Inspection robot intelligent navigation method and device

Abstract

The embodiment of the application provides a kind of intelligent navigation method and device of inspection robot, through innovatively designing map fusion system, through multi-source data and raster processing, the effective expression of environment is realized.Environmental perception mechanism is constructed, combined with semantic analysis and trajectory prediction, reliable obstacle identification model is established.Dynamical programming is introduced, through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured.The method effectively solves the deficiencies of traditional technology in map fusion, environmental perception and dynamic programming, etc., and provides technical support for inspection robot.

Description

Technical Field

[0001] This application relates to the field of path planning, specifically to an intelligent navigation method and device for inspection robots. Background Technology

[0002] Existing navigation methods for inspection robots have significant shortcomings. Traditional systems perform poorly in map fusion and path planning, failing to effectively integrate multi-source information and affecting navigation accuracy.

[0003] Furthermore, existing technologies suffer from bottlenecks in environmental perception and obstacle prediction. Most systems lack robust semantic analysis mechanisms and trajectory prediction strategies, resulting in less than ideal obstacle avoidance performance.

[0004] Existing systems have technical limitations in dynamic programming. The lack of in-depth analysis of interference regions makes it difficult to achieve flexible obstacle avoidance control through state-space optimization, thus impacting navigation efficiency. Solving these problems is crucial for improving the performance of inspection robots. Summary of the Invention

[0005] To address the problems in the existing technology, this application provides an intelligent navigation method and device for inspection robots, which can effectively solve the shortcomings of traditional technologies in map fusion, environmental perception and dynamic planning, and provide technical support for inspection robots.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] Firstly, this application provides an intelligent navigation method for an inspection robot, comprising:

[0008] The historical map information of the inspection robot is input into the map fusion module. Combined with the laser point cloud map and visual map data, the confidence weight of the point cloud map and the visual map is calculated to generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0009] Environmental perception data is collected by a depth camera and a lidar. The environmental perception data is then segmented into targets and classified into scene semantics. The attribute features of the passable area are extracted, a motion prediction model is constructed, the spatial position and volume parameters of dynamic obstacles are calculated, the movement trajectory of the dynamic obstacles is predicted, the movement trajectory is input into a collision detection unit, the degree of interference between the dynamic obstacles and the preset inspection path is determined, and an interference area marker is generated.

[0010] An obstacle avoidance planner is constructed based on the movement trajectory and spatial location. The interference area is mapped to the gridded map, and temporarily impassable units are marked. The obstacle avoidance parameters are adjusted according to the attribute characteristics of the passable area. The optimal obstacle avoidance path is calculated using state space iteration. Temporary obstacle avoidance instructions are generated, and the inspection robot is controlled to execute the temporary obstacle avoidance instructions. After completing the obstacle avoidance, it returns to the preset inspection path.

[0011] Furthermore, it also includes: reading the historical map information of the inspection robot, constructing a multi-source data fusion module, converting the laser point cloud map data into a standard coordinate system, performing noise reduction and spatial clustering processing on the point cloud data, extracting and matching feature points from the visual map data, constructing a map record table containing data source, collection time, and spatial location, and writing the map record table into the fusion database;

[0012] The system reads the map record table from the fusion database, constructs a weight calculation module, performs registration processing on the point cloud map and the visual map, calculates the matching degree coefficient of the overlapping area, generates a confidence weight vector based on the matching degree coefficient, combines the weight vector with the map data in a weighted manner to generate fused map data, and writes the fused map data into the map database.

[0013] Furthermore, it also includes: reading fused map data from the map database, constructing a grid division module, rasterizing the fused map data according to a fixed resolution, marking the grid occupied by static obstacles, generating a grid distribution matrix of passable and impassable units, constructing a grid record table containing grid number, occupancy status, and location information, and writing the grid record table into the planning database.

[0014] The system reads the grid record table from the planning database, constructs a path planning module, calculates the connectivity parameters between passable units, generates a connectivity weight matrix, optimizes the path search based on the weight matrix, generates a preset inspection path, smooths the preset inspection path, and writes the preset inspection path into the navigation control system.

[0015] Furthermore, it also includes: reading the sensor data streams of the depth camera and the lidar, constructing a data acquisition module, synchronizing the depth image data and the laser scanning data in time, generating an environmental perception dataset, performing target segmentation processing on the environmental perception dataset, constructing a perception record table containing target contours, depth information, and scanning intensity, and writing the perception record table into the acquisition database;

[0016] The system reads the perception record table from the collection database, constructs a scene analysis module, performs semantic classification on the environmental perception data, extracts the width parameters and road surface features of the passage area, generates scene attribute vectors, establishes a motion prediction model based on the scene attribute vectors, and writes the motion prediction model into the prediction database.

[0017] Furthermore, it also includes: reading motion prediction models from the prediction database, constructing an obstacle analysis module, calculating the center coordinates and volume boundaries of dynamic obstacles, generating spatial position parameters, performing time-series tracking of the dynamic obstacles, calculating the obstacle movement trajectory based on the motion prediction model, constructing an obstacle record table containing target number, position parameters, and trajectory data, and writing the obstacle record table into the obstacle database;

[0018] The obstacle record table in the obstacle database is read, a collision detection module is constructed, the movement trajectory is mapped to a preset inspection path, the spatial interference degree of the trajectory overlap area is calculated, an interference degree score vector is generated, the location of the interference area is marked based on the score vector, an interference record table containing area number, interference degree, and spatial range is constructed, and the interference record table is written into the interference database.

[0019] Furthermore, it also includes: reading the interference record table in the interference database, constructing an obstacle avoidance planner, inputting the obstacle movement trajectory and spatial location parameters into the planning model, calculating the obstacle avoidance priority coefficient, generating a planning weight matrix, constructing a planning record table containing planning number, weight parameters, and priority level, and writing the planning record table into the planning database;

[0020] Read the planning record table in the planning database, construct a region marking module, convert the interference region identifier into raster coordinates, update the occupancy status in the rasterized map, mark temporarily impassable units, construct a marking record table containing raster number, occupancy status, and time limit identifier, and write the marking record table into the marking database.

[0021] Furthermore, it also includes: reading the marker record table in the marker database, constructing a parameter adjustment module, calculating obstacle avoidance constraints based on the attribute characteristics of the passage area, generating an obstacle avoidance parameter vector, calculating the obstacle avoidance path cost using a state space iterative method, constructing a path record table containing path nodes, turning angles, and speed components, and writing the path record table into the path database;

[0022] The path record table in the path database is read, an instruction generation module is constructed, the optimal obstacle avoidance path is converted into motion control instructions, a temporary obstacle avoidance instruction sequence is generated, the inspection robot is controlled to perform obstacle avoidance actions, the obstacle avoidance execution status is monitored, the temporary path data is cleared after obstacle avoidance is completed, the inspection robot is guided to the preset inspection path, an execution record table containing instruction number, execution status, and return position is constructed, and the execution record table is written into the execution database.

[0023] Secondly, this application provides an intelligent navigation device for an inspection robot, comprising:

[0024] The path planning module is used to input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and the visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0025] The obstacle avoidance determination module is used to collect environmental perception data through a depth camera and a lidar, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate an interference area marker.

[0026] The obstacle avoidance execution module is used to construct an obstacle avoidance planner based on the movement trajectory and spatial location, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iterative calculation to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent navigation method for the inspection robot.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent navigation method for the inspection robot.

[0029] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the intelligent navigation method for the inspection robot.

[0030] As can be seen from the above technical solution, this application provides an intelligent navigation method and device for inspection robots. Through an innovative design of a map fusion system, and by utilizing multi-source data and rasterization processing, it achieves effective representation of the environment. An environmental perception mechanism is constructed, and a reliable obstacle recognition model is established by combining semantic analysis and trajectory prediction. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively solves the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the intelligent navigation method for the inspection robot in the embodiments of this application;

[0033] Figure 2 This is a structural diagram of the intelligent navigation device for the inspection robot in the embodiments of this application;

[0034] Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0035] Figure label:

[0036] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0039] To address the problems existing in current technologies, this application provides an intelligent navigation method and device for inspection robots. Through an innovative map fusion system design, and by utilizing multi-source data and rasterization processing, it achieves effective representation of the environment. An environmental perception mechanism is constructed, combining semantic analysis and trajectory prediction to establish a reliable obstacle recognition model. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively solves the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots.

[0040] To effectively address the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic planning, and to provide technical support for inspection robots, this application provides an embodiment of an intelligent navigation method for inspection robots. See [link to embodiment]. Figure 1 The intelligent navigation method for the inspection robot specifically includes the following:

[0041] Step S101: Input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0042] This embodiment focuses on S101 for two types of inspection environments: park roads and computer room corridors. Inputs include historical map information (including timestamps, poses, and odometry drift statistics) generated from previous robot operations, raw laser point cloud maps, and visual map data (keyframe images, sparse / semi-dense features). First, coordinate unification and quality cleaning are performed in the map fusion module: based on the reference coordinate system of historical trajectories, extrinsic parameters are used to map the current batch of point clouds and images to unified coordinates.

[0043] To avoid old data misleading new paths, confidence labels and drift variances from historical maps are read, and slices with timeouts, high drift, or excessive duplicate coverage are downweighted or removed. Voxel filtering and ground segmentation are performed on the point cloud side to preserve stable walls, curbs, and cabinet facades; robust features and edge segments are extracted on the visual side to construct a cross-source matching feature set, providing a common-view basis for subsequent weight estimation.

[0044] In this embodiment, a layered strategy is adopted in the registration stage. First, the historical map is used as a reference to perform coarse registration on the new point cloud-image pair. The coarse registration is based on the weighted average of the initial ICP value and the visual PNP solution. After rough alignment, fine registration is performed: geometric consistency (point-to-surface residual, line-to-line parallelism) and photometric consistency (local NCC, gradient direction angle) are calculated in the overlapping area to obtain the matching degree coefficient.

[0045] Considering that tree shade and glass surfaces in the park may introduce spurious features, the fusion module incorporates reflectivity and texture repetition as rejection factors, reducing the weight of areas with high reflectivity and low geometric support. The final generated confidence weight vector w, along with the confidence scores w_p (point cloud source) and w_v (visual source), are derived from the joint estimation of the statistics of the overlapping area and historical drift records, conforming to the natural law that "the more consistent the geometry, the smaller the historical drift, and the more robust the current observation, the higher the weight."

[0046] In this embodiment, the map synthesis after weight calculation follows the geometry priority rule: the point cloud geometry is used as the main skeleton, and visual data provides boundary refinement and hole filling.

[0047] For corridor ends or road corners, where point clouds are sparse but visual edges are clear, the system fills in the grid boundaries by adjusting w_v in low-depth uncertain regions. For metal cabinet surfaces, where point clouds are dense and normals are stable, w_p dominates to avoid visual reflection distortion. The fused output is stored in the database as occupancy information: the space is divided into voxels of fixed resolution, and the voxel occupancy probability is accumulated from observational evidence weighted by w_p and w_v, then projected onto the ground to form a two-dimensional occupancy grid. Historical obstacles (long-term walls, bases) are fixed in the base layer, while short-term debris is recorded in a revocable layer for easy updates.

[0048] In this embodiment, during the rasterization stage, the occupancy status, static / dynamic labels, surface roughness, and slope of each grid are recorded. For computer room corridors, the minimum passage width and doorway location are considered; for campus roads, the shoulder boundaries and speed bumps are considered. Connectivity assessment is based on eight-neighborhood or sparse graph extraction, and edge weights are determined by a combination of free space width, distance to obstacles, road surface smoothness, and historical traversal frequency. If the historical map shows that a certain passage is frequently blocked by temporary debris, the traversal cost for that segment is increased in the weight matrix. For reflective areas, although they may appear geometrically open, the cost is increased due to perceptual instability, reducing the path's dependence on perceptually vulnerable areas.

[0049] The path generation in this embodiment adopts a two-stage strategy.

[0050] In the first stage, a heuristic search is performed on the global grid map using the weight matrix as the cost graph to obtain the preset inspection path skeleton. The heuristic function takes into account the Euclidean distance to the target and semantic penalty.

[0051] The second phase involves continuous spatial smoothing of the skeleton, controlling the curvature and obstacle distance constraints to ensure the mobile chassis can still move around corridor corners and sharp bends in the park. To prevent the smoothing from eroding necessary safety clearances, hard constraint boundaries are introduced, setting minimum distances from the edges for path segments near hazardous edges (staircases, steps). Functional waypoints, such as shooting pose and sensor calibration docking points, are also inserted along the path, allowing subsequent task modules to be directly integrated.

[0052] In this embodiment, before finalizing the path, a coverage check is performed on the entire path based on the confidence layer of the fused map. If a segment relies on low-confidence data, the system traces back to the fusion module to query the data source, determining whether there is a risk of single-source dominance or insufficient overlap. If necessary, it requests encrypted raster (to increase resolution or supplement local data) for further evaluation. For loop inspection tasks, the preset path bias selects a closed track that is bidirectionally accessible and has a sufficient turning radius to avoid turning around at narrow ends. For cross-area inspections in parks with access control or elevators, the path cost is additionally incurred by the time cost of passage slots and waiting for doors to open.

[0053] This embodiment provides a formula to illustrate the composition of connectivity weights:

[0054] .

[0055] Where e represents an edge in the graph; D(e) represents the reciprocal of the distance from the edge's centerline to the nearest obstacle, reflecting the safety margin; R(e) represents the road surface roughness or slope index, reflecting driving smoothness; H(e) is the historical passage penalty, representing the probability of congestion / blockage of this edge in historical records; a, b, and c are non-negative coefficients, derived from w_p, w_v, and scene semantics. The smaller C(e) is, the higher the connectivity weight, and the more likely the path search will select this edge. Each parameter has a clear physical meaning and does not involve uninterpretable black-box extrapolations.

[0056] In this embodiment, the generated preset inspection path is written into the navigation control system in a format with timing and action tags, including path node coordinates, desired speed, minimum obstacle clearance distance, arrival conditions, and necessary perception trigger markers. After being written, the navigation side uses this to construct initial values ​​for the tracker and local replanning. Subsequently, if dynamic obstacles are encountered during the execution phase, the system will use this preset path as a reference for the "return to the main line".

[0057] The core problem addressed in this embodiment is how to provide a reliable, interpretable, and executable global path under conditions of multi-source incompleteness and historical drift. Through confidence-weighted fusion, semantically aware weighted connectivity modeling, and two-stage path shaping, the preset inspection path maintains reasonable safety margins and traceability at both ends of narrow corridors and open roads, reducing interference and backtracking in subsequent local obstacle avoidance.

[0058] Step S102: Collect environmental perception data through a depth camera and LiDAR, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate interference area markers.

[0059] This embodiment implements S102 for two inspection scenarios: park roads and computer room corridors. The sensing configuration is a surround-view depth camera and a 16-line LiDAR.

[0060] First, strict temporal and spatial alignment is established at the data level. In this embodiment, depth image frames and laser scanning frames are synchronized to the same reference time through hardware timestamps and extrinsic parameter calibration matrices. Short window interpolation is introduced for conditions such as turning and acceleration to correct the pose drift of the laser beam during the scanning cycle and avoid sampling the boundaries of the same object as spatial stretching. After denoising, the raw data enters the target segmentation and semantic classification process. The segmentation model adopts a geometric and texture fusion approach: continuous clustering is performed on the depth map, and candidate regions are initially delineated based on normal changes and depth gradients; then, the category confidence output by the visual semantic network is used for refinement, generating multiple types of masks such as pedestrians, vehicles, equipment carts, walls, cable trays, curbs, and vegetation. The input of the model is an RGB-D image and a laser intensity projection, and the output is a pixel-level category image and instance ID. This set of inputs is chosen because geometric boundaries and reflection intensity can better reflect the contours of real objects and conform to the physical laws of occlusion and materials.

[0061] Specifically: Local planes are estimated on ground-like masks to calculate slope, roughness, and passability width; in corridor scenes, the parallelism between wall lines and cabinet arrays is utilized to perform RANSAC fitting on the centerline of the passable corridor, obtaining the corridor width and eccentricity; on park roads, curbs and markings are merged to form left and right boundary zones, and risk labels for slippery areas and gravel-covered areas are assessed. Attribute features are organized into scene attribute vectors, with fields including passability bandwidth W, local curvature K, surface friction indicator M, and occlusion probability O. These quantities are derived from geometrically observable or stable appearance statistics and will be subsequently weighted for motion prediction and collision assessment.

[0062] This embodiment calculates the spatial location and volume parameters of dynamic obstacles, establishing a 3D bounding volume for each instance ID. Depth point clouds and radar points are fused in the same coordinate system, and a voxel grid is used for robust fitting of the point set, outputting the center coordinates c, orientation θ, and volume boundaries b (length, width, and height). For pedestrians, cylinders or vertical ellipsoids are used; for vehicles or carts, directed bounding boxes are used. The semantically dependent geometric prototypes are introduced because different objects have different physical forms and motion constraints. Cylinders better fit the lateral swaying trajectory of a person, while directed boxes can represent the frontal direction of a vehicle. To reduce single-frame noise, this embodiment calculates robust statistics of the centroid and orientation within a short time window and records occlusion confidence to avoid false collision reports caused by volume inflation due to partial occlusion.

[0063] This embodiment constructs a motion prediction model to provide interpretable trajectory predictions.

[0064] Considering that the main movements in the inspection environment are mostly low-speed, short-range walking or driving constrained by scene geometry, the model adopts a hybrid approach of "semantic-guided constant velocity + social force": the base layer uses constant velocity / constant acceleration Kalman filtering to track position and velocity; the constraint layer introduces passable boundaries such as the centerline of the passageway and walls as soft constraints to limit the trajectory from crossing impassable areas; the interaction layer uses social force terms to describe the avoidance tendency between pedestrians and robots or other people. The input consists of the previous state vector x (position, velocity, orientation) and the scene attribute vector, as well as the state set of neighboring dynamic objects; the output is the trajectory spline and its confidence interval for the next few seconds. The relationships between parameters follow common sense: when the channel width W decreases or the obstacle density increases, the variance of the lateral velocity prediction decreases, and the upper bound of the velocity magnitude tightens; when the target orientation is consistent with the channel direction, the weight of the forward velocity component increases, and the model will not generate paths that cross walls without reason. Uncertainty is retained in the prediction process so that collision detection considers the worst case without being overly conservative.

[0065] In this embodiment, interference assessment based on the preset inspection path involves projecting the predicted trajectory onto the passageway of the robot's reference trajectory and calculating the spatial and temporal overlap. To avoid misjudgment based on simple distance thresholds, the volume boundary b of the dynamic body is extended outward by a safe time interval on the time axis, forming a spatiotemporal occupancy segment. The preset inspection path is parameterized by arc length, matching the robot's external boundary and velocity curve. The interference degree calculated in this embodiment integrates the minimum time interval, overlap length, and angle difference, and introduces passage attribute weights M and K, increasing the cost of deceleration and detour when the ground is rough or slippery. To facilitate engineering closure, the interference area identifier not only includes the spatial range and score but also the trigger frame, target ID, and prediction confidence, allowing the subsequent obstacle avoidance planner to process obstacles according to priority.

[0066] In situations where multi-source sensor noise and short-term occlusion are unavoidable, single-frame judgments are prone to flickering. This embodiment incorporates a consistency buffer before collision detection: if the predicted trajectory of a target has a consistent direction and stable velocity range within a continuous window, the reliability of its interference score is weighted and increased; if the target is only detected in a few frames and the occlusion probability O is too high, the generation of a hard interference region is delayed, replaced by a soft marker, and a review is required in the next frame to avoid frequently triggering meaningless obstacle avoidance. For pedestrian groups, cluster-level bounding volumes and average velocities are used to evaluate group interference, reducing score jitter caused by members entering and exiting.

[0067] This embodiment provides an expression for the interference score:

[0068] ,

[0069] Where S is the interference score, D is the spatiotemporal distance between the inspection path and the dynamic body trajectory at the closest moment (the smaller the distance, the more dangerous), L is the arc length or duration of spatiotemporal overlap, A is a function of the angle between their travel directions (more dangerous when traveling head-on), R is the passage risk index (synthesized from ground roughness, slipperiness, and occlusion probability), and w1~w4 are weights determined by the scene attribute vector and target semantics. Each parameter has a clear physical meaning and is consistent with common sense regarding actual walking and vehicle movement, ensuring that the score does not deviate from natural laws.

[0070] This embodiment provides an example at a fork in a park road: a pedestrian walks out from the green belt on the left. The segmentation module first provides an incomplete outline at the vegetation boundary, and then completes it within two frames. Motion prediction extends the trajectory towards the sidewalk under the constraint of the channel boundary. The interference score increases rapidly due to the large angle term A, and the system output covers a small section of the interference area in front of the fork. In the example of the computer room corridor, the cart travels in the same direction along the corridor. The model judges it as a low-angle following interference, and the score comes more from the overlap length L, giving a narrow and long interference band. The interference area identifiers in both cases are written into the downstream database, with fields including area number, spatial range, score, and confidence, for direct consumption by the obstacle avoidance planning stage. Thus, S102 connects the three things "what is seen", "where will it go", and "will it collide" into a continuous link, relying on measurable geometric and photometric information, and ensuring that the conclusion conforms to common sense on site through semantic and physical constraints.

[0071] Step S103: Construct an obstacle avoidance planner based on the movement trajectory and spatial location, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iteration to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

[0072] This embodiment implements S103 for online obstacle avoidance scenarios in park roads and computer room corridors. The entry data includes the dynamic body movement trajectory, spatial location and interference area markers output by S102, the rasterized map and preset inspection path generated by S101, and the attribute characteristics of the passage area (channel width, ground friction, slope, occlusion probability, etc.).

[0073] First, the state and constraint model of the obstacle avoidance planner is constructed. The state vector is x=[p_x,p_y,ψ,v], representing planar position, heading angle, and linear velocity, respectively. The control variable is u=[a,ω], corresponding to acceleration and angular velocity. To be consistent with the dynamics of the mobile chassis, a single-track vehicle model with nonholonomic constraints is discretized, with the step size aligned with the perception refresh rate. Hard constraints cover the maximum turning rate, maximum acceleration, and safe profile expansion, while soft constraints are derived from traffic attributes, such as limiting instantaneous curvature in slippery areas and suppressing lateral sway in narrow corridors.

[0074] This embodiment maps the interference region identifier to the same coordinate system of the rasterized map. During mapping, the dynamic body is expanded to the prediction time domain according to the spatiotemporal segments, and each time t corresponds to a "future occupancy layer". A three-dimensional occupancy raster is formed by the index aligned with the robot's time axis.

[0075] For regions with overlapping multiple targets, maximum occupancy confidence synthesis is employed, and morphological dilation is applied at the edges to cover positioning errors. While marking temporarily impassable cells, a "soft prohibition" layer is retained. Interference with insufficient confidence or only temporary occurrences is improved at a cost rather than hard-blocked, avoiding path oscillations. After mapping, a marking record is generated, containing grid number, occupancy status, and validity period. The validity period is estimated by target velocity and grid size to ensure timely release of space after dynamic obstacles are removed.

[0076] This embodiment adjusts obstacle avoidance parameters based on the characteristics of the passage area, forming a combination of cost field and constraints. The passage width W directly affects the feasible zone of the path; the smaller W is, the more inclined to choose a low-curvature trajectory. When the friction indicator M is low, the braking priority is increased. Slope and roughness affect the upper bound of speed and wheel speed difference constraints. To ensure parameter self-consistency, this embodiment configures a lightweight policy network, which acts only as a weight recommender rather than a direct decision-maker. The inputs include W, M, local obstacle density, and historical passage frequency, and the outputs are cost coefficients and curvature / speed priority coefficients. Their inherent relationship conforms to empirical rules. For example, when W narrows and density increases simultaneously, the network recommends increasing the curvature penalty and deceleration weights to avoid sharp turns at narrow bottlenecks. The network output is then subjected to physical upper bound pruning and rule verification to prevent extreme values ​​that violate dynamics.

[0077] This embodiment employs state-space iterative search to find the optimal obstacle avoidance path, specifically using model predictive control in the rolling time domain as its core. The objective function integrates travel cost, obstacle distance, curvature smoothing, and deviation from the "regression mainline." In the reference construction, the preset inspection path is truncated into a local reference band, with the centerline and width defining the allowable deviation range. The dynamic occupancy layer is updated on the predicted horizon, constraining the robot's position to prevent it from falling into the hard prohibition grid, while the soft prohibition region is separated by the obstacle distance potential field. To overcome local minima, this embodiment generates multiple initial guessed trajectories at the beginning of each MPC iteration: straight deceleration, left detour, right detour, and stop / wait. These are then evaluated using the same objective function, and the lowest-cost path is selected as the hot start, enabling rapid strategy switching when encountering obstacles at forks or ends.

[0078] This embodiment requires instruction generation after path generation. Discrete control sequences are translated into speed and angle commands executable by the chassis, with timestamps and fault-tolerance windows, and safety monitoring logic is injected: if a hard prohibition layer is detected encroaching on the planned trajectory in the next frame, braking with the maximum deceleration is immediately triggered; if a soft prohibition layer maintains high confidence for more than three frames and can be bypassed, a lateral bypass branch is prioritized. During execution, tracking errors and signs of slippage are monitored. If speed and attitude deviations exceed thresholds, the MPC time domain is shortened and braking weights are increased to avoid control oscillations due to excessive aggression. Obstacle avoidance is judged as the local dynamic occupancy layer being cleared or the robot returning to the vicinity of the reference zone centerline with a "return to the main line" deviation below a threshold.

[0079] This embodiment considers a narrow passageway case in a computer room corridor: the trolley moves slowly in the same direction and occupies the middle of the passage, and the interference mapping gives a long, impassable strip. During the parameter adjustment phase, because W is relatively small and M is relatively high (good floor friction), the system tends to detour along the edge with low curvature. MPC converges on the initial value of "detour to the right", and the speed curve is compressed to prevent lateral collisions. After detouring to the rear of the trolley and clearing the occupied layer, the deviation term is gradually pulled back to the center line, and the robot smoothly returns to the preset path. In the example of oncoming pedestrians on the park road, the interference angle is large and the duration is short. The policy network increases the deceleration weight, MPC chooses to give way briefly and maintain the center of the lane. After the pedestrian passes, the original speed is restored by acceleration constraint, and there are no extra bends when returning to the main line.

[0080] To enhance interpretability, this embodiment internally records a simplified cost function for ease of implementation:

[0081] .

[0082] In the formula, J represents the total planning cost; J_track represents the deviation cost from the preset inspection path, reflecting the intention to return to the main line; J_obs represents the obstacle distance penalty, derived from the distance between the hard / soft prohibition layer and the potential field; J_smooth represents the control smoothing term, limiting the changes in curvature and angular velocity; J_speed represents the velocity deviation term, constraining the speed within the allowable range of the passage attributes; α, β, γ, and δ are weighting coefficients, whose values ​​are determined by the aforementioned strategy network and rule synthesis. The dimensions and physical meanings of each term are clear, corresponding to both chassis dynamics and on-site risks.

[0083] In this embodiment, the control chain is encapsulated as a temporary obstacle avoidance instruction set and written into the navigation system. The instructions include timing sequences, speed-turn angle curves, emergency strategies, and return trigger conditions. After the task is completed, the system cleans up the temporary path data, retaining only key event summaries and parameter trajectories for offline auditing and online self-calibration. Through this process, the obstacle avoidance planner translates the spatiotemporal risks of S102 into executable motion plans, maintaining safety boundaries in complex but regular inspection scenarios without losing the rhythm of returning to the main task.

[0084] As described above, the intelligent navigation method for inspection robots provided in this application can effectively represent the environment through an innovatively designed map fusion system and multi-source data processing. It constructs an environmental perception mechanism, combining semantic analysis and trajectory prediction to establish a reliable obstacle recognition model. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively solves the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots.

[0085] In one embodiment of the intelligent navigation method for inspection robots in this application, the method may further include the following:

[0086] Step S201: Read the historical map information of the inspection robot, construct a multi-source data fusion module, convert the laser point cloud map data into a standard coordinate system, perform noise reduction and spatial clustering processing on the point cloud data, extract and match feature points from the visual map data, construct a map record table containing data source, collection time, and spatial location, and write the map record table into the fusion database.

[0087] Step S202: Read the map record table in the fusion database, construct a weight calculation module, perform registration processing on the point cloud map and the visual map, calculate the matching degree coefficient of the overlapping area, generate a confidence weight vector based on the matching degree coefficient, combine the weight vector with the map data in a weighted manner to generate fused map data, and write the fused map data into the map database.

[0088] This embodiment addresses long-term inspections of park roads and computer room corridors, implementing S201 and S202. The entry point contains historical map information generated from previous operations, including source type (laser / vision), acquisition batch, timestamp, initial pose value and drift statistics, sensor status, and environmental labels (day / night, wet / dry).

[0089] Specifically, the coordinate reference unification is first completed in the multi-source data fusion module: the latest global reference coordinate system R0 is read, the external reference odometry error of each historical trajectory segment is corrected in a chain, the laser point cloud map is transformed to the standard coordinate system under R0, and the beam motion distortion is compensated based on the rotation center and scanning time. The point cloud data then undergoes noise reduction and spatial clustering: voxelization downsampling reduces redundancy, statistical filtering removes isolated points, and region growing is performed based on normal consistency and density threshold to extract ground, facade, and slender rod clusters; in the computer room corridor, cellular-scale anisotropic voxel grids are used for slender cable trays and cabinet columns to avoid over-smoothing. Robust feature points and line segments are extracted from the visual map data, and a cross-batch keyframe selection strategy is adopted to prioritize the retention of frames with good clarity, texture richness, and exposure balance. Then, geometric consistency verification is performed based on external reference coarse matching to establish a cross-source corresponding anchor point set.

[0090] This embodiment organizes the processed entries into a map record table. Each record specifies the data source (LiDAR / Camera), acquisition time and time-domain batch number, spatial location (pose in R0 coordinates), quality indicators (noise level, echo intensity distribution, image sharpness), semantic overview (corridor / road / equipment area), and coverage boundary. The record table is designed with subsequent weight calculations and backtracking audits in mind; therefore, it retains the original fragment index and preprocessing version number for quick location during reregistration or error tracing. The record table is written to the fusion database, and a secondary index is created based on spatial grid and time batch. Querying any region allows for rapid retrieval of historical multi-source fragments.

[0091] In this embodiment, a weight calculation module is constructed in S202 to register the point cloud map and the visual map. First, overlapping candidate segments are found based on the spatial index of the record table, and the initial pose value under R0 is used to align to a coarse alignment accuracy of 10–30 cm. The geometric registration link adopts a joint strategy of "point-to-surface ICP + line feature alignment": the point-to-surface distance is used to converge the facade clusters, and the horizontal drift of long boundary line segments in corridors and roads is corrected by directional consistency and intersection position constraints; on the photometric side, local NCC and gradient correlation are calculated in the overlapping image blocks to eliminate false high similarity caused by strong reflection or repeated textures. Reflectivity and depth stability are introduced as elimination factors for areas such as shaded areas, glass walls, and bright metal surfaces in the park to weaken their impact on the overall registration. After obtaining stable cross-source correspondence, the matching degree coefficient of the overlapping area is calculated, including geometric residual quantile, directional consistency score, photometric similarity, occlusion consistency, and temporal adjacency consistency, forming a multi-dimensional matching degree vector m.

[0092] This embodiment translates the matching degree vector into a confidence weight vector. The basic principle is "prioritizing geometric credibility, enhancing confidence through temporal consistency, and reducing weights for isolated modalities." The point cloud weight w_p is comprehensively given by the point-to-surface residual, normal stability, echo intensity variance, and ground / facade ratio; the visual weight w_v is given by NCC, gradient correlation, feature repetition penalty, and exposure balance; historical drift statistics and the current batch sensor status are used as global adjustment factors. To avoid being overly sensitive to a single indicator, this embodiment uses interpretable weighted normalization instead of black-box fitting, and limits the weight amplitude within an empirical range. The physical meaning of the weights is clear: high w_p means sufficient geometric support and low measurement noise; high w_v means reliable texture and consistent light intensity. The two complement each other in different scenarios; the facade of the computer room is more dependent on w_p, while the details of the park intersections are more dependent on w_v.

[0093] The occupancy evidence and boundary geometry within the spatial grid are weighted and combined separately: occupancy probability is updated using Bayesian methods, point cloud observations are written with w_p as the prior strength, and visual observations are adjusted with w_v to sharpen the boundaries; the geometric skeleton is fitted to planes and lines using dense areas of the point cloud, and visual observations provide boundary fine-tuning and hole filling. To prevent temporally inconsistent data contamination, this embodiment introduces time-dependent decay and scene semantic adjudication for cross-batch contradictory observations. For example, if the location of a corridor wall is stable over a long period, and a false "opening" is caused by close-range reflection in a certain batch, this anomaly is suppressed during fusion using long-term stable support. The fusion output is fused map data, including occupancy grids, probabilities, boundary feature sets, and confidence layers, and is written to the map database, recording the fusion version and source weight summary.

[0094] The key point of this embodiment lies in the "causal closed loop" of registration and weighting. The matching degree obtained from registration is not only used for one-time weight estimation, but also fed back to the record table to update the reliability statistics of the segment, so that its availability can be judged more quickly when entering the adjacent region next time. Conversely, the weight determines the contribution of each source in the fusion, which directly affects the subsequent rasterization and path connectivity calculation in S101.

[0095] For example, the metal rack array in the narrow passage of the computer room has a clear boundary on the point cloud but visual reflection. The geometric term in m is dominant, w_p is increased, and the fusion boundary is based on straight line fitting. The zebra crossing boundary at the intersection of the park is visually clear but the point cloud is sparse. w_v is increased, and the fusion map retains the texture-guided boundary refinement at this point. This complementarity ensures that subsequent paths do not get too close to the misjudged opening in the visible details.

[0096] This embodiment considers anomalies and sparse data in its engineering details. If a certain area has only single-source historical coverage, the weighting module allows for a degradation mode: if there is no visual data, w_p is estimated separately based on geometric consistency and historical stability, with the default confidence of visual data set to a low value; if there is no point cloud, w_v is estimated based on the stability of image structured light or sparse features, while increasing the confidence threshold of the fusion output to avoid including low-support openings as passable. For low-light scenes in the computer room at night, visual clarity decreases, and the weights converge to be dominated by point clouds; during the backlight period in the park during the day, point clouds are unstable when penetrating the tree canopy, and visual data becomes the boundary reference. These adjustments are consistent with natural conditions.

[0097] For ease of verification, this embodiment uses a simplified combinatorial model to describe the weighting process:

[0098] ,

[0099] Where P_f represents the occupancy evidence of the fused raster, P_p represents the point cloud observation contribution, and P_v represents the visual observation contribution; α and β correspond to the normalized weights of w_p and w_v in that raster, respectively, and they are derived from the matching degree and quality index, satisfying α+β=1. Each parameter in this expression has a clear physical meaning.

[0100] This embodiment ultimately writes the fused map data into a map database, along with overlap, residual statistics, a weight distribution heatmap overview, and anomaly markers. The database maintains a time series of different versions of the same area, allowing S101 to select the latest and most reliable version when constructing the raster map and connectivity weights. Through the linkage of S201 and S202, a stable link is achieved from raw multi-source data to an interpretable fused map, resolving the registration and fusion challenges caused by historical drift, perception instability, and cross-scene differences during long-term inspections. The output directly serves subsequent steps of path generation and online obstacle avoidance.

[0101] In one embodiment of the intelligent navigation method for inspection robots in this application, the method may further include the following:

[0102] Step S301: Read the fused map data from the map database, construct a grid division module, rasterize the fused map data according to a fixed resolution, mark the grid occupied by static obstacles, generate a grid distribution matrix of passable and impassable units, construct a grid record table containing grid number, occupancy status, and location information, and write the grid record table into the planning database.

[0103] Step S302: Read the grid record table in the planning database, construct the path planning module, calculate the connectivity parameters between passable units, generate a connectivity weight matrix, optimize the path search based on the weight matrix, generate a preset inspection path, smooth the preset inspection path, and write the preset inspection path into the navigation control system.

[0104] This embodiment focuses on the long-term inspection tasks of campus roads and computer room corridors, and the entry point is the fused map data in the map database, which includes the occupancy probability layer, the boundary feature set and the confidence layer.

[0105] To avoid directly transferring the scale and precision differences in the fusion process to the planning domain, this embodiment first determines a fixed resolution r and a reference origin in the grid division module. The value of r is based on the chassis outline and the minimum turning radius, so that the cell width can express the safety boundary without producing excessively fine jagged edges.

[0106] Occupation evidence in the fused map is stored in Bayesian form. The module projects this evidence onto the ground and accumulates it on each cell. Elements spanning multiple floors or elevated structures are removed based on a height threshold to prevent ceilings and hanging structures from accidentally entering the driving surface. Considering the voids caused by sparse sensing, this embodiment uses hole-closing operations guided by boundary features to fill small gaps, but retains labels with lower confidence for subsequent weight reduction processing.

[0107] This embodiment establishes the identification of static obstacles based on long-term consistency criteria. The time series of fused data is read, and the occupancy stability of each cell in historical batches is statistically analyzed. Stable cells with semantics such as walls, cabinets, and fixed railings are marked as static occupants. For short-term occurrences with semantics such as pedestrians, carts, and piles of fallen leaves, unstable objects are marked as candidate dynamics and are not hardcoded as static obstacles for the time being.

[0108] To avoid "false openings" caused by visual reflections, in continuous zones with stable point cloud normals on the facade, isolated low-occupancy cells are corrected for occupancy using connectivity. A raster distribution matrix of passable and impassable cells is ultimately generated. In addition to binary states, matrix elements include source confidence and boundary distance, used for subsequent weighting. The raster record table is written cell-by-cell into the planning database, with fields including raster number, center coordinates, occupancy status, source layer summary, and timestamp, supporting incremental updates per spatial block.

[0109] This embodiment reads the grid record table in S302 and constructs a path planning module. The key is that the definition of connectivity parameters must consider safety, drivability, and historical accessibility. The module establishes a grid graph using eight neighborhoods, assigning a cost to the edges between adjacent drivable cells. This cost is a combination of multiple factors: the inverse of the distance to an obstacle represents the safety margin pressure; the reverse term of cell confidence expresses perceived risk; ground attributes (roughness, slope) project onto speed constraints to form energy consumption costs; and historical frequency forms a preference term. In the computer room corridor, cells near cabinet door handles and corners will have increased edge costs due to the risk of collision; after rain, the penalty for curvature terms increases in low-friction areas on park roads, making the path more inclined to go straight to the inner safety belt of the shoulder. These factors are converted into connectivity parameters and summarized in the weight matrix, making the path search favor more robust connectivity links.

[0110] This embodiment employs a heuristic graph search combined with local re-evaluation for path search. At the global level, A... Or D The algorithm calculates the shortest cost path to each inspection point on the weight matrix. The heuristic is not a simple Euclidean distance, but a directed distance projected along the main road / corridor direction, reducing back-and-forth contact with walls. During the search, when encountering large low-confidence areas, the algorithm triggers a candidate channel search, attempting to detour to a higher-confidence corridor or lane. If crossing is unavoidable, the path is marked with a "requires verification" label to facilitate additional detection during execution. For multi-target inspection scenarios, the module introduces a task layer on the global graph, approximating the node access order using a recent insertion strategy or a small-scale TSP (Time-of-Stake) approach, avoiding repeated backtracking at distant points.

[0111] The preset inspection path generated in this embodiment is not a direct splicing of the grid center polylines, but rather geometrically smoothed to meet the chassis kinematics requirements. The smoothing employs curvature-constrained splines to transform the polylines into C2-continuous trajectories. The upper limit of curvature is determined by the chassis limits and the road / corridor width. A "distance-to-edge" term is added to sections near obstacles to prevent the smoothing from pulling the path too close. If there are doorways or intersections at corners, the spline nodes are laterally offset while maintaining accessibility, providing sufficient field of view for vision / radar to detect lateral objects in advance. For areas with narrow bottlenecks, the smoother adds virtual constraint points on both sides to prevent excessive curve undulation within the bottleneck, which could cause tracking sway.

[0112] This embodiment takes into account the execution attributes of the path. In addition to the geometric trajectory, the path object also includes speed and action tags. The speed curve is set with an upper limit based on the grid slope and roughness. The cruising speed is given for long straight sections, and the speed is automatically reduced according to the curvature of the curves. At waypoints in the machine room where nameplates need to be photographed or instruments need to be read, action markers of "passing at low speed or pausing briefly" are added. The subsequent navigation control system can trigger perception tasks according to the markers. To reduce frequent replanning during the execution period, "reserved avoidance sections" are marked in the path. For road sections with a high probability of dynamic interference, a wider feasible band parameter is set so that S103 can detour locally within the band and then return without disrupting the overall progress.

[0113] To enhance the explanation, this embodiment provides an expression for the edge cost:

[0114] .

[0115] In the formula, E_edge represents the edge cost between two adjacent passable cells; d is the distance from the edge centerline to the nearest static obstacle, representing the safety margin; q is the cell confidence level, representing the reliability of the fused perception; s is the projected amount of the ground attribute cost, which can be synthesized from slope and roughness; f is the penalty for historical traversal frequency (a higher value is taken for fewer historical traversals to encourage cautious exploration); a, b, c, and h are non-negative coefficients, set according to the scene semantics (corridor / road) and the weighted summary of the fused map. The lower the cost, the higher the connectivity, and the more likely the path search will select that edge.

[0116] This embodiment verifies the above process in two typical scenarios. At a dead-end corner in the computer room corridor, the wall boundaries in the fused map are straight, and the doorways are accurately marked. After rasterization, the confidence of the cells within the doorways is low, and the weight matrix forms a cost peak at the center of the doorway. A provides the turning path from the corridor centerline near the inner corner. After smoothing, the curvature does not exceed the limit, and a margin from the wall is retained. There are scattered false alarm openings under the shade of trees on the park roads. In the static recognition stage, openings are suppressed with long-term consistency. Connectivity forms a stable channel on the main road surface, and the path does not exhibit zigzag edges when crossing shadows. Finally, the preset inspection path is written into the navigation control system, including node coordinates, spline parameters, speed limit, action labels, and verification marks. Subsequent steps S102 and S103 use this as the reference line and regression target. The interface fields between the path and perception / obstacle avoidance are consistent, facilitating rapid weight switching and anomaly tracing during execution.

[0117] In one embodiment of the intelligent navigation method for inspection robots in this application, the method may further include the following:

[0118] Step S401: Read the sensor data streams from the depth camera and LiDAR, construct a data acquisition module, synchronize the depth image data and laser scanning data in time, generate an environmental perception dataset, perform target segmentation processing on the environmental perception dataset, construct a perception record table containing target contours, depth information, and scanning intensity, and write the perception record table into the acquisition database.

[0119] Step S402: Read the perception record table in the acquisition database, construct a scene analysis module, perform semantic classification on the environmental perception data, extract the width parameter and road surface features of the passage area, generate a scene attribute vector, establish a motion prediction model based on the scene attribute vector, and write the motion prediction model into the prediction database.

[0120] This embodiment focuses on the long-term inspection of campus roads and computer room corridors, implementing S401 and S402. The entry point is the raw data stream from the depth camera and LiDAR. A robust spatiotemporal reference is first established in the data acquisition module. For time synchronization, the hardware timestamps of the two sensors are read, master clock alignment is used, and short-window interpolation correction is introduced during turns or acceleration to prevent pose drift within the radar scan cycle from stretching the same target into a tilted strip. Spatial alignment is based on extrinsic parameters obtained from offline calibration, projecting the depth camera point cloud onto the radar coordinate system, and compensating for the line exposure time of the depth frames during rolling shutter. To suppress sensor noise and environmental interference, the module performs median filtering on the radar intensity and edge-preserving bilateral filtering on the depth map. Subsequently, a joint dense point set is constructed within the overlapping field of view to provide consistent geometric support for downstream segmentation.

[0121] This embodiment performs target segmentation on an environmental perception dataset, employing a two-stage process of "geometric initial segmentation + appearance refinement." In the geometric stage, region growing and normal clustering are performed on the joint point set to coarsely segment the ground, facades, slender members, and volumes. For cabinet rows in the computer room, planar consistency and parallel bundle constraints are used to classify continuous facades into unified instances. On park roads, floating points are eliminated using the height residual of the ground model. In the appearance refinement stage, the grayscale / color texture from the depth camera and radar intensity are projected onto the same pixel grid. A lightweight semantic segmentation network is invoked to output pixel-level categories and instance boundaries. The network input is RGB-D + intensity channels, and the output includes category masks and instance IDs for pedestrians, vehicles / carts, walls, curbs, vegetation, etc. The model is constrained to have clearly interpretable features: intensity and normal variations primarily affect the distinction between metal / glass and solid bodies, while gradient textures are used to distinguish between road markings and cracks. These signals correspond to physical properties, avoiding uninterpretable black-box judgments. Once completed, the target contour (instance ID, polygon or bounding box), depth information (center depth, voxelized point set), and scan intensity statistics are written into the perception record table, along with frame time, sensor pose, and overlapping field of view quality labels, and written into the acquisition database in batches for traceability.

[0122] In this embodiment, the perception record table is read in the scene analysis module to perform semantic classification and extraction of passage area parameters. First, ground-type areas are semantically split into corridor / road categories: for computer room corridors, the lateral passable width W and eccentricity Δ are estimated by fitting the corridor centerline using RANSAC along the main direction parallel to the walls; for park roads, lane boundaries are located using curb and marking masks, and curves are detected based on curvature continuity. Regarding road surface features, slope S, roughness G, and friction indicator M (indirectly inferred from texture micro-contrast and intensity fluctuations) are estimated from ground point sets, and a slippery label is given for water-reflective areas. At this point, a scene attribute vector A = {W, S, G, M, O, R} is generated, where O is the occlusion probability (derived geometrically from instance overlap and field-of-view occlusion), and R is the region's historical traversal reliability (mapped from the confidence layer of the upstream fused map). Each component of the attribute vector is directly related to geometry and appearance statistics; for example, W comes from the distance between boundaries, and M comes from road surface micro-texture and intensity variations. These quantities have stable physical meanings in different scenes.

[0123] This embodiment establishes a motion prediction model based on scene attribute vectors, aiming to provide common-sense trajectory extrapolation for subsequent collision detection and obstacle avoidance. Considering the differences in the types of moving subjects, the model adopts a category-adaptive state-space representation: for pedestrians, a constant velocity + social force extended Kalman filter is used, with the state including position, velocity, orientation, and intention bias. The social force term uses obstacle distance and corridor centerline as potential fields to restrict unreasonable crossings; for vehicles / carts, a constant acceleration + lane constraint model is used, introducing heading stability and minimum turning radius; for service robots and similar targets, a more conservative velocity change bound is considered. The model input is the previous state, scene attribute vector A, and the state set of neighboring dynamic objects, and the output is the trajectory parameters and confidence interval for the next several steps. The relationship between parameters and predictions follows natural laws: when W narrows and S increases, the lateral velocity variance and reachable curvature range tighten; when M is low (slippery), the model gives a shorter braking distance suggestion and reduces the available acceleration; when O increases (more occlusion), the extrapolation confidence interval widens to allow collision detection to use a more conservative envelope.

[0124] This embodiment gates the data quality before generating predictions. If the depth point set of an instance is sparse or the intensity fluctuations are abnormal, the sensor status in the acquisition records is checked to determine if it is transient noise. If so, a short-window robust estimation is used to replace the single-frame value, and the weight of that instance in the model is reduced to avoid unreliable input dominating the results. For false contours caused by glass reflections, depth consistency and multi-view occlusion logic are used to remove them. If necessary, only the appearance trajectory is retained and not used for volume estimation. All instances that pass the gate are registered in the prediction database with the prediction model handle and the latest parameters, as well as the association record with scene attributes, to facilitate downstream collision detection in explaining "why this prediction is made".

[0125] This embodiment uses motion prediction results and traffic attributes together for a structured representation of traffic understanding. For each small segment of corridor or road, a "dynamic occupancy time slice" is calculated in the prediction time domain. This involves squeezing the bounding volume of each dynamic object onto a ground grid according to its future trajectory, generating a spatiotemporal occupancy index. Based on this, a risk summary is attached to each segment (e.g., high-angle oncoming objects, slow-moving tailing in the same direction, increased probability of crossing). These summaries directly correspond to the interference assessment indicators in S102, avoiding redundant estimation. Thus, the scene analysis module not only outputs static geometry and semantics but also provides a quantitative characterization of short-term evolution, facilitating the subsequent weighting of avoidance strategies.

[0126] This example demonstrates a scenario in a computer room corridor: Two maintenance personnel walk slowly side-by-side. During the segmentation phase, they are first geometrically separated into two instances, semantically classified as pedestrians, with the width (W) estimated as narrow. The social force model treats the two as a mutually constrained small group. The lateral component is suppressed by the corridor boundary, and the predicted trajectory is a slow forward movement in the same direction with a narrow confidence interval. On a park road, a cart approaches against the flow of traffic along the sidewalk. The road surface (M) is low (slippery after rain). The vehicle model gives a large braking demand prediction, and the trajectory angle is large with the main lane. The risk summary is labeled as "short-term oncoming crossing." Both outputs are written into the prediction database, retaining the attribute vectors and model weights. The subsequent obstacle avoidance planner can set different detour or yield preferences based on these interpretation fields.

[0127] This embodiment preserves a simplified gating-prediction coupling expression:

[0128] Risk radius ,

[0129] Where r is the safety expansion radius used to generate the dynamic occupancy layer, r0 is the chassis foundation safety radius, W is the local passage width, S is the slope index, M is the friction indicator, O is the occlusion probability, and k1~k4 are non-negative coefficients derived from historical statistics and rule settings. The physical meaning of each parameter in this formula is clear, reflecting the common sense that "the narrower, steeper, smoother, and more easily occluded the terrain, the larger the safety radius." It is used to uniformly expand the envelope before predicting the landing of the grating, reducing the risk of missed collisions.

[0130] The process in this embodiment maintains traceability at the database level: the database stores segmented geometric and intensity evidence, and the prediction database stores scene attributes and model parameters and their sources, which are linked using timestamps and spatial block indexes. Through the connection between S401 and S402, the system starts from the raw quantities of asynchronous sensing and gradually constructs a semantic-motion joint representation consistent with scene geometry, materials, and human / vehicle motion patterns. This provides a stable and interpretable basis for subsequent interference assessment and obstacle avoidance planning, reducing misjudgments and unnecessary stagnation caused by noise and reflections in narrow corridors and complex park environments.

[0131] In one embodiment of the intelligent navigation method for inspection robots in this application, the method may further include the following:

[0132] Step S501: Read the motion prediction model in the prediction database, construct the obstacle analysis module, calculate the center coordinates and volume boundary of the dynamic obstacle, generate spatial position parameters, perform time-series tracking of the dynamic obstacle, calculate the obstacle movement trajectory based on the motion prediction model, construct an obstacle record table containing target number, position parameters, and trajectory data, and write the obstacle record table into the obstacle database.

[0133] Step S502: Read the obstacle record table in the obstacle database, construct a collision detection module, map the movement trajectory to a preset inspection path, calculate the spatial interference degree of the trajectory overlap area, generate an interference degree score vector, mark the location of the interference area based on the score vector, construct an interference record table containing area number, interference degree, and spatial range, and write the interference record table into the interference database.

[0134] This embodiment focuses on two scenarios: park roads and computer room corridors, implementing S501 and S502. The inputs are the motion prediction model in the prediction database and the scene attribute vector index output by S402. First, a unified expression of spatial position parameters is established for each dynamic instance within the obstacle analysis module.

[0135] The latest state estimate and category label are read, and a geometric prototype is selected based on the category: a cylinder / ellipsoid is used for pedestrians, and a directed bounding box is used for carts and vehicles. A 3D point cluster is recovered from the joint point set and depth map of the previous frame, and robust fitting is used to obtain the center coordinates c=(x,y,z), orientation θ, and volume boundary b=(l,w,h). To mitigate the volume inflation caused by occlusion, this embodiment maintains the visibility and point cloud density for each instance. If there is a sudden drop, the volume is kept to change gradually while the uncertainty is increased to avoid writing instantaneous noise into the position parameters. Spatial position parameters are recorded using a dual-channel approach: ground projection and upper height bound, adapting to the difference between the low-precision corridor and the open elevation difference of the park.

[0136] This embodiment couples a motion prediction model with online observations for time-series tracking. The tracking framework maintains the state variables using extended Kalman or unscented Kalman filtering, with the state including position, velocity, heading, and category-related intention terms. Observations are c and θ measured in the current frame, and the measurement noise covariance is jointly determined by visibility, density, and sensor quality labels. When the model predicts that the target is constrained by a narrow channel (narrow W) or slippery (low M), the process noise for velocity and steering is tightened, conforming to the natural law of "narrower space, slipperier road surface, and lower maneuverability." After obtaining the filtered state, it is extrapolated to the planning time domain τ based on the prediction model to form a discrete-time series trajectory, with each time step associated with a volume boundary b(t). The trajectory data and position parameters are written together into the obstacle record table of the obstacle database. The record table fields include target number, timestamp, center coordinates, orientation, boundary size, trajectory polyline / spline, confidence interval, and quality label, facilitating rapid downstream retrieval and auditing.

[0137] This embodiment employs a layered tracking strategy for complex crowds or overlapping targets. First, it maintains ID stability through trajectory continuity and category consistency. Then, it introduces a "cluster" layer for individuals experiencing short-term occlusion. The cluster layer is used to evaluate the overall envelope and velocity of the group, avoiding path decision jitter caused by brief ID drift. Once the occlusion is removed, the individual layer resumes its dominant position. For glass reflection ghost images or specular phantoms, this embodiment negates their observation weights through depth consistency and occlusion geometry, allowing the filter to primarily rely on model predictions to navigate abnormal windows and preventing unreasonable jumps in the record table.

[0138] This embodiment enters S502 to construct the collision detection module. First, the movement trajectory of each target is mapped to the reference coordinate system of the preset inspection path. The path is parameterized with arc length s and includes the robot's outline and velocity curve. The volume boundary b(t) is projected onto the ground to form an occupancy segment that evolves over time. Isotropic or directional expansion is used to address the prediction uncertainty. The expansion radius comes from the risk radius rule in S402 to ensure that the worst-case scenario is not ignored. For each time step, the minimum distance, intersection area, and directional angle between the robot's outline and the occupancy segment at the corresponding time s(t) are calculated and recorded as instantaneous interference indices. Considering the differences in hazards of different categories, this embodiment introduces a semantic weight table: pedestrians facing each other have a higher weight than those moving in the same direction; carts in narrow corridors are more sensitive to lateral clearance; and vehicles on park roads have a higher speed weight.

[0139] To prevent short-term noise from causing unstable judgments, this embodiment performs time-based shaping on instantaneous indicators. Instead of simple averaging, it employs quantile fusion within a sliding window to reduce the impact of single-point spikes. If several consecutive frames of "high-risk" markers appear within the sliding window, the interference level is cumulatively increased over time to reflect the risk of continuous blockage. For multi-target overlay regions, the maximum risk principle combined with overlap length reduction is used to avoid repeated risk calculations for the same space. Finally, the interference curve for each target in the prediction time domain is obtained and projected onto the arc length axis as an interference band along the path.

[0140] This embodiment provides an expression for implementing hierarchical scoring:

[0141] .

[0142] In the formula, S represents the interference score; D represents the minimum safe distance between the robot's outline and the segment occupied by the target, indicating the degree of spatial proximity; L represents the duration or arc length of overlap with the path in the prediction time domain, indicating the duration of obstruction; V represents the component of the target's relative velocity in the tangential direction of the path, indicating the intensity of danger in a head-on or following encounter; Φ is a function of the directional angle, with higher values ​​for head-on encounters and lower values ​​for in-direction encounters; α, β, γ, and η are weighting coefficients, adjusted by the scene attribute vector of S402, such as increasing α when the channel width W decreases and increasing γ when the slipperiness M decreases. All parameters have clear physical meanings and are consistent with the actual situation.

[0143] This embodiment transforms the scoring vector into spatial annotation. For continuous intervals where S exceeds the threshold, interference region locations are generated, recording the arc length range, corresponding time window, target ID set, and score peak. For intervals where the score is close to the threshold but unstable, they are marked as "soft interference," allowing S103 to improve the passage cost at the cost field level without hard blocking. In the scenario of a computer room corridor, slow-moving carts in the same direction will generate long and narrow interference bands, with β contributing significantly. When encountering oncoming pedestrians on park roads, α and η dominate in short-term high-risk intervals, resulting in short and strong interference bands. All interference record tables include a source tracing field (used to look up the obstacle record table and prediction model version) when written to the interference database. If the subsequent obstacle avoidance effect is unsatisfactory, the system can use this to analyze whether it is a prediction bias or a scoring setting bias.

[0144] This embodiment adds extra processing to the boundary conditions. If the preset inspection path passes through a low-confidence area, collision detection will simultaneously issue an "uncertain environment" label, prompting S103 to adopt a more conservative speed and a wider safety margin in that segment. If the target disappears at the end of the prediction time domain, the interference score does not immediately return to zero, but decays with a delay based on the target speed and O (occlusion probability), avoiding the misjudgment of a target that has just left the field of view as risk-free. Through the chained implementation of S501 and S502, the system progressively structures the "object-trajectory-collision-interference" data, preserving both physical and semantic basis, and tightly connecting with the cost and constraint interface of S103, allowing subsequent obstacle avoidance planning to directly consume the interference zone and explain its causes.

[0145] In one embodiment of the intelligent navigation method for inspection robots in this application, the method may further include the following:

[0146] Step S601: Read the interference record table in the interference database, construct an obstacle avoidance planner, input the obstacle movement trajectory and spatial position parameters into the planning model, calculate the obstacle avoidance priority coefficient, generate a planning weight matrix, construct a planning record table containing planning number, weight parameters, and priority level, and write the planning record table into the planning database.

[0147] Step S602: Read the planning record table in the planning database, construct the area marking module, convert the interference area identifier into raster coordinates, update the occupancy status in the rasterized map, mark the temporarily impassable units, construct a marking record table containing raster number, occupancy status, and time limit identifier, and write the marking record table into the marking database.

[0148] This embodiment implements S601 and S602 for two application scenarios: park roads and computer room corridors. The input data is the interference record table in the interference database, which includes the area number, interference degree, spatial range, target ID set, and time window. To translate "risk perception" into "executable planning preferences," this embodiment establishes a unified planning model input in the obstacle avoidance planner: for each interference area, the spatial position parameters and movement trajectory of the corresponding target in S501 are read, and they are discretized into a series of spatiotemporal occupancy segments in the prediction time domain and aligned with the arc length coordinate system of the preset inspection path; the scene attribute vector (passage width W, friction indicator M, slope S, occlusion probability O) is also included as an environmental prior for weight adjustment. The logic behind this is that priority should not only focus on the score, but also consider whether there is enough detour space and whether the road surface allows for sudden braking or sharp turns, to avoid planning solutions that are physically difficult to execute.

[0149] In this embodiment, when calculating the obstacle avoidance priority coefficient, a multi-dimensional risk feature is first constructed for each interference region. Where S_peak is the peak value of the interference score, L is the length or duration of the interference band, D_min is the minimum safe distance, V_rel is the relative tangential velocity, Φ is the directional angle function, and W / M / O are local traffic attributes. The obstacle avoidance priority coefficient π is obtained through a physically meaningful mapping: if W narrows or O increases, π is increased for areas with the same S_peak, indicating that areas with small detour windows or poor visibility should be prioritized; if M is low (slippery), the tendency to "decelerate over detour" will be increased.

[0150] This embodiment introduces a lightweight policy model for weight recommendation. The input is the aforementioned 'r', and the output is a set of cost coefficients and priority levels. The model does not directly provide paths but only suggests weights to avoid black-box decision-making. During model training, successful obstacle avoidance trajectories from offline replay are used as samples. The loss is defined as a partial order relationship of intervening earlier for more dangerous situations and being more conservative for narrower obstacles, ensuring that the output conforms to common sense.

[0151] This embodiment generates a planning weight matrix accordingly. The matrix is ​​defined on a local grid strip, and its elements are combinations of cost weights and velocity constraint weights, derived from the superposition of three types of information:

[0152] First, there is the spatial-temporal potential field from the interference zone. The cost increases sharply near the center line of the interference, and according to V_rel and Φ, the frontal field is widened, while the field in the same direction is lengthened.

[0153] Second, there is static modulation from the traffic attributes: in the low M zone, the curvature penalty is increased and the speed limit is lowered, while in the high S zone, the uphill energy consumption item is increased.

[0154] Thirdly, there is a preference derived from historical connectivity. If a certain historical path is unstable (low R), then a risk bias is added to the matrix.

[0155] Priority levels are represented as discrete levels (e.g., P1, P2, P3), used by downstream systems to prioritize cost updates and replanning for regions P1 when computational resources are strained or conflicts arise. A planning record table is then constructed, and each region is written to the planning database. Fields include planning number, time window, weight parameter summary, priority level, associated interference region, and target ID, ensuring that subsequent S103 or online MPC systems can retrieve data via keystrokes.

[0156] This embodiment considers the handling of overlapping and conflicting areas. At T-junctions in the park, there may be overlapping interference between oncoming pedestrians and lateral pushcarts. Handling these separately would result in mutually canceling cost paths. Before generating the weight matrix, the obstacle avoidance planner performs an overlap consistency check: if the potential fields of two high-priority areas conflict spatially, the system introduces a "conservative blockade" strategy, directly marking the overlapping kernel as temporarily impassable and expanding the detour tendency along the passageway to avoid repeated left and right probing at narrow openings; if high and low priorities overlap, the higher priority is covered, and the lower priority only retains the speed penalty, ensuring unified decision-making.

[0157] This embodiment enters the S602 construction area marking module. The module reads the planning record table from the planning database and converts the interference area identifier into raster coordinates. The conversion process uses the back projection of the path arc length to the map coordinates, and then discretizes it according to the raster resolution; a "future occupancy layer" is generated for each time slice t, and the expansion radius and blocking strategy are determined according to the size of π: high priority uses hard blocking, directly marking the corresponding raster as temporarily impassable; medium priority uses hard blocking in the core and sets cost increase at the outer edge for guidance rather than blocking; low priority only writes soft markings and hands them over to MPC for processing in the cost layer. In order to align the recovery after the target leaves, this embodiment records a time-lapse identifier (TTL) for each raster. The TTL is calculated according to the target speed, O and W. If the speed is high or there is more occlusion, the release is delayed to prevent the path from being opened prematurely as soon as the target leaves the field of view.

[0158] This embodiment also handles the relationship between dynamic outlines and static obstacles on the map during marking. When a trolley in a computer room corridor is only two squares away from a corner, direct expansion might block the passage. The area marking module will query the width of the "reserved avoidance strip" on the preset path generated by S101. If there is still a usable strip for a one-way corridor, the blocking level of the grid near the wall will be reduced, allowing the path to pass in edge mode. If W is too narrow, it will be marked as a hard block and the priority will be increased, prompting S103 to slow down and wait in advance. On park roads, when oncoming pedestrians briefly cross, the area marking will generate a narrow and short hard block strip and write soft cost slopes on both sides of it, making the planning more inclined to give way for short-term pedestrians rather than taking a long detour.

[0159] This embodiment constructs the marking results into a marking record table and writes it into the marking database. The record table stores the raster number, occupancy status (hard block, soft cost, normal), timeliness identifier, source planning number, and timestamp for each raster, and retains a reverse index to the interference area for easy subsequent revocation and accountability. To avoid database write amplification, the module updates in batches using spatial blocks and stores consecutive raster cells of the same type using run-length encoding to ensure efficient database entry for markings in long corridors or long lanes.

[0160] This embodiment makes convergent arrangements for timeliness and refresh. If the interference record is not refreshed in the next cycle (target lost or prediction window empty), the area marking module will not clear it immediately, but will decay it by TTL annealing; if the subsequent interference area position shifts but is still adjacent, the module uses a moving window to translate the existing markings instead of rebuilding, reducing path oscillation. For large-scale low-confidence areas caused by sensor anomalies, this embodiment does not directly block them in S602, but adds the risk to the planning weight matrix and transfers it to S103 to reduce speed, avoiding mistaking environmental uncertainty as a hard obstacle.

[0161] The key point of this embodiment is to unify the scoring, geometry, and access attributes into a dual-track output of weight matrix and raster markers.

[0162] On one hand, the weight matrix provides a continuous cost profile for the rolling planning of S103; on the other hand, the grid marking provides hard boundaries for bottom-level collision avoidance. The two are bound together by the planning number and time window to ensure consistency between execution and cognition. In the example of the park fork, when a pedestrian and a cart entering from the right appear simultaneously, the system hard blocks the overlapping kernel, reserves a soft cost channel to the left, and priority P1 pushes S103 to slow down and then move left to bypass it. In the cross corridor of the computer room, the probability of the pedestrian group being occluded at a distance is relatively high (O). The module delays the release of the blockade and maintains the soft cost band to prevent the robot from accelerating too early in the blind corner. Through the cooperation of S601 and S602, obstacle avoidance has moved from "knowing where the danger is" to "how the map reflects and how the planning balances", forming a closed loop at both the interpretability and execution ends.

[0163] In one embodiment of the intelligent navigation method for inspection robots in this application, the method may further include the following:

[0164] Step S701: Read the marker record table in the marker database, construct the parameter adjustment module, calculate the obstacle avoidance constraint conditions according to the attribute characteristics of the passage area, generate the obstacle avoidance parameter vector, calculate the obstacle avoidance path cost using the state space iteration method, construct the path record table containing path nodes, turning angles, and speed components, and write the path record table into the path database.

[0165] Step S702: Read the path record table in the path database, construct the instruction generation module, convert the optimal obstacle avoidance path into motion control instructions, generate a temporary obstacle avoidance instruction sequence, control the inspection robot to perform obstacle avoidance actions, monitor the obstacle avoidance execution status, clear the temporary path data after completing obstacle avoidance, guide the inspection robot to the preset inspection path, construct an execution record table containing instruction number, execution status, and return position, and write the execution record table into the execution database.

[0166] This embodiment implements S701 and S702 for two scenarios: park roads and computer room corridors. The entry data is a marked record table in the marked database, with fields including grid number, occupancy status (hard blockade / soft cost / normal), time-to-live (TTL) identifier, and source planning number. It also reads scene attributes (passage width W, slope S, friction indicator M, occlusion probability O) and the spline and speed limit of the preset inspection path. The parameter adjustment module first extracts a rolling planning window in the robot's reference frame; the window length depends on the speed and reaction distance. Two paths are established for the grids within the window: a cost layer and a constraint layer. Hard blockade cells form geometrically inaccessible constraints; soft cost cells are not blocked, only the spatial potential field raises the passage cost. The constraint layer also derives dynamic constraint boundaries from W, M, and S. For example, when W narrows, it limits the upper bound of curvature; when M decreases, it compresses the available acceleration range; and when S increases, it increases the uphill energy consumption threshold. The logic of this step is to translate the environment and risk labels into a computable controllable feasible domain; otherwise, even if a low-cost path is found in subsequent iterations, it may be impossible to execute due to dynamic infeasibility.

[0167] This embodiment generates an obstacle avoidance parameter vector p, which uniformly describes the planning preferences and constraints of the current window, including {κ_max, a_min, a_max, v_max, d_safe, λ_obs, λ_track, λ_smooth}. Here, κ_max is the curvature upper limit (determined by W and the minimum turning radius of the chassis), a_min / a_max are the deceleration and acceleration upper limits (adjusted by M and S), v_max is the speed upper limit (determined by soft cost coverage and occlusion O), d_safe is the spatial safety distance (derived from the interferometric band expansion result and the fuselage outline), λ_obs is the obstacle potential field weight, λ_track is the regression principal line preference weight, and λ_smooth is the smoothing control weight. To avoid parameter jumps, this embodiment performs exponential smoothing on p across periods and releases λ_obs around cells where the TTL is about to expire, reducing path jitter at the moment the blockade is lifted.

[0168] This embodiment employs a state-space iterative method to calculate the obstacle avoidance path cost, using a rolling model predictive control framework. The state is x = [p_x, p_y, ψ, v], and the control is u = [a, ω]. A discrete nonholonomic model serves as the dynamic constraint. The objective function consists of four terms: path tracking, obstacle distance, control smoothing, and velocity preference. The obstacle avoidance path cost is minimized under the dual constraints of the grid potential field and curvature constraints. Initial solutions are sampled in parallel from four types of prototype trajectories (straight-line deceleration, left-side detour, right-side detour, short stop) to prevent getting trapped in local minima. For each candidate, a rolling simulation is performed along the predicted horizon, accumulating the cost and eliminating schemes that cross hard blocks or violate κ_max / a_max, retaining the lowest cost as the optimal solution at the current moment. To balance interpretability and engineering verification, this embodiment internally records the detailed decomposition of the cost. A sudden increase in cost can be traced back to determine whether it is caused by the obstacle distance or curvature term, facilitating parameter backtracking.

[0169] In this embodiment, when the path is mapped to a "path record table," steering angle and velocity components are included in addition to geometric nodes. Geometric nodes are indexed by time and recorded as a sequence of positions in the robot chassis reference frame. The steering angle is obtained by differentiating the ψ sequence and trimmed to the upper limit of the angular rate. The velocity components are determined by the velocity profile generated by the constraints of v and a, including the desired cruise and minimum deceleration segments. For narrow corners in a machine room, the record table inserts a pre-deceleration node and a "low-speed passage" label before the corner. For oncoming objects in the park, the record table sets a "yield and slow down" node before the intersection for the control module to recognize. The path record table is written to the path database, with fields including path ID, window timestamp, node sequence, angular velocity reference, velocity reference, and parameter vector summary, achieving consistent indexing across modules.

[0170] This embodiment enters the S702 instruction generation module, which converts the optimal obstacle avoidance path into motion control instructions. The conversion step generates chassis-level instruction pairs from the time-indexed node trajectories using a trajectory tracker (e.g., a pure tracking or MPC tracker).<v_cmd, ω_cmd> Each control segment is equipped with a fault-tolerant time window and an emergency strategy (degrading to braking or fine-tuning detour in case of failure). After the command sequence is issued, monitoring is performed to form a multi-source feedback closed loop: the occupancy grids of the odometer, IMU, and near-end sensing are read, and the trajectory tracking error and remaining obstacle distance are calculated; if the error exceeds the limit or the remaining distance falls below the threshold, rapid replanning or emergency braking is triggered. Considering that the soft cost zone may persist but the risk is not high, this embodiment allows small deviations on the tracker side in exchange for more stable control, and the upper limit of deviation is jointly limited by λ_track and the current speed.

[0171] In this embodiment, the criteria for "completing obstacle avoidance" are set on two lines:

[0172] One possibility is that the occupancy layer will be cleared within the window or the interference score will fall below the threshold and remain there for several cycles.

[0173] Second, the robot has returned to the preset inspection path and the regression deviation and attitude difference are within the threshold.

[0174] Once the conditions are met, the instruction generation module clears the temporary path data, retains only the event summary, and switches the navigation mode back to "main path tracking," re-entering the S701 process for the next intervention event. If the TTL expires during obstacle avoidance but the target is still present, the marker database will refresh the soft / hard status, and the instruction terminal will automatically extend or retract the detour zone accordingly, without manual intervention.

[0175] This embodiment illustrates the behavior in a typical scenario. In the computer room corridor, the trolley occupies the center and W is relatively small. After parameter adjustment, κ_max is reduced, and λ_obs forms a steep potential field on both sides of the trolley. Iteration generates the optimal solution of "slow-speed detour along the edge". The angular velocity sequence in the path record table is smoothed and the velocity sequence is compressed. When a pedestrian briefly crosses the road in the park, M is low and O is slightly high. The parameter vector increases the deceleration weight and shortens the predicted horizon. The control module issues an "early deceleration + short stop" command. After the pedestrian passes, the system automatically returns to the preset inspection path according to the regression rule. The return position and time are marked in the execution record table for post-event verification.

[0176] This embodiment provides a total cost form to guide the mapping of parameters to dynamics:

[0177] ,

[0178] Where J represents the total obstacle avoidance cost; J_track measures the deviation of the current path from the preset inspection path zone, reflecting a regression tendency; J_obs is the integral of the obstacle potential field, reflecting the spatial relationship with hard / soft barriers; J_smooth constraint controls the rate of change, suppressing angular velocity and acceleration jitter; J_speed measures the deviation between the speed reference and the environmentally permissible speed, avoiding excessive speed in low-friction or high-slope areas; α, β, γ, and δ are weighting coefficients derived from the obstacle avoidance parameter vector p and adjusted by W, M, S, and O. The physical meaning of each quantity directly corresponds to kinematic and environmental risks.

[0179] In this embodiment, the execution database records the full-link information for each obstacle avoidance operation, including the instruction number, execution status, triggering reason, minimum safe distance, return position, and timestamp, and establishes a reverse index to the path ID and planning number. Through the division of labor and coupling between S701 and S702, the system translates the spatiotemporal risk information in the marker database into a speed-steering sequence that the chassis can execute, maintaining a consistent driving rhythm in narrow corridors and complex intersections, without sacrificing safety boundaries or excessive stagnation in the face of short-term interventions.

[0180] To effectively address the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic planning, and to provide technical support for inspection robots, this application provides an embodiment of an intelligent navigation device for inspection robots that implements all or part of the aforementioned intelligent navigation method. See [link to embodiment]. Figure 2 The intelligent navigation device for the inspection robot specifically includes the following components:

[0181] The path planning module 10 is used to input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and the visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0182] The obstacle avoidance judgment module 20 is used to collect environmental perception data through a depth camera and a lidar, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate an interference area marker.

[0183] The obstacle avoidance execution module 30 is used to construct an obstacle avoidance planner based on the movement trajectory and spatial position, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iterative calculation to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

[0184] As described above, the intelligent navigation device for inspection robots provided in this application can effectively represent the environment through an innovatively designed map fusion system and multi-source data processing. It constructs an environmental perception mechanism, combining semantic analysis and trajectory prediction to establish a reliable obstacle recognition model. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively solves the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots.

[0185] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic planning, and to provide technical support for inspection robots, this application provides an embodiment of an electronic device for implementing all or part of the intelligent navigation method for the inspection robot. The electronic device specifically includes the following components:

[0186] The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the intelligent navigation device of the inspection robot and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the intelligent navigation method and the intelligent navigation device of the inspection robot in the embodiments, the content of which is incorporated herein, and repeated details will not be described again.

[0187] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0188] In practical applications, the intelligent navigation method for inspection robots can be partially executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0189] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0190] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0191] In one embodiment, the intelligent navigation method function of the inspection robot can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0192] Step S101: Input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0193] Step S102: Collect environmental perception data through a depth camera and LiDAR, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate interference area markers.

[0194] Step S103: Construct an obstacle avoidance planner based on the movement trajectory and spatial location, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iteration to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

[0195] As described above, the electronic device provided in this application, through an innovative map fusion system design and multi-source data processing, effectively represents the environment. It constructs an environmental perception mechanism, combining semantic analysis and trajectory prediction to establish a reliable obstacle recognition model. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively addresses the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots.

[0196] In another embodiment, the intelligent navigation device for the inspection robot can be configured separately from the central processing unit 9100. For example, the intelligent navigation device for the inspection robot can be configured as a chip connected to the central processing unit 9100, and the intelligent navigation method function of the inspection robot can be realized through the control of the central processing unit.

[0197] like Figure 3As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0198] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives input and controls the operation of various components of the electronic device 9600.

[0199] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0200] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0201] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0202] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0203] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0204] Based on different communication technologies, multiple communication modules 9110 can be installed in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0205] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the intelligent navigation method for inspection robots with a server or client execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the intelligent navigation method for inspection robots with a server or client execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0206] Step S101: Input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0207] Step S102: Collect environmental perception data through a depth camera and LiDAR, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate interference area markers.

[0208] Step S103: Construct an obstacle avoidance planner based on the movement trajectory and spatial location, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iteration to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

[0209] As described above, the computer-readable storage medium provided in this application embodiment achieves effective environmental representation through an innovatively designed map fusion system and multi-source data processing. It constructs an environmental perception mechanism, combining semantic analysis and trajectory prediction to establish a reliable obstacle recognition model. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively addresses the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots.

[0210] Embodiments of this application also provide a computer program product capable of implementing all steps in the intelligent navigation method for inspection robots where the execution subject is a server or client as described in the above embodiments. When executed by a processor, this computer program / instruction implements the steps of the intelligent navigation method for inspection robots. For example, the computer program / instruction implements the following steps:

[0211] Step S101: Input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system.

[0212] Step S102: Collect environmental perception data through a depth camera and LiDAR, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate interference area markers.

[0213] Step S103: Construct an obstacle avoidance planner based on the movement trajectory and spatial location, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iteration to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

[0214] As described above, the computer program product provided in this application, through an innovative map fusion system design and multi-source data processing, effectively represents the environment. It constructs an environmental perception mechanism, combining semantic analysis and trajectory prediction to establish a reliable obstacle recognition model. Dynamic programming is introduced, and through interference analysis and state optimization, the accuracy of obstacle avoidance is ensured. This method effectively addresses the shortcomings of traditional technologies in map fusion, environmental perception, and dynamic programming, providing technical support for inspection robots.

[0215] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0216] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0217] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0218] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0219] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An intelligent navigation method for an inspection robot, characterized in that, The method includes: The historical map information of the inspection robot is input into the map fusion module. Combined with the laser point cloud map and visual map data, the confidence weight of the point cloud map and the visual map is calculated to generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system. Environmental perception data is collected by a depth camera and a lidar. The environmental perception data is then segmented into targets and classified into scene semantics. The attribute features of the passable area are extracted, a motion prediction model is constructed, the spatial position and volume parameters of dynamic obstacles are calculated, the movement trajectory of the dynamic obstacles is predicted, the movement trajectory is input into a collision detection unit, the degree of interference between the dynamic obstacles and the preset inspection path is determined, and an interference area marker is generated. An obstacle avoidance planner is constructed based on the movement trajectory and spatial location. The interference area is mapped to the gridded map, and temporarily impassable units are marked. The obstacle avoidance parameters are adjusted according to the attribute characteristics of the passable area. The optimal obstacle avoidance path is calculated using state space iteration. Temporary obstacle avoidance instructions are generated, and the inspection robot is controlled to execute the temporary obstacle avoidance instructions. After completing the obstacle avoidance, it returns to the preset inspection path.

2. The intelligent navigation method for inspection robots according to claim 1, characterized in that, The process of inputting historical map information from the inspection robot into the map fusion module, combining laser point cloud data and visual map data, calculating the confidence weights of the point cloud data and the visual map, and generating a fused map includes: Read the historical map information of the inspection robot, construct a multi-source data fusion module, convert the laser point cloud map data into a standard coordinate system, perform noise reduction and spatial clustering processing on the point cloud data, extract and match feature points from the visual map data, construct a map record table containing data source, collection time, and spatial location, and write the map record table into the fusion database. The map record table in the fusion database is read, a weight calculation module is constructed, the point cloud map and the visual map are registered, the matching degree coefficient of the overlapping area is calculated, a confidence weight vector is generated based on the matching degree coefficient, the weight vector is combined with the map data in a weighted manner to generate fused map data, and the fused map data is written into the map database.

3. The intelligent navigation method for inspection robots according to claim 1, characterized in that, The process of constructing a rasterized map, dividing it into passable and impassable units, generating a preset inspection path based on the connectivity weights of the passable units, and writing the preset inspection path into the navigation control system includes: Read the fused map data from the map database, construct a grid division module, rasterize the fused map data according to a fixed resolution, mark the grid occupied by static obstacles, generate a grid distribution matrix of passable and impassable units, construct a grid record table containing grid number, occupancy status, and location information, and write the grid record table into the planning database. The system reads the grid record table from the planning database, constructs a path planning module, calculates the connectivity parameters between passable units, generates a connectivity weight matrix, optimizes the path search based on the weight matrix, generates a preset inspection path, smooths the preset inspection path, and writes the preset inspection path into the navigation control system.

4. The intelligent navigation method for inspection robots according to claim 1, characterized in that, The process of acquiring environmental perception data through a depth camera and LiDAR, performing target segmentation and scene semantic classification on the environmental perception data, extracting access area attribute features, and constructing a motion prediction model includes: Read the sensor data streams from the depth camera and LiDAR, construct a data acquisition module, synchronize the depth image data and laser scanning data in time, generate an environmental perception dataset, perform target segmentation processing on the environmental perception dataset, construct a perception record table containing target contours, depth information, and scanning intensity, and write the perception record table into the acquisition database. The system reads the perception record table from the acquisition database, constructs a scene analysis module, performs semantic classification on the environmental perception data, extracts the width parameters and road surface features of the passage area, generates scene attribute vectors, establishes a motion prediction model based on the scene attribute vectors, and writes the motion prediction model into the prediction database.

5. The intelligent navigation method for inspection robots according to claim 1, characterized in that, The process of calculating the spatial position and volume parameters of the dynamic obstacle, predicting the movement trajectory of the dynamic obstacle, inputting the movement trajectory into the collision detection unit, determining the degree of interference between the dynamic obstacle and the preset inspection path, and generating interference area markers includes: Read the motion prediction model from the prediction database, construct an obstacle analysis module, calculate the center coordinates and volume boundaries of dynamic obstacles, generate spatial position parameters, perform time-series tracking of the dynamic obstacles, calculate the obstacle movement trajectory based on the motion prediction model, construct an obstacle record table containing target number, position parameters, and trajectory data, and write the obstacle record table into the obstacle database. The obstacle record table in the obstacle database is read, a collision detection module is constructed, the movement trajectory is mapped to a preset inspection path, the spatial interference degree of the trajectory overlap area is calculated, an interference degree score vector is generated, the location of the interference area is marked based on the score vector, an interference record table containing area number, interference degree, and spatial range is constructed, and the interference record table is written into the interference database.

6. The intelligent navigation method for inspection robots according to claim 1, characterized in that, The obstacle avoidance planner constructed based on the movement trajectory and spatial location maps the interference area identifier to the rasterized map and marks temporarily impassable units, including: Read the interference record table in the interference database, construct an obstacle avoidance planner, input the obstacle movement trajectory and spatial location parameters into the planning model, calculate the obstacle avoidance priority coefficient, generate a planning weight matrix, construct a planning record table containing planning number, weight parameters, and priority level, and write the planning record table into the planning database. Read the planning record table in the planning database, construct a region marking module, convert the interference region identifier into raster coordinates, update the occupancy status in the rasterized map, mark temporarily impassable units, construct a marking record table containing raster number, occupancy status, and time limit identifier, and write the marking record table into the marking database.

7. The intelligent navigation method for inspection robots according to claim 1, characterized in that, The process of adjusting obstacle avoidance parameters based on the characteristics of the passage area, calculating the optimal obstacle avoidance path using state-space iteration, generating temporary obstacle avoidance commands, controlling the inspection robot to execute the temporary obstacle avoidance commands, and returning to the preset inspection path after completing obstacle avoidance includes: Read the marker record table in the marker database, construct a parameter adjustment module, calculate obstacle avoidance constraints based on the attribute characteristics of the passage area, generate an obstacle avoidance parameter vector, calculate the obstacle avoidance path cost using the state space iterative method, construct a path record table containing path nodes, turning angles, and speed components, and write the path record table into the path database. The path record table in the path database is read, an instruction generation module is constructed, the optimal obstacle avoidance path is converted into motion control instructions, a temporary obstacle avoidance instruction sequence is generated, the inspection robot is controlled to perform obstacle avoidance actions, the obstacle avoidance execution status is monitored, the temporary path data is cleared after obstacle avoidance is completed, the inspection robot is guided to the preset inspection path, an execution record table containing instruction number, execution status, and return position is constructed, and the execution record table is written into the execution database.

8. An intelligent navigation device for an inspection robot, characterized in that, The device includes: The path planning module is used to input the historical map information of the inspection robot into the map fusion module, combine the laser point cloud map and the visual map data, calculate the confidence weight of the point cloud map and the visual map, generate a fused map, construct a raster map, divide the passable units and inaccessible units, generate a preset inspection path based on the connectivity weight of the passable units, and write the preset inspection path into the navigation control system. The obstacle avoidance determination module is used to collect environmental perception data through a depth camera and a lidar, perform target segmentation and scene semantic classification on the environmental perception data, extract the attribute features of the passable area, construct a motion prediction model, calculate the spatial position and volume parameters of the dynamic obstacle, predict the movement trajectory of the dynamic obstacle, input the movement trajectory into the collision detection unit, determine the degree of interference between the dynamic obstacle and the preset inspection path, and generate an interference area marker. The obstacle avoidance execution module is used to construct an obstacle avoidance planner based on the movement trajectory and spatial location, map the interference area identifier to the raster map, mark temporarily impassable units, adjust obstacle avoidance parameters according to the attribute characteristics of the passable area, use state space iterative calculation to calculate the optimal obstacle avoidance path, generate temporary obstacle avoidance instructions, control the inspection robot to execute the temporary obstacle avoidance instructions, and return to the preset inspection path after completing obstacle avoidance.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent navigation method for the inspection robot according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent navigation method for the inspection robot as described in any one of claims 1 to 7.

Images