Indoor patrol path construction system based on robot vision recognition

Abstract

This invention discloses an indoor inspection path construction system based on robot vision recognition, belonging to the field of indoor inspection path technology. It obtains synchronously calibrated multimodal perception data; acquires a robot six-DOF pose sequence and sparse 3D point cloud; initializes a 3D Gaussian volume parameter set and constructs an initial 3D Gaussian sputtering model; obtains an optimized robot six-DOF pose sequence and an optimized 3D Gaussian volume parameter set; obtains a continuous collision probability field; obtains a collision-free optimized inspection path; and completes real-time autonomous inspection in an indoor environment. This invention avoids the accidental deletion of key obstacles during downsampling, improving the mapping integrity and pose estimation stability of slender obstacles in the indoor environment.

Description

Technical Field

[0001] This invention relates to the field of indoor patrol path technology, and in particular to an indoor patrol path construction system based on robot vision recognition. Background Technology

[0002] As the application of mobile robots in indoor environments continues to expand, autonomous inspection and path planning technologies based on vision and multimodal perception are gradually becoming core components of intelligent robot systems. In existing technologies, common methods typically rely on feature-matching-based simultaneous localization and mapping (SLAM) methods to construct environmental maps and perform path planning based on discrete grid maps or point cloud maps. However, in scenarios with slender obstacles, complex boundary structures, and dynamic occlusion, traditional methods exhibit significant shortcomings in mapping accuracy, collision modeling capabilities, and path planning safety.

[0003] Existing vision- or laser-based mapping methods typically use voxel filtering or uniform downsampling strategies to compress spatial data during point cloud processing. This fails to distinguish the importance of different spatial structures and easily misclassifies slender obstacles such as cables, poles, and thin table legs as noise and removes them, resulting in the loss of key obstacle information and affecting the safety of subsequent path planning.

[0004] Existing path planning methods typically model the environment based on discrete occupancy grids or cost maps, representing the environment as discrete passable or impassable areas. This approach lacks the ability to express the continuity of spatial occupancy. In such discrete representations, collision risks are often hard-segmented using threshold judgments, failing to reflect the continuous risk distribution characteristics of different areas in space. This makes it difficult to obtain smooth paths with safety margins during path planning. Furthermore, when dealing with slender obstacles or boundary areas, paths are prone to hugging edges or misjudging passable areas, thus increasing the collision risk in actual operation.

[0005] In existing technologies, mapping and path planning processes are typically independent and lack a unified constraint mechanism. The mapping phase fails to prioritize key structures for subsequent obstacle avoidance requirements, while the path planning phase struggles to fully utilize the structural information embedded in the mapping process, resulting in insufficient overall system synergy in complex environments. Furthermore, traditional path optimization methods often only consider geometric smoothness or length optimality, lacking the ability to model environmental risk information and dynamic constraints in a unified manner, making it difficult to effectively avoid high-risk areas while ensuring path smoothness. Summary of the Invention

[0006] One objective of this invention is to propose an indoor patrol path construction system based on robot vision recognition. This invention avoids the accidental deletion of key obstacles during downsampling, and improves the mapping integrity and pose estimation stability of slender obstacles in the indoor environment.

[0007] An indoor patrol path construction system based on robot vision recognition according to an embodiment of the present invention includes: The multimodal acquisition module collects multimodal sensing data of the indoor environment and performs time synchronization and calibration processing on the multimodal sensing data to obtain synchronized and calibrated multimodal sensing data. An improved RTAB module is used to process synchronously calibrated multimodal perception data in real time based on an improved RTAB-Map model, thereby obtaining the robot's six-degree-of-freedom pose sequence and sparse 3D point cloud. The Gaussian model initialization module initializes the three-dimensional Gaussian volume parameter set and constructs the initial three-dimensional Gaussian sputtering model using the robot's six-degree-of-freedom pose sequence and the sparse three-dimensional point cloud as priors. The graph-volume joint optimization module performs graph-volume joint incremental optimization on the initial 3D Gaussian sputtering model and the robot's six-DOF pose sequence to obtain the optimized robot's six-DOF pose sequence and the optimized 3D Gaussian volume parameter set. The hierarchical Gaussian management module performs hierarchical Gaussian body management based on the optimized three-dimensional Gaussian body parameter set, generates a hierarchical three-dimensional Gaussian body set, and constructs a continuous collision probability field in three-dimensional space based on the hierarchical three-dimensional Gaussian body set to obtain the continuous collision probability field. The path planning and optimization module performs connectivity analysis using the continuous collision probability field and performs sparse graph search to generate an initial inspection path based on the starting pose and target pose of the inspection task. It then performs gradient-driven trajectory optimization on the initial inspection path to obtain a collision-free optimized inspection path. The online replanning module integrates the collision-free optimized inspection path with robot dynamics constraints to generate a smooth inspection path and send it to the robot control system. During robot execution, the continuous collision probability field is updated, and the smooth inspection path is replanned online to complete real-time autonomous inspection in indoor environments.

[0008] Optionally, the multimodal acquisition module includes: Collect multimodal sensing data of the indoor environment and construct the original multimodal sensing dataset; Perform a unified time base transformation on the original timestamps corresponding to various sensors in the original multimodal sensing dataset to obtain standard timestamps; Using a reference time sequence under a unified time base as the alignment benchmark, cross-sensor time matching is performed on standard timestamps to determine the alignment sampling index of various sensors at each reference time. The observation data of various sensors obtained through time matching are calibrated, and the intrinsic parameter matrix and extrinsic parameter transformation matrix of each type of sensor are obtained respectively. Based on the aligned sampling index, intrinsic parameter matrix and extrinsic parameter transformation matrix, a unified spatiotemporal coordinate representation is performed on the multimodal observation content corresponding to each reference time, generating a synchronously calibrated multimodal sensing dataset.

[0009] Optionally, the improved RTAB module includes: The synchronously calibrated multimodal perception dataset is grouped according to the reference time sequence to construct a synchronous observation frame set. For each synchronous observation frame, slender obstacle edges, passage area boundaries, and stable structure regions are constructed in relation to the robot's indoor inspection path. Inspection sensitivity response values ​​are calculated to obtain an inspection sensitivity response map. Based on the inspection-sensitive response map, inspection-sensitive features are filtered for each synchronous observation frame to generate a subset of inspection-sensitive local feature descriptions, an inspection-sensitive appearance description vector, and an inspection-passage topology description vector. Weighted feature matching is performed on the sensitive local feature descriptor subsets of adjacent reference times, and the relative pose transformation matrix between adjacent reference times is obtained by combining the spatial observation content in the synchronously calibrated multimodal sensing dataset. Pose accumulation is performed based on the relative pose transformation matrix between adjacent reference times to obtain the initial six-degree-of-freedom robot pose sequence; Perform closed-loop detection based on the inspection-sensitive appearance description vector and the inspection-traffic topology description vector to determine the inspection constraint closed-loop set; Based on the initial robot six-DOF pose sequence, the relative pose transformation matrix between adjacent reference times, and the set of roving constraint closed loops, a roving constraint pose graph is constructed, and global optimization is performed on the roving constraint pose graph to obtain the optimized robot six-DOF pose sequence. Based on the optimized robot six-degree-of-freedom pose sequence, a unified coordinate transformation is performed on the spatial observation points corresponding to each reference time. The boundary support value, slender obstacle support value, and multi-frame visible stability value of each spatial observation point in the construction of the robot's indoor inspection path are calculated and weighted and fused into a structure retention contribution. Based on the structure retention contribution, sparse aggregation is performed on the spatial observation points after the unified coordinate transformation to obtain a sparse three-dimensional point cloud.

[0010] Optionally, the Gaussian model initialization module includes: Read the robot's six-degree-of-freedom pose sequence and sparse three-dimensional point cloud, establish the initial correspondence of the three-dimensional Gaussian body point by point according to the sparse three-dimensional points in the sparse three-dimensional point cloud, and generate a seed set of three-dimensional Gaussian bodies. Read the initial value of the center position of the seed of the 3D Gaussian body, and search for the local neighborhood sparse 3D point set in the sparse 3D point cloud with the initial value of the center position as the center. Construct a local structure matrix, and initialize the initial value of the principal axis direction parameter of the corresponding 3D Gaussian body based on the eigenvalue decomposition result of the local structure matrix. Read the support value of the slender obstacle and the contribution of structure preservation in the seed of the three-dimensional Gaussian body, and combine the eigenvalues ​​obtained by the seed of the three-dimensional Gaussian body in the process of local structure matrix eigenvalue decomposition to initialize the initial values ​​of the scale parameters of the corresponding three-dimensional Gaussian body along the three principal axes. Based on the initial values ​​of the scale parameters and principal axis parameters of the corresponding 3D Gaussian body seed along the three principal axes, the initial values ​​of the covariance matrix of the corresponding 3D Gaussian body are constructed. Based on the robot's six-degree-of-freedom pose sequence and a synchronously calibrated multimodal perception dataset, the visible observation set of the three-dimensional Gaussian body corresponding to the seed of the three-dimensional Gaussian body at each reference time is calculated, and the initial value of the color parameter of the three-dimensional Gaussian body corresponding to the seed of the three-dimensional Gaussian body is initialized according to the visible observation set. Based on the boundary support value and multi-frame visible stability value in the 3D Gaussian seed, initialize the initial value of the opacity parameter of the 3D Gaussian corresponding to the 3D Gaussian seed. The initial values ​​of the center position, covariance matrix, principal axis direction parameter, color parameter, and opacity parameter of each 3D Gaussian volume seed are collected to generate the initial value set of 3D Gaussian volume parameters. Based on the initial value set of 3D Gaussian volume parameters, the robot's six-degree-of-freedom pose sequence, and the sparse 3D point cloud, an initial 3D Gaussian sputtering model is constructed.

[0011] Optionally, the graph-volume joint optimization module includes: Based on the initial 3D Gaussian sputtering model, an incremental optimization window is constructed according to the current reference time, and the set of pose nodes and the set of visible 3D Gaussian volumes within the incremental optimization window are determined. Based on the boundary support value, slender obstacle support value, multi-frame visible stability value, and structure preservation contribution inherited by each 3D Gaussian volume in the visible 3D Gaussian volume set, the navigation enhancement weight of the 3D Gaussian volume is calculated, and the color observation residual and depth observation residual are constructed based on the standard observation content corresponding to each reference time within the incremental optimization window. Based on color observation residuals, depth observation residuals, patrol constraint odometer constraint edges, and patrol constraint closed-loop constraint edges, a graph-volume joint incremental optimization objective function is constructed. The graph-volume joint incremental optimization objective function is solved by alternating iterations, updating the pose node parameters and the 3D Gaussian volume parameters. If the change in the graph-volume joint incremental optimization objective function between two consecutive iterations is not greater than the preset convergence threshold, the iteration is stopped. After the incremental optimization window has converged, the updated pose node parameters are written back to the robot's six-DOF pose sequence, and the updated 3D Gaussian volume parameters are written back to the 3D Gaussian volume parameter set, thus obtaining the optimized robot six-DOF pose sequence and the optimized 3D Gaussian volume parameter set.

[0012] Optionally, the hierarchical Gaussian management module includes: Based on the optimized set of three-dimensional Gaussian volume parameters, the visibility frequency and visibility distribution of each three-dimensional Gaussian volume during the observation process at multiple reference times are statistically analyzed, and the information entropy value and visibility value of the corresponding three-dimensional Gaussian volume are calculated by combining the opacity parameter and covariance matrix of the three-dimensional Gaussian volume. Based on the information entropy and visibility values ​​of each 3D Gaussian body, the hierarchical score is calculated, and the 3D Gaussian bodies in the optimized 3D Gaussian body parameter set are hierarchically stored according to the hierarchical score to generate multi-level Gaussian body subsets. Based on the hierarchical score value, updated opacity parameter and structural retention contribution of the multi-level Gaussian body subset, the 3D Gaussian bodies in each level are clipped and filtered to retain the 3D Gaussian bodies corresponding to the navigation key boundaries and clip the low-contribution 3D Gaussian bodies to generate a hierarchical 3D Gaussian body set. A continuous collision probability field is constructed in three-dimensional space based on a hierarchical set of three-dimensional Gaussian bodies. The occupancy contribution value of each level of three-dimensional Gaussian body is calculated for any spatial query point. The occupancy contribution values ​​of each level of three-dimensional Gaussian body are then fused hierarchically to obtain the continuous collision probability value of the spatial query point. Perform normalization mapping on the continuous collision probability values ​​of spatial query points and encapsulate them into a continuous collision probability field.

[0013] Optionally, the path planning optimization module includes: Based on the continuous collision probability field, the corresponding starting position point and target position point are extracted from the starting pose and target pose of the inspection mission, and a sparse graph search node candidate set is established in the workspace of the inspection environment. Based on the continuous collision probability field, perform connectivity analysis on the candidate set of search nodes in the sparse graph to determine the walkable connection relationship between candidate nodes and generate an inspected sparse connected graph. Connect the starting point and the target point to the sparse connected graph of the inspection, and calculate the edge value corresponding to each passable connection in the sparse connected graph of the inspection. Based on the sparse connected graph of the inspection, the starting point, the target point, and the edge values ​​corresponding to each traversable connection, the inspection sparse connected graph search is performed to obtain the initial inspection path formed by the sequential connection of multiple path nodes. Read the initial inspection path and continuous collision probability fields, resample the initial inspection path according to the preset arc length interval, and generate an initial optimized path sampling sequence formed by connecting multiple path optimization sampling points in sequence. Based on the differentiability of the continuous collision probability field, the continuous collision probability value and continuous collision probability gradient at each path optimization sampling point are calculated, and the collision cost term of the path optimization sampling sequence is constructed. Based on the positional relationship between adjacent path optimization sampling points in the initial optimized path sampling sequence, a path smoothing cost term and a path length cost term are constructed, and then weighted and summed with the collision cost term to obtain the gradient-driven trajectory optimization objective function. The gradient-driven trajectory optimization objective function is solved iteratively. The positions of each path optimization sampling point are updated according to the direction of the continuous collision probability gradient and the direction of the path smoothing constraint, while keeping the starting position and the target position fixed, to obtain the updated path optimization sampling sequence. Perform collision-free constraint determination and convergence determination on the updated path optimization sampling sequence, and output the updated path optimization sampling sequence that satisfies the collision-free constraint and convergence condition as the collision-free optimized inspection path.

[0014] Optionally, the online replanning module includes: The starting point, path node sequence, and target point in the collision-free optimized inspection path are sequentially expanded, and the path node sequence is discretely sampled according to a preset arc length interval to obtain the path dynamics sampling sequence. Based on the positional relationship between adjacent path dynamics sampling points in the path dynamics sampling sequence, calculate the discrete acceleration vector, the acceleration rate vector, and the discrete curvature index; Based on the Euclidean distance between adjacent path dynamics sampling points in the path dynamics sampling sequence and the preset reference running speed, the sampling time interval between each path dynamics sampling point in the path dynamics sampling sequence is calculated, and a smooth inspection path generation objective function is constructed based on the discrete curvature index, discrete acceleration vector, and acceleration change rate vector. The objective function for generating the smooth inspection path is solved by constraint optimization. While keeping the starting and target positions fixed, the positions of the intermediate path dynamic sampling points are updated to generate the smooth inspection path. The path tracking reference sequence corresponding to the smooth inspection path is then sent to the robot control system. During robot execution, new multimodal perception data is read, and the continuous collision probability field is locally incrementally updated in combination with the robot's current execution pose to obtain the updated continuous collision probability field. Based on the updated continuous collision probability field, online safety checks and online replanning are performed on the smooth patrol path to complete real-time autonomous patrol in indoor environments.

[0015] Optionally, the online safety verification and online replanning includes: when there are path dynamic sampling points in the smooth patrol path that do not meet the safety constraints, the position point corresponding to the robot's current execution pose is used as the new starting position point, and the position point corresponding to the target pose of the patrol task is used as the target position point, and the updated smooth patrol path is regenerated.

[0016] The beneficial effects of this invention are: This invention introduces a patrol-sensitive response mechanism during the mapping phase and integrates boundary support values, slender obstacle support values, multi-frame visible stability values, and structural preservation contributions throughout the feature extraction, pose estimation, and point cloud sparsification processes. This ensures consistent enhancement of slender obstacle edges, passageway boundaries, and stable structural regions in feature matching weights, graph optimization residual weights, and point cloud preservation strategies, achieving navigation semantic-driven mapping. Compared to traditional methods based on uniform voxel filtering and indiscriminate feature extraction, this invention prioritizes the preservation of slender structures in sparse 3D point clouds by weighting the structural preservation contribution, preventing the accidental deletion of key obstacles during downsampling and improving the mapping integrity and pose estimation stability of slender obstacles in indoor environments.

[0017] This invention transforms the discrete point cloud occupancy representation into a differentiable continuous probability expression by constructing a continuous collision probability field based on a three-dimensional Gaussian volume parameter set. During the initialization and joint optimization of the three-dimensional Gaussian volume, boundary support values, slender obstacle support values, and structural retention contributions are introduced to couple and adjust scale parameters, opacity parameters, and optimization residuals. This enables key navigation regions to form a continuous expression with high occupancy intensity in the continuous collision probability field, allowing for continuous calculation of collision probabilities at any spatial location. Furthermore, a smooth and physically consistent collision probability distribution is constructed through exponential decay and hierarchical fusion mechanisms, providing differentiable risk gradient information during the path planning stage. This effectively solves the problems of discontinuous collision boundaries, path edge-hugging, and non-differentiable risks in discrete maps.

[0018] This invention constructs an integrated framework for graph-volume joint incremental optimization and continuous collision probability-driven path optimization. It introduces unified navigation enhancement weights and continuous collision probability gradients during pose graph constraints, 3D Gaussian volume parameter updates, and path planning, achieving deep coupling between mapping, environment modeling, and path planning. In the path optimization stage, by performing gradient calculations on the continuous collision probability field, collision risk is directly transformed into guiding information for path optimization. This information, along with constraints on path curvature continuity, velocity change rate, and acceleration change rate, constructs a unified optimization objective. This allows the path to actively avoid high-risk areas while satisfying robot dynamics constraints, realizing a closed-loop optimization mechanism from perception to planning. In dynamic environments, incremental updates of the continuous collision probability field and online replanning strategies enable the path to respond to environmental changes in real time, improving the safety, robustness, and real-time performance of indoor patrol tasks. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an indoor patrol path construction system based on robot vision recognition proposed in this invention. Detailed Implementation

[0020] Example 1: Reference Figure 1 An indoor patrol path construction system based on robot vision recognition includes: The multimodal acquisition module collects multimodal sensing data of the indoor environment and performs time synchronization and calibration processing on the multimodal sensing data to obtain synchronized and calibrated multimodal sensing data. In this embodiment, the multimodal acquisition module includes: Collect multimodal sensing data of the indoor environment and construct the original multimodal sensing dataset; Perform a unified time base transformation on the original timestamps corresponding to various sensors in the original multimodal sensing dataset to obtain standard timestamps; Using a reference time sequence under a unified time base as the alignment benchmark, cross-sensor time matching is performed on standard timestamps to determine the alignment sampling index of various sensors at each reference time. The observation data of various sensors obtained through time matching are calibrated, and the intrinsic parameter matrix and extrinsic parameter transformation matrix of each type of sensor are obtained respectively. Based on the aligned sampling index, intrinsic parameter matrix and extrinsic parameter transformation matrix, a unified spatiotemporal coordinate representation is performed on the multimodal observation content corresponding to each reference time, generating a synchronously calibrated multimodal sensing dataset.

[0021] An improved RTAB module is used to process synchronously calibrated multimodal perception data in real time based on an improved RTAB-Map model, thereby obtaining the robot's six-degree-of-freedom pose sequence and sparse 3D point cloud. In this embodiment, the RTAB module is improved, including: The synchronously calibrated multimodal perception dataset is grouped according to the reference time sequence to construct a synchronous observation frame set. For each synchronous observation frame, slender obstacle edges, passage area boundaries, and stable structure regions are constructed in relation to the robot's indoor inspection path. Inspection sensitivity response values ​​are calculated to obtain an inspection sensitivity response map. In Example 1, the synchronous observation frame set consists of synchronous observation frames corresponding to all reference times, and each synchronous observation frame consists of standardized observation data from various sensors corresponding to the same reference time.

[0022] For each synchronous observation frame, observation sampling units are divided according to a preset window or a preset point cloud neighborhood. The weighted change of depth difference and gray level difference between the observation sampling unit and its neighboring observation sampling units is calculated to obtain the obstacle edge response value, which is used to represent the position of the slender obstacle edge, the boundary of the passage area, and the structural outline.

[0023] By comparing the spatial position of the same observation sampling unit in the standard observation content corresponding to different sensors, the cross-modal structural consistency response value is obtained, which is used to represent the stability of the observation sampling unit under multimodal observation.

[0024] By statistically analyzing the proportion of continuously reachable observation sampling units that meet the unobstructed condition in the spatial neighborhood, the passability continuity response value is obtained, which is used to represent the extent of the extension of the continuously passable area during the construction of the robot's indoor inspection path.

[0025] The obstacle edge response value, cross-modal structural consistency response value, and passage continuity response value are weighted and summed to obtain the inspection sensitive response value. The inspection sensitive response values ​​corresponding to all observation sampling units in each synchronous observation frame are spatially arranged to form an inspection sensitive response map.

[0026] Based on the inspection-sensitive response map, inspection-sensitive features are filtered for each synchronous observation frame to generate a subset of inspection-sensitive local feature descriptions, an inspection-sensitive appearance description vector, and an inspection-passage topology description vector. In Example 1, a set of local features for inspection is obtained by filtering the local features corresponding to the observation sampling units whose inspection-sensitive response values ​​are greater than or equal to the inspection-sensitive feature filtering threshold. The inspection-sensitive feature filtering threshold is obtained by statistically analyzing the inspection-sensitive response values ​​of all observation sampling units in the current synchronous observation frame and selecting the response values ​​that meet the preset retention ratio as the threshold.

[0027] If the number of local features in the generated inspection-sensitive local feature set is lower than the preset minimum threshold for maintaining robust pose estimation, then regular local features are extracted in the non-sensitive area of ​​the current synchronous observation frame, and the inspection-sensitive weight of the regular local features is set to the preset lower limit value and then incorporated into the inspection-sensitive local feature set to ensure the robustness of pose calculation in weak texture or open environment.

[0028] Each local feature in the patrol-sensitive local feature set is multiplied by its corresponding patrol-sensitive weight, and a local feature descriptor is extracted to form a patrol-sensitive local feature descriptor set, which is used for point-to-point feature matching between adjacent frames. At the same time, bag-of-words quantization and statistical aggregation are performed on the local feature descriptors with patrol-sensitive weights to obtain a global patrol-sensitive appearance description vector. The patrol-sensitive weight is obtained by dividing the patrol-sensitive response value corresponding to the local feature by the sum of the patrol-sensitive response values ​​corresponding to all local features in the current synchronous observation frame.

[0029] The synchronous observation frame is divided into multiple directional partitions, and the average value of the traffic continuity response of all observation sampling units in each directional partition is calculated. The values ​​are then arranged in order to obtain the patrol traffic topology description vector.

[0030] Weighted feature matching is performed on the sensitive local feature descriptor subsets of adjacent reference times, and the relative pose transformation matrix between adjacent reference times is obtained by combining the spatial observation content in the synchronously calibrated multimodal sensing dataset. In Example 1, the similarity matching of local features between two reference times is performed based on the patrol sensitive local feature description subset, and an initial matching feature pair set is obtained. For each matching feature pair in the initial matching feature pair set, the pairing weight of the matching feature pair is calculated according to its corresponding patrol sensitive weight.

[0031] Based on the synchronously calibrated multimodal perception dataset, the observation points corresponding to each matching feature pair are converted into spatial feature points in the robot body coordinate system. Spatial point pairs of matching feature pairs in the robot body coordinate system are constructed through spatial feature points. The weighted spatial coordinate error is minimized by taking all spatial point pairs of matching feature pairs as constraints. The weighted spatial coordinate error is obtained by weighting and squaring the difference between the pairing weight of each matching feature pair and the spatial coordinates of the matching feature pair after rotation and translation transformations at the current reference time and the previous reference time. The rotation and translation transformation results obtained are combined to generate the relative pose transformation matrix between adjacent reference times.

[0032] ; ; in, Indicates the first The reference time and the first Between the reference times The pairing weights for each matching feature are jointly determined by the patrol sensitivity weights corresponding to the two reference times. These weights are used to ensure higher constraint strength for slender obstacle edges, passageway boundaries, and stable structural regions in relative pose estimation. Indicates from the first The reference time to the 1st The relative rotation matrix at each reference time. Indicates from the first The reference time to the 1st The relative translation vector at each reference time, with units of meters. Indicates the first The reference time and the first The number of effective matching features between reference times, Indicates the first The first reference time corresponding to the first reference time The three-dimensional coordinates of each matching spatial feature point in the robot's body coordinate system. Indicates the first The first reference time corresponding to the first reference time The three-dimensional coordinates of each matching spatial feature point in the robot's body coordinate system. This represents the relative pose transformation matrix.

[0033] Pose accumulation is performed based on the relative pose transformation matrix between adjacent reference times to obtain the initial six-degree-of-freedom robot pose sequence; In Example 1, the pose matrix at the first reference time is set as the identity matrix, and the pose matrices at each reference time are multiplied sequentially by the pose matrix at the previous reference time and the corresponding relative pose transformation matrix to obtain the initial six-degree-of-freedom pose sequence of the robot.

[0034] Perform closed-loop detection based on the inspection-sensitive appearance description vector and the inspection-traffic topology description vector to determine the inspection constraint closed-loop set; In Example 1, the similarity value of the inspection constraint closed loop between any two reference times is calculated and compared with the inspection constraint closed loop judgment threshold to obtain the inspection constraint closed loop set.

[0035] The similarity value of the inspection constraint closed loop is obtained by weighted summation of the normalized inner product of the inspection sensitive appearance description vectors at two reference times and the normalized inner product of the inspection passage topology description vectors.

[0036] By statistically analyzing the similarity values ​​of closed-loop and non-closed-loop samples confirmed during historical inspections, the similarity threshold that minimizes the classification error between closed-loop and non-closed-loop samples is selected, thus obtaining the inspection constraint closed-loop determination threshold.

[0037] Based on the initial robot six-DOF pose sequence, the relative pose transformation matrix between adjacent reference times, and the set of roving constraint closed loops, a roving constraint pose graph is constructed, and global optimization is performed on the roving constraint pose graph to obtain the optimized robot six-DOF pose sequence. In Example 1, the patrol constraint pose graph consists of pose nodes corresponding to multiple reference times and constraint edges connecting the pose nodes. The initial values ​​of the pose nodes are provided by the initial robot six-degree-of-freedom pose sequence. The constraint edges include patrol constraint odometry constraint edges formed by the relative pose transformation matrix between adjacent reference times and patrol constraint closed loop constraint edges formed by the patrol constraint closed loop set.

[0038] The process of performing global optimization on the inspection constraint pose graph includes: using the initial robot six-DOF pose sequence as the initial solution of the optimization variables, calculating the translational and rotational residuals of each constraint edge under the current pose estimation; introducing inspection weighting coefficients during the calculation of the translational and rotational residuals of each constraint edge, which are obtained by calculating the average inspection sensitivity response value corresponding to two reference times, and used to adjust the influence of each constraint edge in the overall optimization; accumulating the weighted squares of the translational and rotational residuals of all constraint edges according to the corresponding inspection weighting coefficients to construct the global optimization objective function; iteratively updating the pose nodes corresponding to each reference time based on the initial robot six-DOF pose sequence, so that the value of the global optimization objective function gradually decreases until the convergence condition is met, and obtaining the optimized robot six-DOF pose sequence.

[0039] Based on the optimized robot six-degree-of-freedom pose sequence, a unified coordinate transformation is performed on the spatial observation points corresponding to each reference time. The boundary support value, slender obstacle support value, and multi-frame visible stability value of each spatial observation point in the construction of the robot's indoor inspection path are calculated and weighted and fused into a structure retention contribution. Based on the structure retention contribution, sparse aggregation is performed on the spatial observation points after the unified coordinate transformation to obtain a sparse three-dimensional point cloud.

[0040] In Example 1, the optimized pose matrix corresponding to the six-DOF pose sequence of the robot at the reference time is read. The three-dimensional coordinates of the spatial observation points in the robot body coordinate system are transformed to the global coordinate system through the corresponding optimized pose matrix to obtain the three-dimensional coordinates of the spatial observation points in the global coordinate system. The unified coordinate transformation process is repeated for all spatial observation points corresponding to all reference times. The three-dimensional coordinates of all spatial observation points corresponding to all reference times in the global coordinate system are collected. Then, a voxel mesh-based presampling and downsampling filter is performed to remove spatially redundant dense clusters of points to obtain the full three-dimensional point cloud.

[0041] For each spatial observation point in the full 3D point cloud, calculate the boundary support value, slender obstacle support value, and multi-frame visible stability value of the spatial observation point.

[0042] Boundary support values ​​are obtained by statistically analyzing the degree of local depth and local direction changes between a spatial observation point and its neighboring spatial observation points at the same reference time. These values ​​are used to identify the degree of support that a spatial observation point provides to the boundary of the passable area and the outline of obstacles.

[0043] The slender obstacle support value is obtained by statistically analyzing the degree of continuous distribution along the main extension direction and the degree of compact distribution along the perpendicular main extension direction in the local neighborhood of the space observation point. It is used to identify the degree to which the space observation point preserves slender obstacles such as cables, poles and thin table legs.

[0044] The multi-frame visibility stability value is obtained by statistically analyzing the number of times a spatial observation point is repeatedly observed at different reference times and the consistency of the repeated observation locations. It is used to represent the degree of repeated visibility of a spatial observation point in continuous patrol observation.

[0045] For each space observation point, the boundary support value, the slender obstacle support value, and the multi-frame visible stability value are weighted and fused to obtain the structural preservation contribution of the space observation point.

[0046] Sparse aggregation is performed on the spatial observation points in the full 3D point cloud according to the reference time sequence. For the current spatial observation point: If the structural retention contribution of a spatial observation point is greater than or equal to the structural retention contribution screening threshold, then the spatial observation point will be directly retained as a candidate sparse point. If the structural retention contribution of a spatial observation point is less than the structural retention contribution screening threshold, the minimum distance between the spatial observation point and historically retained spatial observation points is calculated. When the minimum distance is greater than or equal to the sparse aggregation distance threshold, the spatial observation point is retained as a candidate sparse point. When the minimum distance is less than the sparse aggregation distance threshold, the spatial observation point is discarded.

[0047] The sparse aggregation process is repeated for all spatial observation points to gather all candidate sparse points and obtain a sparse three-dimensional point cloud.

[0048] The improved RTAB-Map model's front-end mapping architecture in this implementation mainly consists of a patrol-sensitive feature extraction module, a weighted visual odometry front-end, a task constraint graph optimization back-end, and a dual-track point cloud sparsification module, all cascaded sequentially. In terms of data flow and hierarchical connections, the input, synchronously calibrated multimodal sensing data is converted into a patrol-sensitive response map in the patrol-sensitive feature extraction module, and based on this, a subset of local feature descriptors with navigation weights, a patrol-sensitive appearance description vector, and a patrol topology description vector are generated. The subset of local feature descriptors with navigation weights flows into the weighted visual odometry front-end, where the phase difference between adjacent reference times is calculated by minimizing spatial errors through weighting. The pose transformation matrix is ​​processed and accumulated at each time step to form an initial six-DOF robot pose sequence. Simultaneously, the inspection-sensitive appearance description vector and the inspection-travel topology description vector are fed into the task constraint graph optimization backend to trigger closed-loop detection. The generated closed-loop constraint edges and odometry constraint edges are combined with the inspection weighting coefficients in the task constraint graph optimization backend to construct a global optimization objective function, and iteratively output a globally consistent optimized six-DOF robot pose sequence. The optimized six-DOF robot pose sequence and the spatial observation points corresponding to each reference time are input into the dual-track point cloud sparsification module. After joint filtering by structure preservation contribution and spatial distance, a sparse three-dimensional point cloud is output for initialization of the three-dimensional Gaussian volume parameter set.

[0049] Compared with the traditional general RTAB-Map model based on pure environmental appearance, the core improvement of this implementation method lies in breaking the current technical status quo of separating mapping and downstream obstacle avoidance tasks. It deeply injects the key geometric attributes of slender obstacle edges and passage boundaries into feature extraction, graph optimization residual weight allocation, and the whitelist retention mechanism for point cloud downsampling. The important role and significant advantage of this improvement is that it enables the front-end pose calculation and back-end optimization to actively tilt towards high-risk collision areas at the mathematical level. It completely solves the industry pain point that traditional voxel filtering tends to filter out slender key obstacle point clouds as noise. While ensuring that features do not degrade in extreme environments, it provides high-quality prior poses and structural seed points with absolute accuracy on the physical boundary for the construction of 3D Gaussian continuous collision fields.

[0050] The Gaussian model initialization module uses the robot's six-degree-of-freedom pose sequence and sparse three-dimensional point cloud as priors to initialize the three-dimensional Gaussian volume parameter set and construct the initial three-dimensional Gaussian sputtering model. In this embodiment, the Gaussian model initialization module includes: Read the robot's six-degree-of-freedom pose sequence and sparse three-dimensional point cloud, establish the initial correspondence of the three-dimensional Gaussian body point by point according to the sparse three-dimensional points in the sparse three-dimensional point cloud, and generate a seed set of three-dimensional Gaussian bodies. In Example 1, the boundary support value, slender obstacle support value, multi-frame visible stability value, and structure preservation contribution value inherited by each sparse 3D point during the sparse aggregation process are synchronously written into the corresponding 3D Gaussian seed.

[0051] Read the initial value of the center position of the seed of the 3D Gaussian body, and search for the local neighborhood sparse 3D point set in the sparse 3D point cloud with the initial value of the center position as the center. Construct a local structure matrix, and initialize the initial value of the principal axis direction parameter of the corresponding 3D Gaussian body based on the eigenvalue decomposition result of the local structure matrix. In Example 1, a set of sparse 3D points in the sparse 3D point cloud is obtained by selecting all sparse 3D points whose Euclidean distance from the initial value of the center position of the current 3D Gaussian seed is not greater than the corresponding local neighborhood search radius.

[0052] The local structure matrix is ​​obtained by averaging the outer product of the offset vectors between each sparse 3D point in the local neighborhood sparse 3D point set and the initial value of the center position of the current 3D Gaussian body seed.

[0053] The eigenvectors obtained by performing eigenvalue decomposition on the local structure matrix determine the initial values ​​of the principal axis direction parameters.

[0054] Read the support value of the slender obstacle and the contribution of structure preservation in the seed of the three-dimensional Gaussian body, and combine the eigenvalues ​​obtained by the seed of the three-dimensional Gaussian body in the process of local structure matrix eigenvalue decomposition to initialize the initial values ​​of the scale parameters of the corresponding three-dimensional Gaussian body along the three principal axes. In Example 1, the purpose of initializing the scale parameters of the corresponding three-dimensional Gaussian body along the three principal axes is to elongate the three-dimensional Gaussian body corresponding to the slender obstacle in the principal extension direction and compress it in the direction perpendicular to the principal extension direction.

[0055] The initial value of the scale parameter in the first principal axis direction is obtained by amplifying the square root of the first eigenvalue. The amplification degree is determined by the support value of the slender obstacle and the contribution of the structure preservation, so that the larger the support value of the slender obstacle, the larger the scale of the three-dimensional Gaussian body in the principal extension direction.

[0056] The initial values ​​of the scale parameters in the second and third principal axis directions are obtained by compressing the square root of the corresponding eigenvalues. The degree of compression is determined by the inverse adjustment term of the slender obstacle support value and the contribution of structural preservation, so that the larger the slender obstacle support value, the smaller the scale of the three-dimensional Gaussian body in the vertical principal extension direction.

[0057] All initial values ​​of scale parameters are superimposed on the corresponding eigenvalues ​​with a proportional term of the sparse aggregation distance threshold to ensure the minimum spatial coverage of the corresponding three-dimensional Gaussian volume.

[0058] Based on the initial values ​​of the scale parameters and principal axis parameters of the corresponding 3D Gaussian body seed along the three principal axes, the initial values ​​of the covariance matrix of the corresponding 3D Gaussian body are constructed. In Example 1, the initial values ​​of the scale parameters in the three principal axes are squared and then arranged in the corresponding principal axes to form a diagonal scale matrix. The initial values ​​of the principal axis parameters are used as the coordinate rotation basis, and the diagonal scale matrix is ​​mapped from the principal axis coordinate system to the global coordinate system. The initial value of the covariance matrix of the corresponding three-dimensional Gaussian body is obtained by combining the initial values ​​of the principal axis parameters, the diagonal scale matrix, and the transpose matrix of the initial values ​​of the principal axis parameters in sequence through matrix multiplication.

[0059] The initial value of the covariance matrix is ​​used to characterize the spatial distribution range and directionality of the corresponding 3D Gaussian volume in the global coordinate system, so that the corresponding 3D Gaussian volume has an anisotropic expression ability consistent with the local spatial distribution direction at the edges of slender obstacles, the boundaries of passage areas, and stable structural regions.

[0060] Based on the robot's six-degree-of-freedom pose sequence and a synchronously calibrated multimodal perception dataset, the visible observation set of the three-dimensional Gaussian body corresponding to the seed of the three-dimensional Gaussian body at each reference time is calculated, and the initial value of the color parameter of the three-dimensional Gaussian body corresponding to the seed of the three-dimensional Gaussian body is initialized according to the visible observation set. In Example 1, the pose matrix corresponding to each reference time is read, and the initial value of the center position of the current 3D Gaussian seed is transformed from the global coordinate system to the sensor observation coordinate system corresponding to each reference time. Imaging projection is performed on the transformed initial value of the center position, and it is determined whether the projected position falls within the effective observation range corresponding to the current reference time. Combined with the depth image of the current reference time in the synchronously calibrated multimodal sensing dataset, the actual observation depth at the projected position is extracted. If the theoretical depth value of the initial value of the center position in the sensor observation coordinate system is not greater than the actual observation depth value plus the set tolerance threshold, it is determined that the unobstructed condition is met. All reference times that meet the visibility condition are collected to obtain the visible observation set of the 3D Gaussian corresponding to the current 3D Gaussian seed.

[0061] For each reference time in the visible observation set, the color observation value in the neighborhood of the projection position is read from the standard observation content corresponding to the current reference time, and the color observation values ​​corresponding to all reference times are averaged and fused to obtain the initial value of the color parameter of the three-dimensional Gaussian body corresponding to the current three-dimensional Gaussian body seed. The initial value of the color parameter is written into the three-dimensional Gaussian body seed as the zeroth order DC component of the spherical harmonic function, and the higher order AC component of the corresponding spherical harmonic function is initialized to zero.

[0062] Based on the boundary support value and multi-frame visible stability value in the 3D Gaussian seed, initialize the initial value of the opacity parameter of the 3D Gaussian corresponding to the 3D Gaussian seed. In Example 1, the multi-frame visible stable value is used as the basic occupancy intensity index. Based on the basic occupancy intensity index, the boundary support value is introduced as a boundary enhancement adjustment term to enhance the basic occupancy intensity index, so that the three-dimensional Gaussian volume located at the boundary of the passage area and the outline of the obstacle has a higher initial value of the opacity parameter.

[0063] By mapping the occupancy intensity index adjusted by the boundary support value to a preset opacity value range, the initial value of the opacity parameter of the corresponding 3D Gaussian body is obtained. This results in 3D Gaussian bodies with larger boundary support values ​​and higher multi-frame visibility stability values ​​forming high occupancy expression regions, while 3D Gaussian bodies with smaller boundary support values ​​or lower multi-frame visibility stability values ​​forming low occupancy expression regions.

[0064] The initial values ​​of the center position, covariance matrix, principal axis direction parameter, color parameter, and opacity parameter of each 3D Gaussian volume seed are collected to generate the initial value set of 3D Gaussian volume parameters. Based on the initial value set of 3D Gaussian volume parameters, the robot's six-degree-of-freedom pose sequence, and the sparse 3D point cloud, an initial 3D Gaussian sputtering model is constructed.

[0065] The graph-volume joint optimization module performs graph-volume joint incremental optimization on the initial 3D Gaussian sputtering model and the robot's six-DOF pose sequence to obtain the optimized robot's six-DOF pose sequence and the optimized 3D Gaussian volume parameter set. In this embodiment, the map-body joint optimization module includes: Based on the initial 3D Gaussian sputtering model, an incremental optimization window is constructed according to the current reference time, and the set of pose nodes and the set of visible 3D Gaussian volumes within the incremental optimization window are determined. In Example 1, taking the current reference time as the center, a preset number of consecutive reference times are selected forward according to the time sequence of the reference time sequence to obtain the incremental optimization window. The pose matrix corresponding to each reference time in the incremental optimization window constitutes the pose node set.

[0066] For each 3D Gaussian volume in the initial 3D Gaussian sputtering model, based on the pose matrix corresponding to each reference time within the incremental optimization window, the center position parameter of the 3D Gaussian volume is transformed from the global coordinate system to the sensor observation coordinate system corresponding to each reference time. Imaging projection is then performed on the transformed center position parameter in the sensor observation coordinate system corresponding to each reference time. It is determined whether the projected position falls within the effective observation range of the corresponding reference time. In conjunction with the depth observation information in the synchronously calibrated multimodal sensing dataset, the consistency between the theoretical depth value and the actual observed depth value at the projected position is judged. When the theoretical depth of a 3D Gaussian body is not greater than the actual observed depth plus a preset tolerance threshold, the 3D Gaussian body is determined to meet the unobstructed condition at the corresponding reference time.

[0067] When a certain 3D Gaussian volume satisfies the no-occlusion condition at at least one reference time within the incremental optimization window, the corresponding 3D Gaussian volume is included in the set of visible 3D Gaussian volumes.

[0068] Based on the boundary support value, slender obstacle support value, multi-frame visible stability value, and structure preservation contribution inherited by each 3D Gaussian volume in the visible 3D Gaussian volume set, the navigation enhancement weight of the 3D Gaussian volume is calculated, and the color observation residual and depth observation residual are constructed based on the standard observation content corresponding to each reference time within the incremental optimization window. In Example 1, for each 3D Gaussian volume in the set of visible 3D Gaussian volumes, the boundary support value, slender obstacle support value, multi-frame visible stability value, and structure preservation contribution value inherited during the initialization phase of the 3D Gaussian volume seed set are read, and then multiplied by the corresponding fusion coefficients based on the basic weight value one and weighted summed to obtain the 3D Gaussian volume navigation enhancement weight.

[0069] Each 3D Gaussian body in the visible 3D Gaussian body set is projected onto the imaging plane corresponding to each reference time within the incremental optimization window. Sampling pixels are selected within the effective observation range corresponding to each reference time. For each sampling pixel, all 3D Gaussian bodies that participate in the rendering of that sampling pixel are counted, and the navigation enhancement weights of the 3D Gaussian bodies corresponding to all 3D Gaussian bodies that participate in the rendering of that sampling pixel are averaged to obtain the navigation enhancement observation weight of the sampling pixel.

[0070] For each sampled pixel, the navigation enhancement observation weight is multiplied by the square norm of the difference between the rendered color value and the actual observed color value at the sampled pixel to obtain the color observation residual.

[0071] For each sampled pixel, the navigation enhancement observation weight is multiplied by the square of the difference between the rendered depth value and the actual observed depth value at the sampled pixel to obtain the depth observation residual.

[0072] Based on color observation residuals, depth observation residuals, patrol constraint odometer constraint edges, and patrol constraint closed-loop constraint edges, a graph-volume joint incremental optimization objective function is constructed. Specifically, the color observation residuals corresponding to all color sampling pixels within the incremental optimization window are divided by the square of the color observation residual standardization coefficient and then summed to obtain the color observation term.

[0073] The depth observation term is obtained by summing the depth observation residuals corresponding to all depth sampling pixels within the incremental optimization window by the square of the depth observation residual standardization coefficient.

[0074] The translation and rotation residuals corresponding to the patrol constraint odometer constraint edge and the patrol constraint closed loop constraint edge are weighted, squared, and accumulated to construct the pose graph constraint term.

[0075] ; in, This represents the pose graph constraint term, used to measure the overall error of the robot's six-DOF pose sequence under the combined influence of the odometry constraint edges and the closed-loop constraint edges of the odometry constraint. This represents a set of constraint edges in the inspection constraint pose graph. This represents the set of constraints on the patrol odometer. This represents the set of closed-loop constraint edges for the inspection constraint. This represents the weighting factor for the inspection. Indicates the first constraint edge of the patrol constraint odometer. The reference time and the first The translational residual vector between reference times, This represents the standardized coefficient of the translation residual corresponding to the constrained edge of the inspection odometer. Indicates the first constraint edge of the patrol constraint odometer. The reference time and the first Rotational residual vector between reference times This represents the normalized coefficient of the rotational residual corresponding to the constraint edge of the inspection constraint odometer. The first constraint edge in the closed-loop constraint of the inspection is represented by the... The reference time and the first The translational residual vector between reference times, This represents the standardized coefficient of the translation residual corresponding to the closed-loop constraint edge of the inspection constraint. The first constraint edge in the closed-loop constraint of the inspection is represented by the... The reference time and the first Rotational residual vector between reference times This represents the normalized coefficient of the rotational residual corresponding to the closed-loop constraint edge of the inspection constraint. Operators for Euclidean norms.

[0076] The navigation enhancement weights of the 3D Gaussian volume are multiplied by the sum of the two regularization results to obtain the navigation key structure regularization components corresponding to the 3D Gaussian volume. The first regularization result is obtained by squaring the differences between the current scale parameter and the initial value of the corresponding scale parameter along the three principal axes of the 3D Gaussian volume, dividing each difference by the square of the standardization coefficient of the scale parameter along the corresponding principal axis, and then summing the results. The second regularization result is obtained by squaring the difference between the current opacity parameter and the initial value of the opacity parameter of the 3D Gaussian volume, and dividing the result by the square of the standardization coefficient of the opacity parameter.

[0077] Repeat the above process for all visible 3D Gaussian volumes and sum them up to obtain the navigation key structure regularization term.

[0078] The color observation term, depth observation term, pose graph constraint term, and navigation key structure regularization term are multiplied by their respective weight coefficients and then summed in a weighted manner to obtain the graph-volume joint incremental optimization objective function.

[0079] The graph-volume joint incremental optimization objective function is solved by alternating iterations, updating the pose node parameters and the 3D Gaussian volume parameters. If the change in the graph-volume joint incremental optimization objective function between two consecutive iterations is not greater than the preset convergence threshold, the iteration is stopped. In Example 1, in each iteration, the current three-dimensional Gaussian volume parameters are first fixed, and the pose nodes corresponding to each reference time within the incremental optimization window are incrementally updated. Then, the updated pose nodes are fixed, and the parameters of each three-dimensional Gaussian volume in the visible three-dimensional Gaussian volume set are incrementally updated.

[0080] After the incremental optimization window has converged, the updated pose node parameters are written back to the robot's six-DOF pose sequence, and the updated 3D Gaussian volume parameters are written back to the 3D Gaussian volume parameter set, thus obtaining the optimized robot six-DOF pose sequence and the optimized 3D Gaussian volume parameter set.

[0081] The updated pose node parameters are written back to the robot's six-DOF pose sequence, and the updated 3D Gaussian volume parameters are written back to the 3D Gaussian volume parameter set, keeping the pose node parameters that have not entered the incremental optimization window and the 3D Gaussian volume parameters that have not entered the visible 3D Gaussian volume set unchanged.

[0082] The hierarchical Gaussian management module performs hierarchical Gaussian body management based on the optimized three-dimensional Gaussian body parameter set, generates a hierarchical three-dimensional Gaussian body set, and constructs a continuous collision probability field in three-dimensional space based on the hierarchical three-dimensional Gaussian body set to obtain the continuous collision probability field. In this embodiment, the hierarchical Gaussian management module includes: Based on the optimized set of three-dimensional Gaussian volume parameters, the visibility frequency and visibility distribution of each three-dimensional Gaussian volume during the observation process at multiple reference times are statistically analyzed, and the information entropy value and visibility value of the corresponding three-dimensional Gaussian volume are calculated by combining the opacity parameter and covariance matrix of the three-dimensional Gaussian volume. In Example 1, for each 3D Gaussian volume, the ratio of its updated opacity parameter to the sum of the updated opacity parameters of all 3D Gaussian volumes is calculated to obtain the opacity percentage. At the same time, the updated covariance matrix is ​​read and the volume scale of the covariance matrix is ​​calculated. The spatial distribution complexity component is obtained by taking the logarithm of the determinant of the covariance matrix. The information entropy component corresponding to the opacity percentage is weighted and summed with the spatial distribution complexity component corresponding to the covariance matrix to obtain the information entropy value.

[0083] The visibility value is calculated by statistically analyzing the visible markers of the current 3D Gaussian body across all reference times and then comparing the number of times each visible marker is marked as visible to the total number of reference times. This visibility value represents the percentage of the current 3D Gaussian body's visible frequency across all reference times.

[0084] Based on the information entropy and visibility values ​​of each 3D Gaussian body, the hierarchical score is calculated, and the 3D Gaussian bodies in the optimized 3D Gaussian body parameter set are hierarchically stored according to the hierarchical score to generate multi-level Gaussian body subsets. In Example 1, the information entropy values ​​of all three-dimensional Gaussian volumes are normalized to obtain the information entropy value normalization result, and the visibility values ​​of all three-dimensional Gaussian volumes are normalized to obtain the visibility value normalization result.

[0085] The information entropy value normalization result and the visibility value normalization result are weighted and summed according to a preset fusion coefficient to obtain the hierarchical score value. The hierarchical score value is used to indicate the retention priority of the current 3D Gaussian body in hierarchical Gaussian body management. According to the preset number of levels and the corresponding level score interval, all 3D Gaussian bodies are divided into different level score intervals based on the hierarchical score value, resulting in multiple levels of Gaussian body index subsets, forming a multi-level Gaussian body subset.

[0086] The preset number of levels and corresponding scoring intervals are determined based on the dynamic balance constraints between the complexity of obstacles and the real-time requirements of onboard computing resources in indoor patrol scenarios. The Gaussian volume is divided into a three- or multi-level architecture consisting of a core structure layer (high-scoring area, preserving key boundaries), an environmental detail layer (medium-scoring area, providing navigation reference), and a redundant noise layer (low-scoring area, performing aggressive pruning). A non-uniform spatial resolution expression mechanism is established through the level scoring intervals, enabling the robot control system to prioritize recalling the continuous collision field analytical expressions in the high-level subset when performing path planning queries to ensure absolute safety of obstacle avoidance. The low-level subset is dynamically loaded according to the computing power redundancy to smooth the local collision gradient. Under the premise of ensuring that slender obstacles are not lost, the real-time query computing power consumption of the continuous collision probability field is controlled at a constant level.

[0087] Based on the hierarchical score value, updated opacity parameter and structural retention contribution of the multi-level Gaussian body subset, the 3D Gaussian bodies in each level are clipped and filtered to retain the 3D Gaussian bodies corresponding to the navigation key boundaries and clip the low-contribution 3D Gaussian bodies to generate a hierarchical 3D Gaussian body set. In Example 1, the hierarchical score, the updated opacity parameter, and the structural retention contribution are weighted and summed according to a preset fusion coefficient to obtain the clipping retention score. For each 3D Gaussian body in each level, the clipping retention score is compared with the clipping retention threshold of the corresponding level. When the clipping retention score is greater than or equal to the clipping retention threshold, the corresponding 3D Gaussian body is retained as a valid Gaussian body in the current level. When the clipping retention score is less than the clipping retention threshold, the corresponding 3D Gaussian body is removed from the current level. The clipping and filtering process is repeated for all levels. The parameters of the 3D Gaussian bodies retained in each level are collected to obtain a hierarchical 3D Gaussian body set.

[0088] A continuous collision probability field is constructed in three-dimensional space based on a hierarchical set of three-dimensional Gaussian bodies. The occupancy contribution value of each level of three-dimensional Gaussian body is calculated for any spatial query point. The occupancy contribution values ​​of each level of three-dimensional Gaussian body are then fused hierarchically to obtain the continuous collision probability value of the spatial query point. In Example 1, the spatial distance weighting term is substituted into the exponential decay function and multiplied by the updated opacity parameter of the corresponding 3D Gaussian body to obtain the occupancy contribution value of the 3D Gaussian body at the current spatial query point.

[0089] For each level, the occupancy contribution values ​​of all 3D Gaussian bodies within that level at the current spatial query point are accumulated to obtain the level occupancy contribution value corresponding to that level. The level occupancy contribution values ​​corresponding to each level are then weighted and summed according to the preset level fusion weight to obtain the continuous collision probability value of the spatial query point. The continuous collision probability value is used to represent the continuous probability intensity of the robot colliding at the spatial query point.

[0090] Perform normalization mapping on the continuous collision probability values ​​of spatial query points and encapsulate them into a continuous collision probability field.

[0091] The path planning and optimization module performs connectivity analysis using the continuous collision probability field and performs sparse graph search to generate an initial inspection path based on the starting pose and target pose of the inspection task. It then performs gradient-driven trajectory optimization on the initial inspection path to obtain a collision-free optimized inspection path. In this embodiment, the path planning optimization module includes: Based on the continuous collision probability field, the corresponding starting position point and target position point are extracted from the starting pose and target pose of the inspection mission, and a sparse graph search node candidate set is established in the workspace of the inspection environment. In Example 1, the position point corresponding to the initial pose of the inspection mission is taken as the starting position point, and the position point corresponding to the target pose of the inspection mission is taken as the target position point. The spatial position is discretely sampled within the workspace of the inspection environment to obtain multiple candidate spatial position points.

[0092] For each candidate spatial location point, the normalized continuous collision probability value of the candidate spatial location point in the continuous collision probability field is read. Candidate spatial location points whose normalized continuous collision probability value is less than or equal to the candidate node collision probability threshold are retained to obtain the sparse graph search node candidate set. The candidate node collision probability threshold is determined by statistically analyzing the normalized continuous collision probability values ​​of all candidate spatial location points in the historical inspection process and selecting the maximum allowable collision probability value that meets the preset safe passage probability quantile constraint.

[0093] Based on the continuous collision probability field, perform connectivity analysis on the candidate set of search nodes in the sparse graph to determine the walkable connection relationship between candidate nodes and generate an inspected sparse connected graph. Connect the starting point and the target point to the sparse connected graph of the inspection, and calculate the edge value corresponding to each passable connection in the sparse connected graph of the inspection. In Example 1, the starting position point and the target position point are merged into the candidate set of search nodes for the sparse connected graph to obtain the search node set for the sparse connected graph.

[0094] For each walkable connection in the sparse connected graph, calculate the Euclidean distance between the two candidate nodes, and read the average and maximum collision probability values ​​of the corresponding connected path. Sum the three probability values ​​with weights to obtain the edge generation value corresponding to the walkable connection.

[0095] Based on the sparse connected graph of the inspection, the starting point, the target point, and the edge values ​​corresponding to each traversable connection, the inspection sparse connected graph search is performed to obtain the initial inspection path formed by the sequential connection of multiple path nodes. Read the initial inspection path and continuous collision probability fields, resample the initial inspection path according to the preset arc length interval, and generate an initial optimized path sampling sequence formed by connecting multiple path optimization sampling points in sequence. Based on the differentiability of the continuous collision probability field, the continuous collision probability value and continuous collision probability gradient at each path optimization sampling point are calculated, and the collision cost term of the path optimization sampling sequence is constructed. In Example 1, for each path optimization sampling point, based on the parameterized expression corresponding to the continuous collision probability field, the partial derivatives of the three components of the spatial coordinates of the path optimization sampling point are calculated respectively to obtain the partial derivative components along the three coordinate directions at the path optimization sampling point. The three partial derivative components are then combined to obtain the continuous collision probability gradient at the path optimization sampling point.

[0096] For each intermediate path optimization sampling point in the initial optimized path sampling sequence, a collision probability barrier mapping process is performed on the corresponding normalized continuous collision probability value to obtain the collision cost value corresponding to each intermediate path optimization sampling point. The collision cost values ​​corresponding to all intermediate path optimization sampling points are accumulated to obtain the collision cost term of the path optimization sampling sequence.

[0097] Based on the positional relationship between adjacent path optimization sampling points in the initial optimized path sampling sequence, a path smoothing cost term and a path length cost term are constructed, and then weighted and summed with the collision cost term to obtain the gradient-driven trajectory optimization objective function. In Example 1, for each intermediate path optimization sampling point in the initial optimized path sampling sequence, the discrete second-order position difference between the current path optimization sampling point and its previous and next path optimization sampling points is calculated, and the square norm of the discrete second-order position difference is accumulated to obtain the path smoothing cost term. The path smoothing cost term is used to characterize the degree of discrete curvature change of adjacent path optimization sampling points in the path optimization sampling sequence.

[0098] For each adjacent path optimization sampling point in the initial optimized path sampling sequence, calculate the Euclidean distance between adjacent path optimization sampling points, and sum all the Euclidean distances to obtain the path length cost term.

[0099] The gradient-driven trajectory optimization objective function is solved iteratively. The positions of each path optimization sampling point are updated according to the direction of the continuous collision probability gradient and the direction of the path smoothing constraint, while keeping the starting position and the target position fixed, to obtain the updated path optimization sampling sequence. Perform collision-free constraint determination and convergence determination on the updated path optimization sampling sequence, and output the updated path optimization sampling sequence that satisfies the collision-free constraint and convergence condition as the collision-free optimized inspection path.

[0100] The online replanning module integrates the collision-free optimized inspection path with robot dynamics constraints to generate a smooth inspection path and send it to the robot control system. During robot execution, the continuous collision probability field is updated, and the smooth inspection path is replanned online to complete real-time autonomous inspection in indoor environments.

[0101] In this embodiment, the online replanning module includes: The starting point, path node sequence, and target point in the collision-free optimized inspection path are sequentially expanded, and the path node sequence is discretely sampled according to a preset arc length interval to obtain the path dynamics sampling sequence. Based on the positional relationship between adjacent path dynamics sampling points in the path dynamics sampling sequence, calculate the discrete acceleration vector, the acceleration rate vector, and the discrete curvature index; In Example 1, for each adjacent path dynamics sampling point in the path dynamics sampling sequence, a discrete velocity vector is obtained by calculating the position difference between the next sampling point and the current sampling point and dividing it by the corresponding sampling interval time. For each adjacent discrete velocity vector in the path dynamics sampling sequence, a discrete acceleration vector is obtained by calculating the difference between the next discrete velocity vector and the current discrete velocity vector and dividing it by the average value of the adjacent sampling time intervals.

[0102] For each adjacent discrete acceleration vector in the path dynamics sampling sequence, the acceleration rate vector is obtained by calculating the difference between the next discrete acceleration vector and the current discrete acceleration vector and dividing it by the average value of the adjacent average sampling time interval.

[0103] For each intermediate path dynamics sampling point in the path dynamics sampling sequence, the discrete second-order position difference between the current path dynamics sampling point and its previous and next sampling points is calculated, and the Euclidean norm of the discrete second-order position difference is divided by the square of the length of the adjacent path segment to obtain the discrete curvature index.

[0104] Based on the Euclidean distance between adjacent path dynamics sampling points in the path dynamics sampling sequence and the preset reference running speed, the sampling time interval between each path dynamics sampling point in the path dynamics sampling sequence is calculated, and a smooth inspection path generation objective function is constructed based on the discrete curvature index, discrete acceleration vector, and acceleration change rate vector. In Example 1, the difference between discrete curvature indices corresponding to adjacent path dynamics sampling points in the path dynamics sampling sequence is squared and accumulated to obtain the curvature continuity cost term. The squares of the Euclidean norms of each discrete acceleration vector in the path dynamics sampling sequence are accumulated to obtain the velocity change rate cost term. The squares of the Euclidean norms of each acceleration change rate vector in the path dynamics sampling sequence are accumulated to obtain the acceleration change rate cost term.

[0105] The path preservation cost term is obtained by summing the squared Euclidean distances between each path dynamics sampling point in the path dynamics sampling sequence and the corresponding collision-free optimized inspection path node.

[0106] The curvature continuity cost, velocity rate of change cost, acceleration rate of change cost, and path preservation cost are multiplied by their respective weighting coefficients and then summed in a weighted manner to obtain the objective function for generating the smooth inspection path.

[0107] The objective function for generating the smooth inspection path is solved by constraint optimization. While keeping the starting and target positions fixed, the positions of the intermediate path dynamic sampling points are updated to generate the smooth inspection path. The path tracking reference sequence corresponding to the smooth inspection path is then sent to the robot control system. During robot execution, new multimodal perception data is read, and the continuous collision probability field is locally incrementally updated in combination with the robot's current execution pose to obtain the updated continuous collision probability field. Based on the updated continuous collision probability field, online safety checks and online replanning are performed on the smooth patrol path to complete real-time autonomous patrol in indoor environments.

[0108] In this embodiment, online safety verification and online replanning are performed, including: when there are path dynamic sampling points in the smooth inspection path that do not meet the safety constraints, the position point corresponding to the robot's current execution pose is used as the new starting position point, and the position point corresponding to the target pose of the inspection task is used as the target position point, and the updated smooth inspection path is regenerated.

[0109] Example 2: In a closed indoor inspection area, the robot needs to start from the inspection starting point, pass through a narrow passage, pass through the boundaries of multiple equipment workstations, and bypass areas with dense distribution of hanging cables and thin table legs, and continue to maintain online inspection after reaching the predetermined inspection target point. The area is characterized by medium spatial scale, complex layout, and significant changes in the width of local passages. It has both large areas of flat walls and a large number of slender structural obstacles, including hanging cables with a diameter of about 8 mm to 15 mm, metal support rods with a diameter of about 18 mm to 25 mm, table and chair legs with a diameter of about 20 mm, and semi-transparent partitions with thin edges.

[0110] To ensure the accuracy of parameters during implementation, before formal deployment, the implementers first compiled a set of historical inspection samples from similar indoor inspection tasks. These samples included 48 complete inspection paths: 32 with normal passage, 9 with risks of grazing against slender obstacles, and 7 with instances of local collisions or sudden stops. Simultaneously, corresponding multimodal sensing data was collected, comprising approximately 31,200 frames of standardized observation data, 1.82 million valid local feature records, and statistical data on approximately 960,000 connected path segments. Based on these historical inspection samples, the implementers conducted statistical analysis on candidate node collision probability thresholds, average collision probability thresholds, maximum collision probability thresholds, node connectivity distance thresholds, and online replanning trigger thresholds. The selected thresholds were: candidate node collision probability threshold 0.18, average collision probability threshold 0.14, maximum collision probability threshold 0.31, node connectivity distance threshold 0.85 meters, and online replanning trigger threshold 0.36. This threshold combination achieved a passable path determination accuracy of 96.7% in historical inspection samples, and kept the proportion of slender obstacles misidentified as passable areas to within 3.1%.

[0111] At the start of this patrol mission, the robot's RGB image sensor, depth sensor, and point cloud sensor were activated simultaneously to perform multimodal perception of the patrol environment. The robot moved forward at a constant speed along the patrol path, acquiring 742 raw image frames, 738 raw depth frames, and 371 raw point cloud frames during the initial data acquisition phase. After a unified time base transformation, 706 reference times were formed. Each reference time corresponds to a set of standardized observation data, serving as the basic input for subsequent mapping and path construction. At the 164th reference time, a suspended cable appeared approximately 1.12 meters to the right front of the robot. The cable's width in the image was approximately 6 pixels, its continuous projection length in the depth map was approximately 0.34 meters, and it corresponded to approximately 28 sparse points in the point cloud. At the 247th reference time, two metal support rods with a spacing of approximately 0.19 meters appeared 0.76 meters to the left of the robot. At the 305th reference time, the boundary of the passage area in front of the robot and a set of thin table legs formed a narrow slit passage, the narrowest point of which was approximately 0.48 meters. The implementers focused on observing key areas as locations most prone to mapping distortion and path misjudgment using traditional methods during the inspection process.

[0112] After the multimodal sensing data enters the RTAB-Map real-time processing stage, the system constructs synchronous observation frames according to reference times. The number of observation sampling units in each synchronous observation frame is not fixed; during this inspection, each frame contains an average of approximately 13,800 observation sampling units. For the 164th reference time, the system detected that the weighted changes in grayscale and depth differences of the observation sampling units corresponding to the cable area were significantly higher than those in the background area. The obstacle edge response value reached 0.83, the consistency response value of the same spatial location in image and depth observation was 0.74, and the passage continuity response value was 0.39. The weighted average of these three values ​​resulted in an inspection sensitivity response value of 0.69. In contrast, the inspection sensitivity response value of the adjacent blank wall area was only 0.11. After filtering based on the inspection sensitivity response map, a total of 1,934 inspection sensitivity local features were retained in this frame. Additionally, since this frame was located near a weak texture area, the system extracted 186 regular local features from non-sensitive areas, bringing the total number of local features included in pose estimation to 2,120. Compared with the default feature extraction strategy of traditional RTAB-Map, which extracts only 1678 effective features at the same reference time, less than 104 of which fall directly near the edges of the cable and support rod. In contrast, the method of this invention obtains 463 effective features related to slender obstacles and passage boundaries at the same location.

[0113] During pose estimation between adjacent reference times, the system performs weighted feature matching on the subset of sensitive local features described during the inspection. Between reference times 164 and 165, there are initially 682 matched feature pairs. After spatial consistency filtering, 547 valid matched feature pairs are retained, of which high-weighted matched feature pairs related to suspended cables, support rods, and passage boundaries account for 38.6%. Based on the spatial point pairs formed by these matched feature pairs, the system calculates the relative pose transformation between adjacent frames, where the translation component is 0.184 meters and the yaw change after conversion of the rotation component is approximately 1.7 degrees. As the inspection progresses, a total of 705 sets of valid adjacent frame constraints are formed within the entire 706 reference times, and 6 sets of inspection constraint closures are triggered when returning to the vicinity of the starting channel. The traditional method has a cumulative pose drift error of about 0.78 meters in this scenario, while the method of this invention controls the cumulative drift error to within 0.17 meters after closed-loop optimization. Especially in areas with dense distribution of slender structures, the jitter amplitude of local pose estimation is reduced from 0.21 meters in the traditional method to 0.05 meters.

[0114] In the sparse 3D point cloud generation stage, the system performs a unified coordinate transformation on all spatial observation points corresponding to all reference times and performs sparse aggregation based on the contribution of structure preservation. The total number of all spatial observation points in the global coordinate system is approximately 1.82 million. The traditional voxel method retains 247,000 points at a voxel resolution of 0.05 meters, but only 12 valid points remain in the overhanging cable region and only 57 valid points remain in the thin table leg region. The method of this invention generates a sparse 3D point cloud of 314,000 points in the same scene, of which 96 valid points are retained in the overhanging cable region and 214 valid points are retained in the thin table leg region. For a table leg boundary point, its boundary support value is 0.88, the support value for slender obstacles is 0.79, the multi-frame visibility stability value is 0.84, and the structure retention contribution is 0.86. Therefore, it is directly retained during sparse aggregation. However, for a point on a flat wall surface, its boundary support value is 0.07, the support value for slender obstacles is 0.03, the multi-frame visibility stability value is 0.42, and the structure retention contribution is 0.18. This point is discarded when its distance from historically retained points is less than the sparse aggregation distance threshold. This result directly demonstrates that the present invention effectively addresses the problem of slender obstacles being easily lost during downsampling.

[0115] In the initialization stage of the 3D Gaussian volume, the system establishes a seed set of 3D Gaussian volume point by point using a sparse 3D point cloud of 314,000 points, generating a total of 314,000 3D Gaussian volume seeds. For a sparse 3D point located on a suspended cable, the system searches for 23 local neighborhood sparse 3D points within a radius of 0.06 meters. After constructing the local structure matrix, the eigenvalue decomposition results reflect obvious single principal axis extension characteristics. Subsequently, combined with the slender obstacle support value of 0.93 and the structure retention contribution of 0.91 inherited by this point, the initial values ​​of the three principal axis scale parameters are approximately 0.041 meters, 0.006 meters, and 0.005 meters, respectively, showing obvious slender anisotropic structure. For a stable wall boundary point, the initial values ​​of the corresponding three principal axis scale parameters are approximately 0.071 meters, 0.053 meters, and 0.018 meters, which are closer to the planar structure expression. By combining color observations and boundary support values ​​at multiple reference times, the system assigns initial values ​​for both color and opacity parameters to the aforementioned cable Gaussian volume. The initial opacity parameter of this cable Gaussian volume reaches 0.91, while the initial opacity parameters of Gaussian volumes in ordinary background regions are mostly concentrated between 0.42 and 0.58. This design enables critical boundaries and high-risk collision regions to have higher occupancy representation intensity in the initial 3D Gaussian sputtering model.

[0116] In the graph-volume joint incremental optimization phase, the system constructs an incremental optimization window centered on the current reference time. For example, when optimizing at the 356th reference time, 24 consecutive reference times are selected backward to form the incremental optimization window, which contains 24 pose nodes and approximately 48,000 visible 3D Gaussian bodies. The system calculates the navigation enhancement weights of the 3D Gaussian bodies based on the boundary support values ​​inherited by each 3D Gaussian body, the support values ​​of slender obstacles, the multi-frame visible stability values, and the contribution of structural preservation. The navigation enhancement weight of a cable Gaussian body reaches 2.87, the navigation enhancement weight of a table leg boundary Gaussian body is 2.41, while the navigation enhancement weight of a regular wall Gaussian body is only 1.16. In the subsequent calculation of color observation residuals and depth observation residuals, the influence of these key navigation regions on the optimization objective function is significantly amplified. After nine alternating iterations, the color observation term decreased from the initial 12.83 to 4.07, the depth observation term decreased from 8.21 to 2.34, and the pose graph constraint term decreased from 5.12 to 1.49, indicating that the overall graph-volume joint incremental optimization objective function converged. The optimized robot six-DOF pose sequence was written back, further reducing the global pose error. Simultaneously, the center position, principal axis directions, scale, and opacity of the 3D Gaussian volume were consistently updated.

[0117] After the optimized set of 3D Gaussian volume parameters was formed, the system entered the hierarchical Gaussian volume management stage. The visibility and information entropy values ​​of each of the 314,000 optimized 3D Gaussian volumes were statistically analyzed. The results showed that approximately 68,000 3D Gaussian volumes had high information entropy and high visibility, mainly corresponding to the boundaries of passable areas, the edges of slender obstacles, and key spatial contours; approximately 142,000 3D Gaussian volumes were in the medium information content range, mainly corresponding to the environmental detail layer; and the remaining approximately 104,000 3D Gaussian volumes were mostly low-contribution background areas. Based on the hierarchical scoring values, the system formed a three-layer architecture: a core structure layer, an environmental detail layer, and a redundant noise layer. Then, based on the updated opacity parameters and the contribution of structure retention, a pruning and filtering process was performed, retaining a hierarchical set of approximately 197,000 3D Gaussian volumes. Compared to directly retaining all 3D Gaussian volumes without layering, this method reduces the query time for subsequent continuous collision probability fields from an average of 2.84 milliseconds per point to 0.41 milliseconds, while maintaining the retention rate of key obstacle boundaries at over 95%.

[0118] The system constructs a continuous collision probability field based on a hierarchical set of 3D Gaussian volumes. For any spatial query point, the system calculates the occupancy contribution value of each level of 3D Gaussian volume at that query point and performs hierarchical fusion to obtain a continuous collision probability value. At a spatial query point 0.02 meters directly below the suspended cable, the continuous collision probability value reaches 0.93; at 0.18 meters beside the cable, it drops to 0.37; and in the area at the center of the passage, approximately 0.42 meters from the nearest obstacle, the continuous collision probability value is only 0.06. This continuously varying result is significantly different from the abrupt, either passable or impassable expression in traditional voxel grid methods, enabling subsequent path search and trajectory optimization to have continuous risk perception capabilities.

[0119] In the initial inspection path generation stage, the system extracts the starting position point and the target position point based on the starting pose and target pose of the inspection task, and performs spatial position discrete sampling in the inspection environment workspace, generating a total of 8400 candidate spatial position points. For each candidate spatial location point, its normalized continuous collision probability value in the continuous collision probability field is read. Points with a normalized continuous collision probability value less than or equal to 0.18 are retained, resulting in 2876 candidate points for sparse graph search nodes. Connectivity analysis is performed on any two candidate nodes, and a straight-line connecting path is constructed and sampled between the candidate nodes. Taking a group of candidate nodes with a distance of 0.64 meters as an example, the system selects 9 connectivity sampling points on the connecting line. The average collision probability value of these 9 sampling points is 0.09, and the maximum collision probability value is 0.17, which meets the passability condition, so a passable connection relationship is established. On the other hand, another group of candidate nodes that cross the cable area has an average collision probability value of 0.26 and a maximum collision probability value of 0.88, so they are judged as impassable. The resulting sparse connectivity graph contains 2878 nodes and 13642 passable connections.

[0120] After connecting the starting and target locations, the system calculates the edge cost value for each traversable connection. For a connection with a length of 0.58 meters, an average collision probability of 0.07, and a maximum collision probability of 0.13, its edge cost value is calculated as 0.58 multiplied by the distance weight coefficient, plus 0.07 multiplied by the average collision probability weight coefficient, and plus 0.13 multiplied by the maximum collision probability weight coefficient. Although a connection crossing a narrow boundary is shorter, its edge cost value is significantly higher due to an average collision probability of 0.19 and a maximum collision probability of 0.34. The system then performs a sparse connected graph search starting from the starting location. During the search, the system prioritizes expanding the candidate path with the lowest overall search cost value and calculates the path cost value after the candidate path reaches the target location. The resulting initial search path consists of 27 path nodes with a total length of approximately 18.9 meters. The traditional grid-based A* method yields a path length of 17.8 meters in the same scenario, which seems shorter. However, two segments of the path are geometrically close to the boundaries of the cable and table leg, with a minimum safe distance of only 0.11 meters. In contrast, the method of this invention generates an initial inspection path with a minimum safe distance of 0.39 meters.

[0121] In the collision-free optimization inspection path stage, the system resamples the initial inspection path at a preset arc length interval of 0.22 meters, resulting in 93 path optimization sampling points. For a path optimization sampling point near the cable boundary, its normalized continuous collision probability value is 0.41, and the magnitude of the continuous collision probability gradient vector formed by taking partial derivatives along the three coordinate directions is 1.28, pointing away from the cable boundary. The system performs collision probability barrier mapping on all intermediate path optimization sampling points, transforming high collision probability regions into high cost-value regions. Then, based on the positional relationship between adjacent path optimization sampling points, a path smoothing cost term and a path length cost term are constructed, which are weighted and summed with the collision cost term to form the gradient-driven trajectory optimization objective function. After 31 iterations, the collision cost term decreased from the initial 28.4 to 6.7, the path smoothing cost term decreased from 13.2 to 4.9, and the minimum safe distance of the collision-free optimized inspection path increased to 0.46 meters.

[0122] In the smooth inspection path generation stage, the system reads the collision-free optimized inspection path, reparameterizes the path according to arc length, and obtains 96 path dynamics sampling points. With a reference operating speed of 0.55 m / s, the sampling time interval between adjacent path dynamics sampling points is automatically allocated according to the distance between adjacent sampling points. Subsequently, the system calculates the discrete velocity vector, discrete acceleration vector, acceleration rate of change vector, and discrete curvature index point by point. Before optimization, the maximum discrete acceleration vector magnitude reaches 1.12 m / s², exceeding the preset velocity rate of change constraint threshold of 0.80 m / s²; the maximum acceleration rate of change vector magnitude reaches 1.97 m / s³, exceeding the acceleration rate of change constraint threshold of 1.40 m / s³. The system constructs a smooth inspection path generation objective function using curvature continuity cost, velocity rate of change cost, acceleration rate of change cost, and path preservation cost, and iteratively updates the intermediate path dynamics sampling points while keeping the starting and target positions unchanged. After 16 optimizations, the maximum discrete acceleration vector magnitude decreased to 0.73 m / s², the maximum acceleration rate of change vector magnitude decreased to 1.08 m / s³, and the peak curvature change decreased by about 48%. A smooth inspection path was generated, and a path tracking reference sequence containing path dynamics sampling points, discrete velocity vectors, and discrete acceleration vectors was generated and sent to the robot control system.

[0123] During the robot's smooth patrol path execution, the system continuously reads new multimodal perception data and performs local incremental updates. When the robot reaches the middle of the original path, a temporarily placed box-shaped obstacle appears approximately 2.6 meters ahead; this obstacle was not present in the previous map. The system constructs a local update region within a 3.2-meter radius, centered on the robot's current pose. Within this region, it recalculates the collision occupancy observations corresponding to newly added observations for spatial query points and merges them with the original continuous collision probability field. After the update, the normalized continuous collision probability values ​​for several spatial query points within the affected region increase from the original 0.04-0.09 to 0.67-0.91. The system rereads the updated normalized continuous collision probability values ​​for all path dynamics sampling points along the smooth patrol path and selects the maximum value as the path safety indicator. The path safety index reached 0.72, exceeding the online replanning trigger threshold of 0.36. Therefore, the system immediately used the position corresponding to the robot's current pose as the new starting position and the original target position as the new target position, and re-executed the path generation and optimization process. The entire online replanning process took 0.48 seconds. The updated smooth patrol path successfully avoided the newly added obstacle, and the robot did not stop abruptly or experience any edge collisions.

[0124] To verify the effectiveness of the method of this invention, the implementers used the same inspection environment, the same robot platform, and the same initial task conditions, and performed 20 inspection tasks each using the method of this invention and the traditional RTAB-Map + voxel downsampling + grid A* + spline smoothing method. Experimental results showed that the traditional method resulted in 6 instances of missed detection of slender obstacles leading to path edge-grazing, 3 instances of triggering emergency stops, and 2 instances of replanning failure after the appearance of new obstacles. The number of complete successful executions was 12, with a success rate of 60.0%. In contrast, the method of this invention only triggered low-speed avoidance once when a new obstacle was extremely close to the channel boundary, but it achieved complete successful execution 19 out of 20 tasks, with a success rate of 95.0%. Regarding the retention rate of slender obstacles, the traditional method averages 43.8%, while the method of this invention averages 92.4%. Regarding the cumulative pose drift error, the traditional method averages 0.74 meters, while the method of this invention averages 0.16 meters. Regarding the single-point query time for the continuous collision probability field, the traditional method averages 2.63 milliseconds, while the method of this invention averages 0.44 milliseconds. Regarding the replanning response time, the traditional method averages 1.72 seconds, while the method of this invention averages 0.48 seconds. Regarding the minimum safe path spacing, the traditional method averages 0.13 meters, while the method of this invention averages 0.44 meters. Regarding the peak path curvature, the traditional method averages 3.21 per meter, while the method of this invention averages 1.74 per meter.

[0125] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An indoor patrol path construction system based on robot vision recognition, characterized in that, include: The multimodal acquisition module collects multimodal sensing data of the indoor environment and performs time synchronization and calibration processing on the multimodal sensing data to obtain synchronized and calibrated multimodal sensing data. An improved RTAB module is used to process synchronously calibrated multimodal perception data in real time based on an improved RTAB-Map model, thereby obtaining the robot's six-degree-of-freedom pose sequence and sparse 3D point cloud. The Gaussian model initialization module initializes the three-dimensional Gaussian volume parameter set and constructs the initial three-dimensional Gaussian sputtering model using the robot's six-degree-of-freedom pose sequence and the sparse three-dimensional point cloud as priors. The graph-volume joint optimization module performs graph-volume joint incremental optimization on the initial 3D Gaussian sputtering model and the robot's six-DOF pose sequence to obtain the optimized robot's six-DOF pose sequence and the optimized 3D Gaussian volume parameter set. The hierarchical Gaussian management module performs hierarchical Gaussian body management based on the optimized three-dimensional Gaussian body parameter set, generates a hierarchical three-dimensional Gaussian body set, and constructs a continuous collision probability field in three-dimensional space based on the hierarchical three-dimensional Gaussian body set to obtain the continuous collision probability field. The path planning and optimization module performs connectivity analysis using the continuous collision probability field and performs sparse graph search to generate an initial inspection path based on the starting pose and target pose of the inspection task. It then performs gradient-driven trajectory optimization on the initial inspection path to obtain a collision-free optimized inspection path. The online replanning module integrates the collision-free optimized inspection path with robot dynamics constraints to generate a smooth inspection path and send it to the robot control system. During robot execution, the continuous collision probability field is updated, and the smooth inspection path is replanned online to complete real-time autonomous inspection in indoor environments.

2. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The multimodal acquisition module includes: Collect multimodal sensing data of the indoor environment and construct the original multimodal sensing dataset; Perform a unified time base transformation on the original timestamps corresponding to various sensors in the original multimodal sensing dataset to obtain standard timestamps; Using a reference time sequence under a unified time base as the alignment benchmark, cross-sensor time matching is performed on standard timestamps to determine the alignment sampling index of various sensors at each reference time. The observation data of various sensors obtained through time matching are calibrated, and the intrinsic parameter matrix and extrinsic parameter transformation matrix of each type of sensor are obtained respectively. Based on the aligned sampling index, intrinsic parameter matrix and extrinsic parameter transformation matrix, a unified spatiotemporal coordinate representation is performed on the multimodal observation content corresponding to each reference time, generating a synchronously calibrated multimodal sensing dataset.

3. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The improved RTAB module includes: The synchronously calibrated multimodal perception dataset is grouped according to the reference time sequence to construct a synchronous observation frame set. For each synchronous observation frame, slender obstacle edges, passage area boundaries, and stable structure regions are constructed in relation to the robot's indoor inspection path. Inspection sensitivity response values ​​are calculated to obtain an inspection sensitivity response map. Based on the inspection-sensitive response map, inspection-sensitive features are filtered for each synchronous observation frame to generate a subset of inspection-sensitive local feature descriptions, an inspection-sensitive appearance description vector, and an inspection-passage topology description vector. Weighted feature matching is performed on the sensitive local feature descriptor subsets of adjacent reference times, and the relative pose transformation matrix between adjacent reference times is obtained by combining the spatial observation content in the synchronously calibrated multimodal sensing dataset. Pose accumulation is performed based on the relative pose transformation matrix between adjacent reference times to obtain the initial six-degree-of-freedom pose sequence of the robot. Perform closed-loop detection based on the inspection-sensitive appearance description vector and the inspection-traffic topology description vector to determine the inspection constraint closed-loop set; Based on the initial robot six-DOF pose sequence, the relative pose transformation matrix between adjacent reference times, and the set of roving constraint closed loops, a roving constraint pose graph is constructed, and global optimization is performed on the roving constraint pose graph to obtain the optimized robot six-DOF pose sequence. Based on the optimized robot six-degree-of-freedom pose sequence, a unified coordinate transformation is performed on the spatial observation points corresponding to each reference time. The boundary support value, slender obstacle support value, and multi-frame visible stability value of each spatial observation point in the construction of the robot's indoor inspection path are calculated and weighted and fused into a structure retention contribution. Based on the structure retention contribution, sparse aggregation is performed on the spatial observation points after the unified coordinate transformation to obtain a sparse three-dimensional point cloud.

4. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The Gaussian model initialization module includes: Read the robot's six-degree-of-freedom pose sequence and sparse three-dimensional point cloud, establish the initial correspondence of the three-dimensional Gaussian body point by point according to the sparse three-dimensional points in the sparse three-dimensional point cloud, and generate a seed set of three-dimensional Gaussian bodies. Using the initial value of the center position of the 3D Gaussian solid seed as the center, search for a local neighborhood sparse 3D point set in the sparse 3D point cloud, construct a local structure matrix, and initialize the initial values ​​of the principal axis direction parameters of the corresponding 3D Gaussian solid based on the eigenvalue decomposition results of the local structure matrix. Read the support value of the slender obstacle and the contribution of structure preservation in the seed of the three-dimensional Gaussian body, and combine the eigenvalues ​​obtained by the seed of the three-dimensional Gaussian body in the process of local structure matrix eigenvalue decomposition to initialize the initial values ​​of the scale parameters of the corresponding three-dimensional Gaussian body along the three principal axes. Based on the initial values ​​of the scale parameters and principal axis parameters of the corresponding 3D Gaussian body seed along the three principal axes, the initial values ​​of the covariance matrix of the corresponding 3D Gaussian body are constructed. Based on the robot's six-degree-of-freedom pose sequence and a synchronously calibrated multimodal perception dataset, the visible observation set of the three-dimensional Gaussian body corresponding to the seed of the three-dimensional Gaussian body at each reference time is calculated, and the initial value of the color parameter of the three-dimensional Gaussian body corresponding to the seed of the three-dimensional Gaussian body is initialized according to the visible observation set. Based on the boundary support value and multi-frame visible stability value in the 3D Gaussian seed, initialize the initial value of the opacity parameter of the 3D Gaussian corresponding to the 3D Gaussian seed. The initial values ​​of the center position, covariance matrix, principal axis direction parameter, color parameter, and opacity parameter of each 3D Gaussian volume seed are collected to generate the initial value set of 3D Gaussian volume parameters. Based on the initial value set of 3D Gaussian volume parameters, the robot's six-degree-of-freedom pose sequence, and the sparse 3D point cloud, an initial 3D Gaussian sputtering model is constructed.

5. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The graph-volume joint optimization module includes: Based on the initial 3D Gaussian sputtering model, an incremental optimization window is constructed according to the current reference time, and the set of pose nodes and the set of visible 3D Gaussian volumes within the incremental optimization window are determined. Based on the boundary support value, slender obstacle support value, multi-frame visible stability value, and structure preservation contribution inherited by each 3D Gaussian volume in the visible 3D Gaussian volume set, the navigation enhancement weight of the 3D Gaussian volume is calculated, and the color observation residual and depth observation residual are constructed based on the standard observation content corresponding to each reference time within the incremental optimization window. Based on color observation residuals, depth observation residuals, patrol constraint odometer constraint edges, and patrol constraint closed-loop constraint edges, a graph-volume joint incremental optimization objective function is constructed. The graph-volume joint incremental optimization objective function is solved by alternating iterations, updating the pose node parameters and the 3D Gaussian volume parameters. If the change in the graph-volume joint incremental optimization objective function between two consecutive iterations is not greater than the preset convergence threshold, the iteration is stopped. After the incremental optimization window has converged, the updated pose node parameters are written back to the robot's six-DOF pose sequence, and the updated 3D Gaussian volume parameters are written back to the 3D Gaussian volume parameter set, thus obtaining the optimized robot six-DOF pose sequence and the optimized 3D Gaussian volume parameter set.

6. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The hierarchical Gaussian management module includes: Based on the optimized set of three-dimensional Gaussian volume parameters, the visibility frequency and visibility distribution of each three-dimensional Gaussian volume during the observation process at multiple reference times are statistically analyzed, and the information entropy value and visibility value of the corresponding three-dimensional Gaussian volume are calculated by combining the opacity parameter and covariance matrix of the three-dimensional Gaussian volume. Based on the information entropy and visibility values ​​of each 3D Gaussian body, the hierarchical score is calculated, and the 3D Gaussian bodies in the optimized 3D Gaussian body parameter set are hierarchically stored according to the hierarchical score to generate multi-level Gaussian body subsets. Based on the hierarchical score value, updated opacity parameter and structural retention contribution of the multi-level Gaussian body subset, the 3D Gaussian bodies in each level are clipped and filtered to retain the 3D Gaussian bodies corresponding to the navigation key boundaries and clip the low-contribution 3D Gaussian bodies to generate a hierarchical 3D Gaussian body set. A continuous collision probability field is constructed in three-dimensional space based on a hierarchical set of three-dimensional Gaussian bodies. The occupancy contribution value of each level of three-dimensional Gaussian body is calculated for any spatial query point. The occupancy contribution values ​​of each level of three-dimensional Gaussian body are then fused hierarchically to obtain the continuous collision probability value of the spatial query point. Perform normalization mapping on the continuous collision probability values ​​of spatial query points and encapsulate them into a continuous collision probability field.

7. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The path planning optimization module includes: Based on the continuous collision probability field, the corresponding starting position point and target position point are extracted from the starting pose and target pose of the inspection mission, and a sparse graph search node candidate set is established in the workspace of the inspection environment. Based on the continuous collision probability field, perform connectivity analysis on the candidate set of search nodes in the sparse graph to determine the walkable connection relationship between candidate nodes and generate an inspected sparse connected graph. Connect the starting point and the target point to the sparse connected graph of the inspection, and calculate the edge value corresponding to each passable connection in the sparse connected graph of the inspection. Based on the sparse connected graph of the inspection, the starting point, the target point, and the edge values ​​corresponding to each traversable connection, the inspection sparse connected graph search is performed to obtain the initial inspection path formed by the sequential connection of multiple path nodes. Read the initial inspection path and continuous collision probability fields, resample the initial inspection path according to the preset arc length interval, and generate an initial optimized path sampling sequence formed by connecting multiple path optimization sampling points in sequence. Based on the differentiability of the continuous collision probability field, the continuous collision probability value and continuous collision probability gradient at each path optimization sampling point are calculated, and the collision cost term of the path optimization sampling sequence is constructed. Based on the positional relationship between adjacent path optimization sampling points in the initial optimized path sampling sequence, a path smoothing cost term and a path length cost term are constructed, and then weighted and summed with the collision cost term to obtain the gradient-driven trajectory optimization objective function. The gradient-driven trajectory optimization objective function is iteratively solved to obtain the updated path optimization sampling sequence. Perform collision-free constraint determination and convergence determination on the updated path optimization sampling sequence, and output the updated path optimization sampling sequence that satisfies the collision-free constraint and convergence condition as the collision-free optimized inspection path.

8. The indoor patrol path construction system based on robot vision recognition according to claim 1, characterized in that, The online replanning module includes: The starting point, path node sequence, and target point in the collision-free optimized inspection path are sequentially expanded, and the path node sequence is discretely sampled according to a preset arc length interval to obtain the path dynamics sampling sequence. Based on the positional relationship between adjacent path dynamics sampling points in the path dynamics sampling sequence, calculate the discrete acceleration vector, the acceleration rate vector, and the discrete curvature index; A smooth inspection path is constructed based on the discrete curvature index, discrete acceleration vector, and acceleration rate of change vector to generate the objective function; The objective function for generating a smooth inspection path is constrained and optimized to generate a smooth inspection path. The path tracking reference sequence corresponding to the smooth inspection path is then sent to the robot control system. During robot execution, new multimodal perception data is read, and the continuous collision probability field is locally incrementally updated in combination with the robot's current execution pose to obtain the updated continuous collision probability field. Based on the updated continuous collision probability field, online safety checks and online replanning are performed on the smooth patrol path to complete real-time autonomous patrol in indoor environments.

9. The indoor patrol path construction system based on robot vision recognition according to claim 8, characterized in that, The online safety verification and online replanning process includes: when there are path dynamic sampling points in the smooth patrol path that do not meet the safety constraints, the position point corresponding to the robot's current execution pose is used as the new starting position point, and the position point corresponding to the target pose of the patrol task is used as the target position point, and the updated smooth patrol path is regenerated.

Images