes of the present invention. Figure 3This is a flowchart illustrating the uncertainty-driven adaptive sampling module in a preferred embodiment of the uncertainty-driven robust mapping method for indoor scenes of the present invention. Figure 4 This is a flowchart illustrating the uncertainty-driven robust mapping method for indoor scenarios according to the present invention, specifically the loss reweighting module based on uncertainty. Figure 5 This is a diagram showing the relationship between software deployment and data flow in a preferred embodiment of the uncertainty-driven robust mapping method for indoor scenarios of the present invention. Figure 6 This is a flowchart illustrating the entire execution steps of the uncertainty-driven robust mapping method for indoor scenes according to a preferred embodiment of the present invention. Figure 7 This is a structural diagram of a preferred embodiment of the uncertainty-driven robust mapping system for indoor scenes of the present invention. Figure 8 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] This invention is applicable to high-fidelity 3D mapping tasks in complex indoor scenes, particularly suitable for rendering large areas of weakly textured planes, partially occluded edges, specular or glass reflection areas, densely distributed furniture areas, and unseen viewpoint rendering scenes where the training and testing perspectives deviate significantly. The system input includes image sequences, corresponding pose information, and depth observations, point cloud observations, or their equivalent geometric anchoring information used to establish geometric priors. The system processing includes multimodal alignment, unified coordinate modeling, hybrid scene representation construction, local geometric reliability scale acquisition and transfer, adaptive surface sampling, loss reweighting, and joint optimization. The system output is a scene map that can be used for rendering from new perspectives, as well as continuous geometric results that can be used for geometric queries, surface extraction, and structural analysis.

[0022] The uncertainty-driven robust mapping method for indoor scenes described in the preferred embodiment of the present invention, such as... Figure 1 , Figure 2 and Figure 5 As shown, the uncertainty-driven robust mapping method for indoor scenes includes the following steps: Step S10: Input image frames, pose information, and depth or point cloud observations. Perform multimodal time alignment on all input data, and unify the different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information. Output pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system.

[0023] Specifically, step S10 is achieved through... Figure 7 This is achieved through a multimodal input and coordinate unification module, which addresses the problem that image, depth observation or point cloud observation, and pose information belong to different coordinate systems and cannot be directly modeled together.

[0024] The input consists of image frames, pose information, and depth or point cloud observations (depth observations are typically pixel-based depth maps, with each pixel corresponding to a distance value; point cloud observations are sets of 3D points, each point directly providing spatial coordinates, usually not corresponding one-to-one with image pixels. Both can serve as geometric anchoring information, but differ in data structure, projection / backprojection, and validity checking methods); the processing logic first performs multimodal temporal alignment (e.g., using image timestamps as a reference, matching other modalities within an allowed time window, and interpolating pose when necessary, to temporally align image frames, poses, and depth maps or point clouds at the same observation time, forming a set of aligned observation tuples that can be jointly modeled), and then unifies the different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information; the output consists of pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system.

[0025] The specific processing procedure is as follows: In one feasible implementation, image frames, pose information, and depth or point cloud observations can be matched according to timestamps; when the timestamps of each sensor are not completely consistent, the pose information is interpolated; then, the sensor coordinate system is transformed to the camera coordinate system using sensor extrinsic parameters, and the camera pose is transformed to the world coordinate system; finally, rays are generated according to camera intrinsic parameters, and consistency checks are performed on the depth back-projection points. The direct technical effect of this multimodal input and coordinate unification module is that it provides a consistent spatial reference for subsequent explicit rendering branches, implicit distance field branches, and local confidence assessments.

[0026] Step S20: Maintain explicit surface parameters in space, and establish a continuous distance field network. Process the pixel observations, spatial points, ray origins, ray directions, pose relationships and validity information under a unified coordinate system, and output the explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query position, gradient and local geometric reliability scale term under the current iteration.

[0027] Specifically, step S20 is achieved through... Figure 7This is achieved through a dual-representation mapping module, which uses a joint modeling approach of explicit Gaussian surface representation and implicit signed distance field representation to express indoor scenes. Explicit branches are used to efficiently output rendering results such as color, depth, and transparency, while implicit branches are used to query the signed distance from any 3D point to the surface, gradient, and scale terms representing local geometric reliability.

[0028] The input consists of pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in a unified coordinate system. The processing logic maintains updatable explicit surface parameters in space while simultaneously establishing a continuous distance field network. The output includes the explicit Gaussian parameters and NSDF parameters (Neural Signed Distance Field) for the current iteration, as well as the SDF value (Signed Distance Field), gradient, and local geometric reliability scale term obtained at the query location. The explicit parameters are responsible for differentiable rasterization and appearance or depth rendering, while the NSDF network handles continuous distance and gradient queries. These two parameters are linked at the same spatial sampling point through SDF geometric coupling loss: the SDF value constrains the Gaussian surface to approach the zero-level set, and the scale term output by the NSDF provides the basis for adaptive sampling standard deviation and confidence adjustment factors.

[0029] The specific processing steps are as follows: Gaussian primitives and NSDF networks are initialized or constrained using depth or point cloud geometry anchoring; the explicit Gaussian branch outputs color, depth, and transparency, while the NSDF branch queries the SDF value, gradient, and local reliability scale for each sampling point; finally, explicit-implicit geometric coupling is established using the SDF value at the sampling point. The direct technical advantage of this dual-representation mapping module is that it simultaneously possesses efficient rendering capabilities, continuous geometric representation capabilities, and the ability to estimate local uncertainties.

[0030] Step S30: Query the NSDF parameters at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, query the NSDF parameters at each sampling point location to obtain the sampling point-level local geometric reliability scale term.

[0031] Specifically, step S30 is achieved through... Figure 7 This is achieved through a local confidence assessment module, which is used to quantify the differences in geometric reliability between different rays or different local regions.

[0032] The inputs are the explicit Gaussian parameters, NSDF parameters, SDF value obtained at the query location, gradient, and local geometric reliability scale term for the current iteration. The processing logic is to read and pass the local geometric reliability scale term in stages within one iteration: first, read the ray-level local geometric reliability scale term at the current estimated surface location. r The data is then passed to the adaptive sampling module; after the adaptive sampling module generates sampling points, the sampling point-level scale term is read at each sampling point location. m is then passed to the loss reweighting module; this module does not calculate the confidence adjustment factor for the sampling points. The output is a ray-level local geometric reliability metric. r and sampling point level local geometric reliability metric m .

[0033] Specific processing steps: First stage: Query the NSDF at the current estimated surface location to obtain... r And pass it to the adaptive sampling module; in the second stage, the adaptive sampling module generates a set of sampling points { After that, NSDF is queried at each sampling point location to obtain m This information is then passed to the loss reweighting module. The direct technical effect of this local confidence assessment module is that it provides the system with a reliable basis for dynamically allocating computing resources and geometric penalty strength.

[0034] Step S40: Input the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. Calculate the adaptive sampling standard deviation of each ray based on the ray-level local geometric reliability scale term. Increase the sampling perturbation range of the surface neighborhood for high uncertainty regions and decrease the perturbation range for high certainty regions. Form a stable and controllable heteroscedastic sampling distribution within the clipping boundary and output the set of sampling points generated around the current estimated surface position.

[0035] Specifically, step S40 is achieved through... Figure 7 This is achieved through an adaptive sampling module (i.e., an uncertainty-driven adaptive sampling module), which is used to address the problem that a fixed sampling scale cannot adapt to the local geometric uncertainty differences in indoor scenes.

[0036] like Figure 3 As shown, the inputs are the current estimated surface location of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. r The processing logic is based on the ray-level local geometric reliability scale term. r The adaptive sampling standard deviation for each ray is calculated. For regions of high uncertainty, the sampling perturbation range in the surface neighborhood is increased; for regions of high certainty, the perturbation range is decreased, resulting in a stable and controllable heteroscedasticity sampling distribution within the clipping boundary. This adjustment employs a continuous relationship. Smaller sizes correspond to lower geometric reliability and larger sampling range. A larger sampling size corresponds to higher geometric reliability and a narrower sampling range, eliminating the need for hard thresholds to delineate high / low uncertainty regions. The output is a set of sampling points generated around the currently estimated surface location. The direct technical effect of this adaptive sampling module is that it allocates more of the limited sampling budget to truly discriminative high uncertainty regions, while reducing invalid sampling in stable regions.

[0037] Under a fixed sampling strategy, sampling points are generated around the current surface position along the ray direction: ; in, The first random perturbation applied along the ray depth direction under a fixed sampling strategy is used to generate neighborhood sampling points near the current surface estimate. The standard deviation of the system's fundamental perturbation is used to describe the default sampling scale under a fixed sampling strategy; The location of the perturbed ray parameters under a fixed sampling strategy is used to determine the depth of the sampling point in the ray direction; The current estimated surface location depth of the ray under a fixed sampling strategy; These are three-dimensional sampling points obtained by mapping the positions of ray parameters under a fixed sampling strategy. Ray origin under a fixed sampling strategy; This refers to the ray direction vector under a fixed sampling strategy.

[0038] Based on ray-level local geometric reliability metric r Construct the adaptive sampling standard deviation for each ray : ; For adaptive sampling standard deviation Perform top and bottom bounds clipping: ; Here, clip(·) is a clipping operation used to clip... Constraints on [ , Within this range, ensure that the sampling range is stable and controllable; This is the lower bound of the adaptive sampling standard deviation; This is the upper bound of the adaptive sampling standard deviation.

[0039] Under heteroscedastic sampling conditions, surface neighborhood sampling points are generated around the currently estimated surface position for each ray: ; in, For the adaptive sampling strategy, the first Ray 1 The second random perturbation applied to each sampling point along the ray depth direction; For the first adaptive sampling strategy Ray 1 The location of the disturbed ray parameter corresponding to each sampling point; For the first The current estimated surface location depth of the ray is used as the center for adaptive sampling; For the first adaptive sampling strategy Ray 1 Each sampling point corresponds to a three-dimensional sampling point; For the first The starting point of each ray is determined by the current camera center or projection center; For the first The ray direction vector of a ray; The number of surface neighborhood sampling points generated for each ray.

[0040] Step S50: Input the SDF value at the sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. Through a confidence mapping function that is monotonically increasing and has saturation characteristics, the sampling point-level local geometric reliability scale term is uniquely mapped to the confidence adjustment factor. The original geometric constraint weights are rewritten as reweighted geometric constraint weights, and the reweighted geometric coupling loss term is output.

[0041] Specifically, step S50 is achieved through... Figure 7 It is implemented using a loss reweighting module (i.e., a loss reweighting module based on uncertainty), which is used to address the problem that uniform strong geometric constraints in low-confidence regions can easily amplify erroneous gradients.

[0042] The inputs are the SDF values ​​at the sampling points, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. m The processing logic involves using a monotonically increasing confidence mapping function with saturation characteristics only in the loss reweighting module to apply the scaling term at the sampling points. m The unique mapping is the confidence adjustment factor. The original geometric constraint weights are rewritten as reweighted geometric constraint weights. This maintains flexible constraints in low-confidence regions and strong constraints in high-confidence regions. The output is the reweighted geometric coupling loss term, such as... Figure 4 As shown. The direct technical effect of this loss reweighting module is to reduce the misleading effect of low-confidence supervision on the optimization direction and suppress local structural collapse and training oscillations.

[0043] Local geometric reliability scale term at the sampling point level m Mapped to confidence adjustment factor: ; in, For the first The confidence adjustment factor corresponding to each sampling point is used to adjust the strength of the geometric constraint; τ is the scale parameter (confidence transition scale) for the transition from the low confidence region to the high confidence region, which is used to control the rate of weight change; This serves as the lower limit of the confidence adjustment factor, ensuring that the low-confidence region retains the minimum geometric constraint weight and avoids the complete loss of geometric traction.

[0044] The original geometric constraint weights are reweighted using a confidence adjustment factor: ; in, These are the reweighted geometric constraint weights after local confidence modulation; For the first The original geometric constraint weights of each sampling point.

[0045] Introduce reweighted weights into the explicit-implicit geometric coupling loss: ; in, The reweighted explicit-implicit geometric coupling loss (i.e., the GS-SDF geometric coupling term, where GS stands for Gaussian Splatting). For the first The signed distance field (SDF value) is taken at each sampling point.

[0046] Step S60: Input image reconstruction error, geometric error, geometric coupling loss term and sampling point set, perform explicit forward rendering and implicit distance field query respectively in the same training iteration, and update the explicit and implicit parameters after comprehensively calculating the loss.

[0047] Specifically, step S60 is achieved through... Figure 7 This is achieved through a joint optimization module that couples explicit rendering, implicit distance fields, adaptive sampling, and reweighted geometric constraints into a unified optimization process.

[0048] The inputs are image reconstruction error (derived from the comparison between the GS rendering result and the real image), geometric error (derived from the geometric supervision / regularization of NSDF itself (SDF querying of spatial points or sampling points combined with depth / point cloud geometric anchoring information)), the reweighted explicit-implicit geometric coupling loss (GS-SDF geometric coupling term) output by the loss reweighting module, and the set of sampling points. The processing logic involves performing explicit forward rendering and implicit distance field querying separately within the same training iteration, and updating the explicit and implicit parameters after comprehensively calculating the above three types of losses. The direct technical effect of this joint optimization module is to form a dynamic closed loop that enhances exploration in high-uncertainty regions, moderately reduces penalties, and concentrates sampling in high-certainty regions to maintain strong constraints.

[0049] Step S70: Output the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map for new perspective rendering and the continuous geometric results for geometric query, surface extraction and structural analysis generated based on the explicit Gaussian parameters and the NSDF parameters.

[0050] Specifically, step S70 is achieved through... Figure 7 This is achieved through the map output module in [the system / platform].

[0051] The software operation data flow of this invention is as follows: sensor data (multimodal) input → spatiotemporal alignment and coordinate unification → hybrid scene representation construction, forward rendering and distance field query → local geometric reliability scale acquisition and transmission → adaptive surface sampling → confidence mapping and loss reweighting → joint optimization → map output, such as Figure 2 As shown, this data stream explicitly defines the calling relationships between input, processing, and output, and forms a robust closed loop in which scale terms obtained from the same local geometric reliability scale field at different query locations drive sampling and loss reweighting respectively.

[0052] like Figure 6 As shown, the specific implementation process of the present invention includes the following steps.

[0053] Step S1: Acquire the image sequence of the indoor scene, the corresponding pose information, and the optional depth or point cloud geometric observations, and unify the different modal observations into the world coordinate system.

[0054] Step S2: Establish a dual-representation scene model consisting of an explicit Gaussian surface representation and an implicit signed distance field representation, and initialize the explicit surface parameters, distance field network parameters, and related caches.

[0055] Step S3: Perform forward rendering and distance field query based on the current dual-representation scene model to obtain scale terms for color, depth, transparency, surface position, and local geometric reliability.

[0056] Step S4: Calculate the adaptive sampling standard deviation for each ray using the ray-level local geometric reliability scale term, and perform upper and lower bound clipping on the standard deviation.

[0057] Step S5: Generate surface neighborhood sampling points along the local depth direction around the currently estimated surface position of each ray, so that the high uncertainty region can obtain a wider sampling coverage and the high certainty region can obtain a more concentrated local sampling.

[0058] Step S6: Uniquely map the sampling point-level local geometric reliability scale term to the confidence adjustment factor, and rewrite the original explicit-implicit geometric constraint weights into reweighted geometric constraint weights.

[0059] Step S7: Input the image reconstruction loss, geometric error (NSDF's own geometric supervision / regularization), and the reweighted explicit-implicit coupling loss into the joint optimization module to update the explicit Gaussian parameters and implicit distance field parameters.

[0060] Step S8: Repeat steps S3 to S7 to allow the high uncertainty region to gradually obtain stable geometric support under more thorough exploration and flexible constraints, and to achieve faster convergence in the high certainty region.

[0061] Step S9: When the rendering quality, local structural stability, and overall geometric consistency meet the set conditions, output a map result that can be used for new perspective rendering and geometric reconstruction.

[0062] This invention takes image sequences, pose information, and depth or point cloud observations used to establish geometric priors as inputs to construct a dual-representation joint mapping architecture consisting of explicit Gaussian surface representation and implicit signed distance field representation. The explicit branch is used to undertake efficient rendering tasks, while the implicit branch is used to provide continuous geometric manifold, surface distance, and local confidence representation, thereby balancing appearance expressiveness and geometric usability within the same system.

[0063] This invention employs two core mechanisms in the joint optimization process: First, it constructs an uncertainty-driven adaptive sampling module that dynamically adjusts the sampling perturbation range of the surface neighborhood based on the ray-level local geometric reliability scale term, enabling more thorough geometric exploration in high-uncertainty regions while maintaining more concentrated and efficient local sampling in high-certainty regions. Second, it constructs an uncertainty-based loss reweighting module that maps the sampling point-level local geometric reliability scale term to a confidence adjustment factor, thereby forming dynamic weights for explicit-implicit geometric coupling constraints, ensuring flexible constraints in low-confidence regions and strong constraints in high-confidence regions. Through these technical solutions, the rendering fidelity, boundary structure stability, detail recovery capability, and training convergence robustness under unseen indoor viewpoints can be improved, enhancing structural recovery capability, training convergence stability, and overall mapping quality under unseen viewpoints without altering high-fidelity rendering capabilities.

[0064] This invention can be deployed in UAV onboard computing units, ground workstations, edge servers, or a combined collaborative deployment environment. In multimodal acquisition scenarios, the onboard unit is mainly responsible for the acquisition and synchronization of image, pose, and depth / point cloud observations, while the ground workstation or server is mainly responsible for local geometric reliability scale acquisition and transmission, adaptive sampling, loss reweighting, joint optimization, and map result output. When computing resources meet the requirements, the entire process can also be deployed and executed on a single workstation, such as... Figure 5 As shown.

[0065] Table 1 lists the key parameter configurations of the present invention in a preferred embodiment. Each parameter can be adjusted according to the complexity of the scene, the intensity of local occlusion, the level of geometric noise, and computing resources, but does not change the core technical idea of ​​the present invention to drive adaptive sampling and loss reweighting simultaneously with unified local confidence.

[0066] Table 1: Optimal Implementation Parameters

[0067] By combining and comparing key modules, it can be verified that adaptive sampling and loss reweighting have a synergistic effect on improving the rendering quality and optimizing stability of indoor unseen viewpoints. Table 2 shows the combined comparison results represented by Peak Signal-to-Noise Ratio (PSNR), where a higher PSNR value indicates a lower reconstruction error under extrapolated viewpoints and a rendering result that is closer to the actual observation.

[0068] Table 2: Summary of Key Module Combination Effects

[0069] This invention, besides being applicable to high-fidelity mapping and rendering of unseen perspectives in indoor scenes, can also be extended to applications such as robot indoor navigation, warehouse digital twins, indoor inspection, embodied intelligent environment modeling, augmented reality scene reconstruction, and high-fidelity indoor simulation. For tasks requiring simultaneous consideration of appearance realism, local structural stability, and robust training convergence, this invention exhibits high versatility. The rendered maps and continuous geometric results output by this invention can further serve downstream tasks such as collision detection, path planning, line-of-sight analysis, surface extraction, scene editing, asset reconstruction, and task planning, possessing clear software productization and engineering transformation value.

[0070] Furthermore, such as Figure 7 As shown, based on the above-described uncertainty-driven robust mapping method for indoor scenes, this invention also provides a corresponding uncertainty-driven robust mapping system for indoor scenes, wherein the uncertainty-driven robust mapping system for indoor scenes includes: The multimodal input and coordinate unification module 51 is used to input image frames, pose information and depth or point cloud observations, perform multimodal time alignment on all input data, and unify different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information. The output is pixel observations, spatial points, ray origin, ray direction, pose relationship and validity information in the unified coordinate system. The dual representation mapping module 52 is used to maintain explicit surface parameters in space and establish a continuous distance field network. It processes pixel observations, spatial points, ray origins, ray directions, pose relationships and validity information under a unified coordinate system, and outputs explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query position, gradient and local geometric reliability scale terms under the current iteration. The local confidence assessment module 53 is used to query the NSDF parameters at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, it queries the NSDF parameters at each sampling point location to obtain the sampling point-level local geometric reliability scale term. The adaptive sampling module 54 is used to input the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. Based on the ray-level local geometric reliability scale term, it calculates the adaptive sampling standard deviation of each ray, increases the sampling perturbation range of the surface neighborhood for high uncertainty regions, decreases the perturbation range for high certainty regions, and forms a stable and controllable heteroscedastic sampling distribution within the clipping boundary. It outputs a set of sampling points generated around the current estimated surface position. The loss reweighting module 55 is used to input the SDF value at the sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, it uniquely maps the sampling point-level local geometric reliability scale term to a confidence adjustment factor, rewrites the original geometric constraint weights into reweighted geometric constraint weights, and outputs the reweighted geometric coupling loss term. The joint optimization module 56 is used to take the image reconstruction error, geometric error, geometric coupling loss term and sampling point set as input. In the same training iteration, it performs explicit forward rendering and implicit distance field query respectively, and updates the explicit and implicit parameters after comprehensively calculating the loss. The map output module 57 is used to output the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map for new perspective rendering and the continuous geometric results for geometric query, surface extraction and structural analysis generated based on the explicit Gaussian parameters and the NSDF parameters.

[0071] Among them, the multimodal input and coordinate unification module is responsible for forming the data foundation under the unified world coordinate system; the dual representation mapping module is responsible for simultaneously maintaining the explicit Gaussian surface representation and the implicit distance field representation; the local confidence evaluation module is responsible for obtaining and transmitting the local geometric reliability scale term output by NSDF; the adaptive sampling module is responsible for dynamically adjusting the sampling range of the surface neighborhood based on the ray-level local geometric reliability scale term; the loss reweighting module is responsible for mapping the sampling point-level local geometric reliability scale term to the confidence adjustment factor and dynamically adjusting the geometric coupling constraint strength; the joint optimization module is responsible for unifying the image reconstruction target, geometric target and reweighted coupling target to perform parameter updates; and the map output module is responsible for outputting the rendering results and geometric results.

[0072] Furthermore, such as Figure 8 As shown, based on the above-mentioned uncertainty-driven robust mapping method and system for indoor scenes, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 8 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0073] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores an uncertainty-driven robust mapping program 40 for indoor scenes, which can be executed by the processor 10 to implement the uncertainty-driven robust mapping method for indoor scenes in this application.

[0074] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the uncertainty-driven robust mapping method for indoor scenes.

[0075] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminal's processor 10, memory 20, and display 30 communicate with each other via a system bus.

[0076] In one embodiment, when the processor 10 executes the uncertainty-driven robust mapping program 40 for indoor scenes stored in the memory 20, it implements the steps of the uncertainty-driven robust mapping method for indoor scenes as described above.

[0077] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an uncertainty-driven robust mapping program for indoor scenes, and the uncertainty-driven robust mapping program for indoor scenes, when executed by a processor, implements the steps of the uncertainty-driven robust mapping method for indoor scenes as described above.

[0078] In summary, this invention provides an uncertainty-driven robust mapping method, system, terminal, and computer-readable storage medium for indoor scenes. The method includes: inputting image frames, pose information, and depth or point cloud observations; performing multimodal time alignment on all input data; unifying different modal observations to a world coordinate system based on sensor extrinsic parameters and pose information; outputting pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in a unified coordinate system; maintaining explicit surface parameters in space and establishing a continuous distance field network; processing pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system; outputting explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query location, gradient, and local geometric reliability scale term under the current iteration; querying NSDF parameters at the current estimated surface location to obtain ray-level local geometric reliability scale term; after generating a set of sampling points, querying NSDF parameters at each sampling point location to obtain sampling point-level local geometric reliability scale term; inputting the current estimated surface location of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term; and then... The adaptive sampling standard deviation of each ray is calculated based on the ray-level local geometric reliability scale term. For high-uncertainty regions, the sampling perturbation range of the surface neighborhood is increased; for high-certainty regions, the perturbation range is decreased. A stable and controllable heteroscedastic sampling distribution is formed within the clipping boundary, and the set of sampling points generated around the currently estimated surface position is output. The input consists of the SDF value at the sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, the sampling point-level local geometric reliability scale term is uniquely mapped to a confidence adjustment factor, and the original... The geometric constraint weights are rewritten as reweighted geometric constraint weights, and the reweighted geometric coupling loss term is output. The input includes image reconstruction error, geometric error, geometric coupling loss term, and a set of sampling points. Explicit forward rendering and implicit distance field query are performed separately in the same training iteration. After comprehensively calculating the loss, the explicit and implicit parameters are updated. The output includes the updated explicit Gaussian parameters and NSDF parameters, as well as a scene map for rendering from a new perspective and continuous geometric results for geometric query, surface extraction, and structural analysis generated based on the explicit Gaussian parameters and NSDF parameters. This invention can improve rendering fidelity, boundary structure stability, detail recovery capability, and training convergence robustness under unseen indoor perspectives.

[0079] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0080] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0081] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An uncertainty-driven robust mapping method for indoor scenes, characterized in that, The uncertainty-driven robust mapping method for indoor scenes includes: Input image frames, pose information, and depth or point cloud observations. Perform multimodal time alignment on all input data and unify different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information. Output pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system. Explicit surface parameters are maintained in space, and a continuous distance field network is established to process pixel observations, spatial points, ray origins, ray directions, pose relationships and validity information under a unified coordinate system. The output is the explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query position, gradient and local geometric reliability scale term under the current iteration. Query the NSDF parameters at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, query the NSDF parameters at each sampling point location to obtain the sampling point-level local geometric reliability scale term. Input the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. Calculate the adaptive sampling standard deviation of each ray based on the ray-level local geometric reliability scale term. Increase the sampling perturbation range of the surface neighborhood for high uncertainty regions and decrease the perturbation range for high certainty regions. Form a stable and controllable heteroscedastic sampling distribution within the clipping boundary and output the set of sampling points generated around the current estimated surface position. The input sample point is the SDF value, the original geometric constraint weights, and the sample point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, the sample point-level local geometric reliability scale term is uniquely mapped to a confidence adjustment factor. The original geometric constraint weights are rewritten as reweighted geometric constraint weights, and the reweighted geometric coupling loss term is output. Input image reconstruction error, geometric error, geometric coupling loss term and sampling point set, perform explicit forward rendering and implicit distance field query respectively in the same training iteration, and update explicit and implicit parameters after comprehensive loss calculation; Output the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map generated based on the explicit Gaussian parameters and NSDF parameters for rendering from a new perspective and the continuous geometric results for geometric querying, surface extraction and structural analysis.

2. The uncertainty-driven robust mapping method for indoor scenes according to claim 1, characterized in that, The process of performing multimodal time alignment on all input data and unifying different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information specifically includes: Match image frames, pose information, and depth or point cloud observations by timestamp; When the timestamps of the various sensors are not completely consistent, the pose information is interpolated. The sensor coordinate system is transformed to the camera coordinate system using sensor extrinsic parameters, and the camera pose is transformed to the world coordinate system. Rays are generated based on camera intrinsic parameters, and consistency checks are performed on depth backprojection points.

3. The uncertainty-driven robust mapping method for indoor scenes according to claim 1, characterized in that, The explicit surface parameters are used to handle differential rasterization and appearance or depth rendering, while the continuous distance field network is used to handle continuous distance and gradient queries.

4. The uncertainty-driven robust mapping method for indoor scenes according to claim 1, characterized in that, The SDF value is used to constrain the Gaussian surface to a near-zero level set, and the local geometric reliability scale term is used to provide a basis for the adaptive sampling standard deviation and confidence adjustment factor.

5. The uncertainty-driven robust mapping method for indoor scenes according to claim 1, characterized in that, The input includes the current estimated surface position of the ray, the ray origin and direction vector, and a ray-level local geometric reliability scale term. Based on this scale term, an adaptive sampling standard deviation for each ray is calculated. For regions with high uncertainty, the sampling perturbation range in the surface neighborhood is increased; for regions with high certainty, the perturbation range is decreased. A stable and controllable heteroscedastic sampling distribution is formed within the clipping boundary. The output is a set of sampling points generated around the currently estimated surface position, specifically including: Under a fixed sampling strategy, sampling points are generated around the current surface position along the ray direction: ； in, The first random perturbation applied along the ray depth direction under a fixed sampling strategy is used to generate neighborhood sampling points near the current surface estimate. The standard deviation of the system's fundamental perturbation is used to describe the default sampling scale under a fixed sampling strategy; The location of the perturbed ray parameters under a fixed sampling strategy is used to determine the depth of the sampling point in the ray direction; The current estimated surface location depth of the ray under a fixed sampling strategy; These are three-dimensional sampling points obtained by mapping the positions of ray parameters under a fixed sampling strategy; Ray origin under a fixed sampling strategy; The ray direction vector under a fixed sampling strategy; Based on ray-level local geometric reliability metric r Construct the adaptive sampling standard deviation for each ray : ； For adaptive sampling standard deviation Perform top and bottom bounds clipping: ； Here, clip(·) is a clipping operation used to clip... Constraints on [ , Within this range, ensure that the sampling range is stable and controllable; This is the lower bound of the adaptive sampling standard deviation; This is the upper bound of the adaptive sampling standard deviation; Under heteroscedastic sampling conditions, surface neighborhood sampling points are generated around the currently estimated surface position for each ray: ； in, For the adaptive sampling strategy, the first Ray 1 The second random perturbation applied to each sampling point along the ray depth direction; For the first adaptive sampling strategy Ray 1 The location of the disturbed ray parameters corresponding to each sampling point; For the first The current estimated surface location depth of the ray is used as the center for adaptive sampling; For the first adaptive sampling strategy Ray 1 Each sampling point corresponds to a three-dimensional sampling point; For the first The starting point of each ray is determined by the current camera center or projection center; For the first The ray direction vector of a ray; The number of surface neighborhood sampling points generated for each ray.

6. The uncertainty-driven robust mapping method for indoor scenes according to claim 5, characterized in that, The SDF value at the input sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term are uniquely mapped to a confidence adjustment factor through a monotonically increasing confidence mapping function with saturation characteristics. The original geometric constraint weights are then rewritten as reweighted geometric constraint weights, and the reweighted geometric coupling loss term is output, specifically including: The sampling point-level local geometric reliability scale term m Mapped to confidence adjustment factor: ； in, For the first The confidence adjustment factor corresponding to each sampling point is used to adjust the strength of the geometric constraint; τ is the scale parameter for the transition from the low confidence region to the high confidence region, used to control the rate of weight change. This is the lower limit of the confidence adjustment factor; The original geometric constraint weights are reweighted using a confidence adjustment factor: ； in, These are the reweighted geometric constraint weights after local confidence modulation; For the first The original geometric constraint weights of each sampling point; Introduce reweighted weights into the explicit-implicit geometric coupling loss: ； in, The weighted explicit-implicit geometric coupling loss; For the first The signed distance field values ​​are taken at each sampling point.

7. The uncertainty-driven robust mapping method for indoor scenes according to claim 6, characterized in that, The reweighted geometric constraint weights are used to maintain flexible constraints in low-confidence regions and strong constraints in high-confidence regions.

8. An uncertainty-driven robust mapping system for indoor scenes, characterized in that, The uncertainty-driven robust mapping system for indoor scenes is used to implement the uncertainty-driven robust mapping method for indoor scenes according to any one of claims 1-7, wherein the uncertainty-driven robust mapping system for indoor scenes comprises: The multimodal input and coordinate unification module is used to input image frames, pose information, and depth or point cloud observations. It performs multimodal time alignment on all input data and unifies different modal observations to the world coordinate system based on sensor extrinsic parameters and pose information. The output is pixel observations, spatial points, ray origins, ray directions, pose relationships, and validity information in the unified coordinate system. The dual-representation mapping module is used to maintain explicit surface parameters in space and establish a continuous distance field network. It processes pixel observations, spatial points, ray origins, ray directions, pose relationships and validity information under a unified coordinate system, and outputs explicit Gaussian parameters, NSDF parameters, SDF values ​​obtained at the query location, gradient and local geometric reliability scale term under the current iteration. The local confidence assessment module is used to query the NSDF parameters at the current estimated surface location to obtain the ray-level local geometric reliability scale term. After generating the sampling point set, it queries the NSDF parameters at each sampling point location to obtain the sampling point-level local geometric reliability scale term. The adaptive sampling module is used to take into account the current estimated surface position of the ray, the ray origin and direction vector, and the ray-level local geometric reliability scale term. Based on the ray-level local geometric reliability scale term, it calculates the adaptive sampling standard deviation of each ray, increases the sampling perturbation range of the surface neighborhood for high uncertainty regions, decreases the perturbation range for high certainty regions, and forms a stable and controllable heteroscedastic sampling distribution within the clipping boundary. It outputs a set of sampling points generated around the current estimated surface position. The loss reweighting module is used to input the SDF value at the sampling point, the original geometric constraint weights, and the sampling point-level local geometric reliability scale term. Through a monotonically increasing confidence mapping function with saturation characteristics, it uniquely maps the sampling point-level local geometric reliability scale term to a confidence adjustment factor, rewrites the original geometric constraint weights into reweighted geometric constraint weights, and outputs the reweighted geometric coupling loss term. The joint optimization module is used to take the input image reconstruction error, geometric error, geometric coupling loss term and sampling point set as input. In the same training iteration, it performs explicit forward rendering and implicit distance field query respectively, and updates the explicit and implicit parameters after comprehensively calculating the loss. The map output module is used to output the explicit Gaussian parameters and NSDF parameters updated in the current iteration, as well as the scene map generated based on the explicit Gaussian parameters and NSDF parameters for rendering from a new perspective and the continuous geometric results for geometric query, surface extraction and structural analysis.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and an uncertainty-driven robust mapping program for indoor scenes stored in the memory and executable on the processor. When the uncertainty-driven robust mapping program for indoor scenes is executed by the processor, it implements the steps of the uncertainty-driven robust mapping method for indoor scenes as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an uncertainty-driven robust mapping program for indoor scenes, which, when executed by a processor, implements the steps of the uncertainty-driven robust mapping method for indoor scenes as described in any one of claims 1-7.

Images