A method for planning an inspection path based on cooperation of a navigation path and an obstacle avoidance strategy

Abstract

The application relates to the technical field of path planning, in particular to a kind of inspection path planning method based on navigation path and obstacle avoidance strategy cooperation, comprising the following steps: based on the preset inspection target set and the inspection environment map, initial navigation path is generated using graph search algorithm, and task graph structure is constructed;Path deviation graph is constructed and deviation node set is output;The dynamic traffic confidence of local obstacle avoidance path segment is calculated;Obstacle avoidance semantic label is generated according to dynamic traffic confidence;Based on the deviation node set and the remaining unfinished target, local reordering algorithm is executed on the task graph, the inspection target access order is updated, the new optimal path is generated in combination with obstacle avoidance semantic label and current pose.The application introduces the dynamic traffic cost adjusted by semantic label on the basis of considering path length, realizes the cooperative optimization of global optimality of path reconstruction and obstacle avoidance adaptability, and considers the feasibility, shortestness and real-time performance of the algorithm.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and in particular to an inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy. Background Technology

[0002] With the widespread application of intelligent inspection robots in critical infrastructure such as power, chemical, coal mining, and rail transportation, their automated inspection capabilities have become one of the key technologies for ensuring equipment operation safety and improving maintenance efficiency. Existing inspection systems typically rely on preset global path planning algorithms or offline constructed task graphs to generate the optimal path from the starting point to each inspection target in sequence. However, in real-world complex environments, the inspection process is often interfered with by dynamic obstacles, causing the original path segment to be interrupted, and the robot cannot complete the task smoothly according to the original plan.

[0003] Some existing solutions attempt to correct paths using local obstacle avoidance algorithms. However, these methods primarily aim for successful obstacle avoidance and fail to consider factors such as the order of subsequent targets and the overall path length in conjunction with the global inspection task. This often results in issues such as "detours," overlapping or missed tasks, and a lack of overall optimization capabilities. Furthermore, current path reconstruction processes are mostly based on geometric obstacle avoidance information, lacking semantic-level differentiation and weight adjustment for obstacle event types. This prevents differentiated processing for different scenarios such as continuous blockages, temporary obstructions, and high-frequency fluctuations, leading to unstable path recovery.

[0004] Furthermore, after a path is disrupted, there is still a lack of dynamic reconfiguration frameworks with clear logical loops and strong scalability, which address how to quickly determine which inspection targets need to be reordered, how to intelligently adjust the path graph structure based on obstacle types, and how to achieve efficient searching. Especially when requirements such as ensuring closed-loop task execution, highly reliable navigation, and dynamic feedback updates are needed, existing methods struggle to balance real-time performance with global optimization. Summary of the Invention

[0005] This invention provides an inspection path planning method based on the collaboration of navigation path and obstacle avoidance strategy, and an inspection path adaptive adjustment method with semantic understanding, local scheduling and global path dynamic reconstruction capabilities, so as to improve the autonomous inspection level of robot system in complex dynamic scenarios.

[0006] An inspection path planning method based on the collaboration of navigation path and obstacle avoidance strategy includes the following steps: S1, based on the preset inspection target set and inspection environment map, uses a graph search algorithm to generate an initial navigation path and constructs a task graph structure with each inspection target as a node and the navigation path segment as an edge, which is used to represent the path connectivity and task execution order; S2, during the inspection process, based on the robot's body sensors, obstacle avoidance behavior events are captured in real time, and it is determined whether the current navigation path segment is interrupted. If it is interrupted, the shortest path segment from the current position to the nearest inspection target is calculated, a path deviation map is constructed, and a set of deviation nodes is output. Based on the historical backtracking trajectory of the navigation path segment interruption location, the corresponding local obstacle avoidance path segments are extracted from the 3D map. Combined with the point cloud change sequence and image residual map constructed during the obstacle avoidance process, the dynamic passability confidence of the local obstacle avoidance path segments is calculated. Obstacle avoidance semantic labels are generated based on the dynamic passability confidence, including temporary obstruction, continuous blockade and high-frequency fluctuation area, to provide semantic-level obstacle avoidance area constraints for subsequent path reconstruction. S3: Based on the set of deviated nodes and the remaining unfinished targets, execute a local reordering algorithm on the task graph to update the access order of the inspected targets, and generate a new optimal path by combining the obstacle avoidance semantic labels and the current pose to guide subsequent inspections.

[0007] Optionally, the graph search algorithm adopts the A* heuristic graph search algorithm, which uses the weighted sum of path distance and estimated travel cost as the cost function to perform a global path search between each inspection target and generate an initial navigation path set connecting each inspection target.

[0008] Optionally, the construction of the task graph structure includes an undirected weighted graph with all inspection targets as nodes and the initial navigation path segment connecting any two targets as edges, which serves as the task graph structure. The weight of each edge is the actual passage cost of its corresponding navigation path segment. Based on the task graph structure, a minimum spanning tree algorithm is used to generate an optimal or near-optimal access sequence covering all nodes, thereby determining the initial inspection task execution order.

[0009] Optionally, the robot's sensors include lidar, ultrasonic, or vision sensors. During the inspection process, the lidar, ultrasonic, or vision sensors on the robot body detect signals in real time. When an obstacle is detected to have entered the safe distance threshold of the current navigation path segment or the proportion of the path ahead being occupied exceeds a preset threshold, it is determined that the current navigation path segment has been interrupted. If the path is interrupted, the robot's current position is taken as the starting point, and all unfinished inspection targets in the task graph structure are taken as potential endpoints. On the environmental map subgraph after removing occupied edges, the shortest path algorithm is used to calculate the shortest path to each potential endpoint, and the path with the smallest total cost is selected. The inspection target corresponding to this path is recorded as the nearest inspection target, and the shortest path is the shortest path segment.

[0010] Optionally, the construction of the path deviation graph includes taking the most recently inspected target as the root node, taking subsequent inspected targets that could not be visited in the original order due to path interruption as deviation nodes, forming a directed weighted graph with a single source pointing to multiple unfinished targets, recording the topological relationship and cost increment between each deviation node and the original planned access sequence, and forming the deviation node set.

[0011] Optionally, S2 further includes, for the location where the navigation path segment is interrupted, retracing the historical trajectory during the robot's obstacle avoidance behavior, extracting the actual driving trajectory used to bypass the obstacle from the pre-built 3D map, as a local obstacle avoidance path segment; associating the time period of the obstacle avoidance process, extracting continuous frame data of the LiDAR point cloud in this area to form a point cloud change sequence, and extracting the inter-frame image residual map of the visual sensor in the same area at different times; based on the point cloud change sequence, calculating the stability index of the obstacle contour; based on the image residual map, analyzing the dynamic change intensity of the texture area; fusing the stability index and the dynamic change intensity, and using a pre-trained neural network evaluation model, outputting the probability value that the local obstacle avoidance path segment can be traversed again in the future time window, as the dynamic traversal confidence.

[0012] Optionally, obstacle avoidance semantic labels are generated based on the predefined interval to which the dynamic access confidence level belongs: When the confidence level is higher than the first threshold, it is marked as a temporary block. When the confidence level is below the second threshold, it is marked as a persistent block. When the confidence level is in the fluctuation range and the historical labeling frequency is high, it is marked as a high-frequency fluctuation region; the semantic label serves as a semantic-level obstacle avoidance region constraint condition for subsequent path reconstruction.

[0013] Optionally, the execution of the local reordering algorithm includes: Based on the set of deviated nodes and all remaining unfinished inspection targets, a local optimization window is defined on the task graph structure. The definition of the local optimization window includes taking the robot's current position as the starting point and including the nodes in the set of deviated nodes and the nodes that were originally located after the most recently interrupted inspection target and have not yet been visited into the target subset to be reordered. With the goal of minimizing the estimated total cost of starting from the current position, visiting all nodes in the target subset, and finally returning to the preset destination, a tabu search algorithm with variable neighborhood search is used to solve the problem on the task graph structure, and the updated inspection target access order is output. Among them, the calculation of the estimated total cost of the path is combined with the obstacle avoidance semantic label to dynamically weight and adjust the passage cost of the relevant edges.

[0014] Optionally, based on the updated inspection target access order and the constraints provided by the obstacle avoidance semantic tags, a new optimal path is generated, including the following weight update rules: For areas marked as continuously blocked, the corresponding path segments are temporarily set to impassable status in the environment map, and are temporarily removed or given infinite weight in the task graph structure. For edges marked as temporary obstructions or high-frequency fluctuation regions, their weights in the task graph structure are updated to the product of the original passage cost and a dynamic penalty coefficient based on label type and confidence level.

[0015] Optionally, S3 further includes taking the robot's current pose as the starting point of the path, running the graph search algorithm again on the task graph structure after applying the weight update rules, calculating the collision-free path of each inspection target point in the updated sequence segment by segment, and finally synthesizing a complete new global optimal path to guide the robot's subsequent inspection navigation.

[0016] The beneficial effects of this invention are: This invention introduces an obstacle avoidance semantic labeling mechanism to achieve dynamic identification and semantic-level classification of different types of obstacle events. During the path reconstruction process, it differentiates the relevant edges in the path graph structure according to the label type, which enhances the ability to cope with dynamic and complex inspection scenarios. It can effectively avoid the problems of invalid paths or obstacle avoidance failures in practical applications of traditional static path planning methods, thereby ensuring task continuity.

[0017] This invention does not simply reposition itself to avoid obstacles after a path interruption. Instead, it dynamically defines a local optimization window based on the current robot state and remaining task nodes, and employs a combined optimization algorithm that integrates tabu search and variable neighborhood search to rearrange the task order for the subset of targets to be visited. This algorithm introduces a dynamic passage cost adjusted by semantic labels, taking into account path length, achieving a synergistic optimization of global optimality in path reconstruction and obstacle avoidance adaptability, thus balancing the algorithm's feasibility, shortest path, and real-time performance.

[0018] This invention constructs a dynamic accessibility confidence model by fusing multi-source features such as point cloud stability and the intensity of dynamic image changes, thereby enabling dynamic evaluation of path segment reliability and identification of deviation nodes. Combined with a multi-stage dynamic weight update mechanism and an A* algorithm-based segmented path reconstruction mechanism in the task graph structure, a complete, collision-free, updated optimal inspection path can be output, effectively supporting the robot's rapid recovery and closed-loop navigation after task interruption. Furthermore, the constructed path deviation map can serve as crucial data support for subsequent scheduling optimization and environmental modeling. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention 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 only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the path planning method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of path reordering and dynamic path generation in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] like Figures 1-2 As shown, an inspection path planning method based on the collaboration of navigation path and obstacle avoidance strategy includes the following steps: S1, based on the preset inspection target set and inspection environment map, uses a graph search algorithm to generate an initial navigation path and constructs a task graph structure with each inspection target as a node and the navigation path segment as an edge, to represent path connectivity and task execution order.

[0023] Specifically as follows: S11, Initial Navigation Path Set Generation: The task is to find an optimal, passable path between every two target points and aggregate these paths into an initial path set. This is based on a pre-defined inspection target set. Inspection Environment Map This paper introduces the A* heuristic graph search algorithm, a classic algorithm in path planning. It combines the cost of actually traversed paths (historical cost) with the heuristically estimated distance from the current position to the target point, thus finding the optimal path in the shortest time while avoiding blind traversal. Using the A* heuristic graph search algorithm, the optimal path between any two inspected targets is calculated pairwise, forming an initial navigation path set. The path cost function is defined as follows: ;in, Representing a path Total distance, This indicates the estimated travel costs along the route, including congestion and road condition complexity. As a weighting factor, satisfying It is used to make a trade-off between path length and traffic risk. This represents the total cost function used for priority queue sorting in the A* algorithm.

[0024] An inspection environment map is a digital representation of the spatial structure, access conditions, and obstacle distribution of the inspection area where the robot is located. It can be a two-dimensional planar map or a three-dimensional stereo map, depending on the complexity of the application scenario. This map contains the following core elements: Spatial grid or node graph structure: The inspection area is divided into several grids or defined as a navigation graph with discrete nodes connected, which is used by the path search algorithm to perform planning in the graph space.

[0025] The markings for passable and impassable areas correspond to the passageways, corridors, outdoor paths, etc. that robots can pass through in the actual scene, as well as impassable areas such as obstacles, walls, equipment, and fences.

[0026] Cost layer or semantic layer: used to label the passage cost of each area, such as ground slope, complexity, ground material, passage frequency, etc., or additional semantic information, such as flammable areas, slippery areas, dangerous equipment edges, etc.

[0027] Before robot deployment, personnel use equipment such as LiDAR, depth cameras, and inertial navigation IMUs, along with the SLAM algorithm, to create a one-time inspection environment map. The A* algorithm needs to determine which areas are passable and which are not based on the nodes, grids, or raster structures in the map; the passage cost of each path segment in the A* algorithm is calculated based on the passage cost of different areas in the map; in actual operation, if an obstacle is suddenly detected in a section of the map that should be passable, it is considered a path interruption, triggering the obstacle avoidance mechanism.

[0028] The above defines the inspection target set as follows: The A* heuristic graph search algorithm requires computation from... arrive The path, from arrive The path until from arrive ( The path is as follows. For each pair of target points, perform the following steps: 1. Initialize search: On the map The starting point is set as follows: The endpoint is set as And the map is gridded or constructed as a node graph.

[0029] 2. Define the cost function: The quality of a path depends not only on its geometric distance but also on its traversability; therefore, a cost function is defined. The cost function measures the overall cost of a path and is the core criterion for the A* algorithm to determine which path is better.

[0030] 3. Perform an A* search, continuously expanding the search path according to the following rules: Known cost of the current point It is the cumulative cost of getting from the starting point to this point; Heuristic estimation This is an estimate of the straight-line distance from this point to the target point; Total cost Prioritize expansion The smallest node.

[0031] Take from the priority queue each time Expand the smallest node until the target point is reached. .

[0032] 4. After the path search is completed, output the optimal path between the pair of target points. Add it to the set Repeat this process until all arrive All paths have been calculated.

[0033] S12, Constructing the Task Graph Structure: The goal is to transform the path relationships into a task graph structure, facilitating subsequent task sorting and path optimization. Specifically, the initial navigation path set calculated in S11 is transformed into a graph, with inspection targets being nodes in the graph, and the path segment between any two targets being an edge. The cost of traversing a path is represented by the edge weights, which constitute the task graph structure. This graph serves as the foundational data structure for subsequent decisions regarding path reconstruction, sorting, and obstacle avoidance semantic constraints. It's important to note that path edges are not simply lines; they are meaningful, undirected edges, representing the path from... arrive and from arrive The path between them is symmetrically feasible and consists of weighted edges, the weight of which is the cost of navigating the path. It may represent the length of the path, or it may be the actual consumption calculated based on terrain and travel conditions.

[0034] Specifically, all inspection targets As a set of nodes in a graph, the initial navigation path set Each path segment Mapping to an undirected weighted graph using an undirected weighted edge ,and ;in, Represents a node There is no direction between the two sides, and the passage cost is included. For the edge The weight of the navigation path segment The actual cost of passage, This is the task graph structure, used to represent the reachability and cost relationships between inspection targets. S12 transforms the path search results into structured graph data, preparing the data structure for subsequent steps. In minimum spanning tree sorting, a path that covers all targets and has the minimum total cost is selected from this graph. In path reconstruction and local reordering, when encountering obstacle avoidance interruptions or nodes that cannot be reached, the task order can be dynamically planned again on this graph. Because edges have weights, other cost factors can also be introduced for weighted combinations.

[0035] S13, Initial Inspection Task Execution Order Generation: In S11 and S12, we obtained the optimal paths between all inspection targets, as well as an undirected weighted graph with inspection targets as nodes and path traversal costs as edge weights. We need to select a set of connections with the minimum total cost from this task graph to form a path sequence covering all targets. For this, the minimum spanning tree algorithm is used. A spanning tree is not a complete path in the traditional sense, but rather a set of edges connecting all targets, without loops, and with the minimum total cost. It can serve as the basis for subsequently determining the inspection access order. That is, it is based on the task graph structure. The minimum spanning tree algorithm is used to generate an approximately optimal path sequence that covers all nodes. Output the order in which nodes in the minimum spanning tree are visited. This refers to the initial inspection task execution order, used to guide the priority access targets of the navigation path. Let be the set of edges in the minimum spanning tree.

[0036] Minimum spanning tree is a classic concept in graph theory. It refers to finding a set of edges in an undirected weighted graph such that each node is connected at least once, no closed cycles are formed, and the total weight of these edges is minimized. In our inspection task graph, this means selecting the path segments with the minimum total travel cost that cover all targets from all target-to-target paths, and using these segments to construct the initial inspection path sequence. The specific execution process is as follows: 1. Initialization: Starting from any inspection target, such as T1, create two sets: one is the set of connected nodes, which initially contains only T1; the other is the set of remaining nodes to be connected, which initially contains all other targets. Initialize an empty edge set to store the edges of the final spanning tree.

[0037] 2: Find the edge with the minimum cost: Between the currently connected nodes and the unconnected nodes, find the edge with the minimum cost, add this edge to the set of edges of the spanning tree, add the unconnected node connected by this edge to the set of connected nodes, and remove the node from the set of unconnected nodes.

[0038] 3: Repeatedly search for the minimum cost edge until all inspection targets are connected, ultimately forming a loop containing n. A spanning tree with 1 edge, where n is the number of inspection targets.

[0039] 4: By performing a depth-first traversal on the spanning tree, a reasonable visiting order can be generated. Although this order is not the absolute shortest traversal path because the minimum spanning tree does not contain cycles, it is a low-cost, highly connected, and easily reconstructable initial inspection sequence.

[0040] S2, during the inspection process, based on the robot's body sensors, obstacle avoidance behavior events are captured in real time, and it is determined whether the current navigation path segment is interrupted. If it is interrupted, the shortest path segment from the current position to the nearest inspection target is calculated, a path deviation map is constructed, and a set of deviation nodes is output. Based on the historical backtracking trajectory of the navigation path segment interruption location, the corresponding local obstacle avoidance path segments are extracted from the 3D map. Combined with the point cloud change sequence and image residual map constructed during the obstacle avoidance process, the dynamic passability confidence of the local obstacle avoidance path segments is calculated. Obstacle avoidance semantic labels are generated based on the dynamic passability confidence, including temporary obstruction, continuous blockade and high-frequency fluctuation area, to provide semantic-level obstacle avoidance area constraints for subsequent path reconstruction.

[0041] Specifically as follows: S21, when the robot performs inspection tasks, it constantly relies on onboard sensors, including LiDAR, ultrasonic sensors, and vision, to perceive changes in the environment ahead in real time. To ensure safe movement, it needs to determine in real time whether the robot's current navigation path segment is still passable. Once the path segment is blocked by obstacles or poses a safety hazard, it needs to be immediately judged as a path interruption, thus triggering an obstacle avoidance or replanning mechanism. The specific judgment criteria have two complementary conditions: First, insufficient safe distance. Real-time collection of obstacle distance data from LiDAR or ultrasonic sensors. When the detected value is less than the set safe distance threshold, it means that there may be a sudden obstacle ahead, and the robot cannot pass safely, thus judging the path as interrupted. Second, excessive occupation of the path area ahead. Based on obstacle detection results generated by vision or radar, it is determined what proportion of the path area ahead is occupied by obstacles. If this proportion exceeds a preset threshold, it is also judged as a path interruption. These two conditions are OR; either condition must be met to trigger a path interruption response. Specifically, during the inspection process, real-time data is collected from the sensors on the robot itself, including: LiDAR distance measurement ; Ultrasonic distance measurement ; Visual recognition of obstacle outline area .

[0042] When any of the following conditions are met at any given moment: or If so, it is determined that the current navigation path segment has been interrupted.

[0043] in, This indicates the safety distance threshold set for the path segment, with a value ranging from 1.5 to 3 meters. It can also be set to 1.2 to 1.5 times the robot's deceleration and braking distance. This represents the area of ​​the forward region occupied by what is identified as an obstacle. A rectangular perception region can be defined for the forward navigation path segment. Obstacle regions are marked using visual detection boxes. Polygonal contours are fitted to these obstacle regions, and their areas are calculated. The sum of the areas of all obstacle regions gives the total area. , The total area of ​​the current navigation path segment. The threshold for determining the path occupancy ratio is 0.3. This means that if more than 30% of the path ahead is obstructed, the path is considered impassable. If the robot travels at a high speed and has a large inertia, the threshold should be set lower. If the path environment is narrow and has poor fault tolerance, such as an industrial pipe gallery, the threshold should also be set lower.

[0044] S22, Shortest Path Calculation: When the robot encounters an obstacle during inspection that prevents it from continuing its current navigation path (i.e., the path is interrupted), it needs to quickly decide which inspection target to move to next and which path to take to bypass the obstacle. In the original task graph, all the lines connecting the inspection targets were represented as edges, and each edge represented a path segment. Now, we identify those path segments that are actually occupied by obstacles or are impassable, and remove the corresponding edges from the task graph, forming a subgraph. This subgraph represents the currently passable graph structure for subsequent path search. Therefore, if the path is interrupted, all edges containing path segments blocked by obstacles are removed from the current map. , forming a subgraph Then, using the robot's current position as the starting point, all uninspected target points are selected as candidate endpoints, and path searching is performed sequentially. The total path cost for each candidate target point is calculated in the subgraph, which is the path cost based on the robot's current pose. Starting with the set of all unfinished inspection targets For each potential endpoint, execute Dijkstra's algorithm to calculate the target point. Total cost of the shortest path: ; This represents an edge in the path that connects two inspection targets. For the edge The cost of travel. This formula represents the cost of the robot moving from its current position. Departure, to any inspection target Among all possible paths, find the one with the minimum total cost and calculate its total cost, which is the sum of the weights (travel costs) of each edge along the path. Select the target point with the minimum total cost. The corresponding path is the shortest path segment.

[0045] S23, Constructing a Path Deviation Graph: In S22, the nearest inspection target and its corresponding shortest path segment, which were detoured due to path interruption, were identified. However, this raises the question of what to do with other target nodes that were originally ranked after the nearest inspection target—can they still be inspected in the original planned order, or do they need to be reordered? To solve this problem, it is necessary to record the deviation in the target access order caused by path interruption and construct a path deviation graph for subsequent dynamic task reconstruction. Therefore, with… As the root node, the unvisited targets in the original planned visit sequence that are after the breakpoint on the current path are denoted as the set of deviation nodes. And record the cost difference of its deviation from the path: ; This indicates that under the current circumstances, a new route should be replanned for access. The total cost, Indicates access in the original task order The original total cost, This represents the set of deviated targets that could not be visited in the original planned order after the path was broken. A path deviation map is constructed from this. Simply put, a path deviation map is... As the root node, move towards each node that deviates from the target. Create a directed edge connecting all And label the graph structure of the cost increment.

[0046] Specifically, the operation logic for this step is as follows: 1. Select the current access anchor point: The most recently inspected target determined in the previous step is used as the root node of the path deviation map.

[0047] 2. Identify the set of deviated targets: Extract all unfinished targets after the current breakpoint in the original planned inspection path; these targets are considered deviated targets because they may not be able to continue to be visited in the original order after the path changes.

[0048] 3. Calculate the cost difference: For each deviation from the target, calculate separately: In the current situation, the total cost of replanning the path from the current location to the target is the current cost; The total cost of visiting the target in the original order in the original task graph, i.e., the original cost; The difference between the two is the increase in path deviation cost caused by path interruption.

[0049] 4. Take the most recently inspected target as the root; connect each deviated target into child nodes, and label each edge with the corresponding cost increment, forming a graph structure representing the current task path deviation status. This path deviation graph is used for subsequent dynamic path reordering, providing basic data for re-optimizing the target access order.

[0050] S24, Local Obstacle Avoidance Path Segment Extraction: When a robot's navigation path is interrupted by an obstacle during inspection, it will perform obstacle avoidance maneuvers and deviate from the planned path to find a local passage around the obstacle. This process may include dynamic behaviors such as turning, deceleration, reversing, and detouring. To analyze the current traffic conditions, it's necessary to review the robot's actual movement path during this obstacle avoidance period. This path is no longer an ideal straight line or curve in the planned path, but rather the robot's actual trajectory during obstacle avoidance. Therefore, it's necessary to extract this actual obstacle avoidance trajectory and extract the map area traversed by this trajectory from the 3D map to form a local obstacle avoidance path segment for subsequent risk assessment. In other words, it's necessary to extract the historical travel trajectory corresponding to the robot's obstacle avoidance actions from the time the path segment is interrupted until traffic resumes. and from 3D maps A local map segment corresponding to the specified time period is extracted as a local obstacle avoidance path fragment. The historical driving trajectory is the robot's actual movement path from the moment the path is interrupted until it rejoins the main path or resumes normal passage. This trajectory is usually continuously recorded by the robot's own localization module and is a continuous, timestamped sequence of positions. Additionally, the 3D map is a three-dimensional structured representation containing a wealth of spatial details about the environment, reflecting the spatial distribution of objects, obstacle shapes, and scene surface textures in the robot's working environment. It is typically constructed in the following ways: LiDAR point cloud: provides high-precision spatial geometric contours; Depth camera: Acquires the depth value of each pixel to reconstruct the scene structure; Visual-SLAM system: Integrates camera and IMU pose to generate dense reconstructed maps.

[0051] Each region in the 3D map will contain: 3D coordinate point set (point cloud); Surface normal vectors and texture information; Timestamps or dynamic attributes, such as the frequency of recent updates; Semantic attributes related to local obstacle avoidance.

[0052] Simply put, this step involves analyzing historical driving trajectories. The spatial range traversed is located, and the corresponding spatial sub-blocks in the dense 3D map are identified. Information such as 3D point clouds, surface features, and timestamps contained in the sub-blocks are extracted to form a complete map and spatial segment of the trajectory that corresponds to the actual obstacle avoidance process. This is used for subsequent dynamic risk assessment and passage confidence evaluation.

[0053] S25. After the obstacle avoidance behavior is completed, it is necessary to determine whether the robot can pass through this path again in the future, or whether this path has the reliability of repeated passage. This is determined by dynamic passage confidence. Therefore, firstly, within the obstacle avoidance time window from interruption to resumption of passage mentioned above: Extracting point cloud change sequences across consecutive time frames: ; The start and end times of the obstacle avoidance time window. Extracting the inter-frame image residual map sequence: ;in For time Visual image frames, This represents the inter-frame interval time of the image residual.

[0054] S26. To quantify the aforementioned passage judgment, two core features are extracted using the point cloud change sequence and image residual map sequence collected by the robot during obstacle avoidance: Feature 1: Point cloud stability index The stability of the obstacle's shape is calculated by measuring the shape variability of the obstacle's outline over time. This indicates whether the obstacle's shape is stable in space during obstacle avoidance. If the point cloud outline of an obstacle changes significantly over time, it indicates that the obstacle is dynamic or unstable, such as a person or vehicle; while if the outline remains almost unchanged, it indicates that the obstacle is static, such as a wall or pillar. By comparing the point cloud data of consecutive frames within the obstacle avoidance time window, the degree of shape variation is analyzed, and the stability of the point cloud outline is extracted as an indicator. The calculation is as follows: Extract all consecutive point cloud frames within the obstacle avoidance time window: , For each frame of the point cloud, contour extraction is performed through boundary fitting to obtain the boundary polygons or voxel distribution of obstacles; the IoU change rate, center drift distance, and shape similarity index of the contour shape are calculated between consecutive frames; the average of the above changes is taken as the shape variability of the point cloud; the variability is negatively correlated and normalized to obtain... , A larger value indicates greater stability.

[0055] Feature 2: Intensity of dynamic changes in the image Texture energy fluctuation assessment based on image residual sequences: This method uses residual maps between consecutive image frames to determine the degree of visual scene change. The residual map is a map of the absolute value difference between pixel values ​​in two frames. If the difference is large, it indicates drastic environmental changes, such as occlusions entering or leaving the frame; if the difference is small, it indicates a relatively stable environment. The texture change energy of the residual map over the entire time period is statistically analyzed to obtain an intensity index that measures visual dynamics. The calculation is as follows: Calculate the residual image between every two frames: For each residual image, calculate the energy of the pixel intensity: ; Calculate the mean or fluctuation range of the residual map energy over the entire obstacle avoidance period to obtain This formula represents the image as a W×H two-dimensional pixel matrix. The average of the squared residuals at all pixel positions (x, y) is used to represent the overall intensity of change between two frames. W and H represent the width and height of the image, used for energy normalization. This represents the position coordinates of each pixel in the image.

[0056] Input the two features mentioned above into the pre-trained model: ; For the pre-trained multi-feature fusion confidence evaluation model, the evaluation model learns the following rules: The more stable the point cloud outline and the less unchanged the image, the higher the probability of passage. The greater the point cloud fluctuations and the more frequent the image changes, the greater the risk of passage.

[0057] The evaluation model will consider both inputs and output a probability value Ppass in the interval [0,1]. That is, it outputs the probability that the local obstacle avoidance path segment can be traversed again in the future window, which is the dynamic passability confidence. .

[0058] Specifically, the evaluation model employs a feedforward neural network with two hidden layers, characterized by small parameter count, fast inference, and suitability for edge deployment. The network hierarchy is as follows: The input layer takes two normalized features as input: Point cloud stability index Sc∈[0,1]; The intensity of dynamic changes in the image, Sr∈[0,1]; The feature vector [Sc, Sr] is formed.

[0059] Hidden layer 1: 16 neurons, ReLU activation function, outputting a 16-dimensional intermediate feature vector; Hidden layer 2: 8 neurons, ReLU activation function, output 8-dimensional intermediate feature vector; Output layer: 1 neuron, sigmoid activation function, output dynamic pass confidence. During the model training phase, historical obstacle avoidance data is used for supervised learning, labeled as "whether it successfully passed again" (1 or 0), to optimize the cross-entropy loss function.

[0060] S27, Obstacle Avoidance Semantic Label Generation: The probability value calculated in S26 is simply a numerical value and lacks semantic interpretation capabilities. To make it easier for subsequent modules such as path reconstruction, task scheduling, and risk labeling to use, this value is semantically classified, that is, the probability value is divided into different risk label types, facilitating the use of rule logic or optimization framework calls. Specifically, based on dynamic passage confidence... Generate obstacle avoidance semantic labels based on three preset judgment intervals: like Marked as temporary obstruction; indicates that the current obstacle is likely temporary, such as a person walking by or a temporary object. It is believed that the path may automatically become passable again after a short time, so it can be waited for first and not reconstructed immediately. like Marked as persistently blocked; this indicates that the obstacle may be long-term, and it is believed that the path cannot be restored to traffic in the short term. This path should be avoided and a substantial path reconstruction should be carried out. like If the historical frequency of occurrence is high, it is marked as a high-frequency fluctuation area; this indicates that the passage status of the path is unstable, with frequent switching between passage and blockage, belonging to a dynamic high-risk area, which can be set as low priority or last retry.

[0061] in, The high threshold for the confidence level of passage (the temporary blocking boundary) is set at 0.75-0.85, which means that a temporary block is considered only when the probability of successful passage is greater than 75%-85%. The threshold value for the confidence level of passage (the boundary of continuous blocking) is 0.30-0.40, which means that a continuous blocking is determined only when the probability of passage is less than 30%-40%.

[0062] S3: Based on the set of deviated nodes and the remaining unfinished targets, execute a local reordering algorithm on the task graph to update the access order of the inspected targets, and generate a new optimal path by combining the obstacle avoidance semantic labels and the current pose to guide subsequent inspections.

[0063] S31, Local Optimization Window Definition: During the robot's inspection task, if the original path is interrupted due to obstacles or sudden changes in the path, it is necessary to locally adjust the order of subsequent target visits. This involves replanning the visit order of which target points, i.e., defining a local optimization window, specifically adjusts the part affected by the path interruption, rather than recalculating globally, improving efficiency and preserving existing feasible path segments. Specifically, based on the robot's current position... Offset node set And the target node that was most recently interrupted in the original plan The set of inspection target nodes that have not yet been completed. Together, we construct a local subset of objectives to be optimized: ; This formula represents a subset of inspection targets that need to be optimized locally. It is a set consisting of two parts: all nodes that have deviated from the original path order, i.e., the set of nodes that have deviated from the path order. And all target points that were originally scheduled after the nearest path jump point and have not yet been visited. This represents the subset of inspection targets that needs to be reordered. Indicates the target point The numbering in the original task access order. This indicates the target point visited most recently, i.e., the anchor point after obstacle avoidance.

[0064] For example: Suppose the original access order is: 1→ 2→ 3→ 4→ 5→ 6; The robot has just finished visiting during the execution process. 2. Prepare to go to 3. As a result, the path was interrupted by an obstacle along the way. The nearest feasible target was determined to be... 5. So it jumped over there, at this point 3. 4 belongs to the set of path deviation nodes. 6 is the original plan in The goals after 5 are not yet completed, so the local optimization window should include... 3, 4, 6. Rearrange their access order.

[0065] S32, Local Reordering Optimization: Rearrange the order in which objectives are visited to minimize the total cost of starting from the current position, visiting all local objectives, and finally returning to the destination. In other words, starting from the current position... The path starts at the system's preset endpoint. To find the destination, perform a local path optimization problem with the goal of minimizing the total cost: This formula represents: in all possible access sequences In the process, we need to find an optimal sequence that starts from the robot's current position. Start by visiting all points in the target subset in sequence. ), and finally reach the preset endpoint of the mission. The goal is to minimize the total cost of this entire path. The cost function is: This formula expresses the total cost of the path, which is the weighted cost of each edge in the path, denoted as . The summation here is different from the ordinary path cost. The impact of obstacle avoidance semantic labels is already included; for example, the cost of temporarily blocking a path will be penalized more severely, and persistent blocking may be assigned an infinite value. The cost of each edge segment is also included. To optimize the path order by using a dynamically weighted passage cost adjusted for semantic labels, and to find a better solution within an acceptable timeframe while avoiding getting trapped in local optima, a combination of tabu search and variable neighborhood search algorithms is employed. The process is as follows: Initialization: The initial solution is based on the current path order or the order of the most recent visits; basic neighborhood operation methods are constructed, such as swapping the positions of two targets, inserting or moving a target, etc.

[0066] Variable Neighborhood Search: Dynamically switch neighborhood structures, including using simple operations such as single-point swaps in the first few iterations. If the optimization stalls, switch to larger perturbation operations such as three-point swaps or segment shifts. Then, in each round, starting from the current solution, generate several candidate solutions in the selected neighborhood, evaluate the total cost of these solutions, and select the current best solution as the best solution in the neighborhood.

[0067] Tabu search: A taboo list is set for each perturbation operation, including the most recent N node swaps or insertions, which will be temporarily prohibited to avoid repeated oscillations near local optima; if the new solution is better than the historical best solution, the taboo restrictions are allowed to be broken and the solution is accepted.

[0068] Termination and Output: Set the maximum number of iterations or the number of stalled iterations. The search terminates when there is no significant improvement, and the current optimal path order is output. .

[0069] S33, Obstacle Avoidance Semantic Label Constraint Modeling: When the robot encounters an obstacle during inspection, causing the path to be interrupted, the dynamic passability confidence of each path segment has been obtained by analyzing historical obstacle avoidance behavior, and obstacle avoidance semantic labels have been generated accordingly: continuous blockage, temporary obstruction, high-frequency fluctuation area. Based on these semantic labels, the passability weight or passability of the edges in the task graph will be adjusted to achieve proactive perception and response to obstacle risks in path planning, as detailed below: S331, Handling persistently blocked edges: For all edges marked as persistently blocked regions by semantic tags... In the task map middle: That is, temporarily removing the edge from the graph structure or assigning it an infinite weight to indicate that it is impassable.

[0070] Specifically, for path segments that are determined to be continuously blocked areas, that is, edges with extremely low dynamic passability confidence, they are considered to be impassable in the short term. There are two ways to handle them: one is to assign the edge an infinite weight, indicating that it is extremely discouraged to pass through; the other is to remove the edge directly from the task graph, which is equivalent to disconnecting the connection in the graph structure, ensuring that the search algorithm no longer considers this path. This is a hard constraint, which is to exclude impassable path segments.

[0071] S332, Temporary Blocking and High-Frequency Fluctuation Region Edge Processing: For edges marked as temporary blocking or high-frequency fluctuation regions... Update its weight in the task graph as follows: ;in, For nodes in the original graph arrive The cost of passage, The penalty coefficient is... This represents the dynamically adjusted passage cost after considering semantic labels, based on the semantic label type of the path segment and its corresponding confidence level. Definition; The penalty coefficient is defined as: ; in, For the edge The dynamic passability confidence of the path segment indicates that the penalty for temporary obstruction depends on the passability confidence of that path segment; the higher the uncertainty, the higher the cost. This represents the variance statistics of historical travel states in high-frequency fluctuation regions, indicating that for high-frequency fluctuation regions, the degree of penalty depends on the severity of the fluctuations in the historical states of the path. It is a non-linear mapping function called sigmoid.

[0072] Specifically, for path segments with a medium to high probability of passage, i.e. temporary obstructions and high-frequency fluctuation areas, instead of directly blocking them, their cost is increased through dynamic penalties, causing them to be downgraded in path planning, but they can still be used when necessary. The original cost of these edges is multiplied by a dynamic penalty coefficient, which varies with the semantic label type and confidence value. This is a soft constraint that reflects the tolerance control of path passage.

[0073] In the above formula This is the dynamic penalty amplification factor. The value range is [0.5, 2.0]. If the path can be restored after a short stop in the scenario, such as when people or cars pass by temporarily in the factory area, a smaller value of 0.5 to 1.0 can be selected. If delayed recovery is not tolerated, or if forced avoidance is required, a larger value of 1.5 to 2.0 can be selected. The principle is that the tighter the system response time and the shorter the task time window, the larger the value should be. The value range is [1.0, 3.0]. If the robot path has strong fault tolerance, you can choose 1.0 to 1.5. If the scene has dense mixed traffic of people and vehicles and large traffic fluctuations, you can choose 2.0 to 3.0.

[0074] S34, based on the updated graph structure after incorporating semantic label adjustments, formally generates a new, practically navigable, complete inspection path to replace the original interrupted path and guide the robot to continue performing the inspection task. We have completed this: S31: Identify which target nodes need to be reordered; S32: A new access order has been optimized. ; S33: The edges in the path were dynamically weighted based on their semantic labels; Now, having obtained the robot's current location, a new target access order, and an updated task map, we recalculate the path segments based on this information and piece them together to form a complete new inspection path.

[0075] Specifically, based on the aforementioned dynamic weighted graph , at current position Starting with the A* heuristic graph search algorithm, the updated target access order is determined. Calculate the optimal collision-free path segments connecting each target in sequence: This formula means: each term This represents the optimal path segment calculated from point a to point b. These path segments are calculated based on the adjusted graph (obstacle avoidance semantic labels are now in effect). The plus sign indicates that the path segments are sequentially concatenated to form the final path, which is then combined into a new complete inspection path.

[0076] use The specific process of the heuristic search algorithm generating paths segment by segment is as follows: 1. From the robot's current position Set off.

[0077] 2. Calculate in segments according to the new order: The new target access order is... ;use The algorithm computes path segments sequentially within the updated task graph structure: From current location to ; ; ...... from To the preset destination ; 3. By piecing together all the above path segments in sequence, a complete collision-free path is obtained. This is for the robot to travel along the route later.

[0078] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0079] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for inspection path planning based on the coordination of navigation path and obstacle avoidance strategy, characterized in that, Includes the following steps: S1, based on the preset inspection target set and inspection environment map, uses a graph search algorithm to generate an initial navigation path and constructs a task graph structure with each inspection target as a node and the navigation path segment as an edge, which is used to represent the path connectivity and task execution order; S2, during the inspection process, based on the robot's body sensors, obstacle avoidance behavior events are captured in real time, and it is determined whether the current navigation path segment is interrupted. If it is interrupted, the shortest path segment from the current position to the nearest inspection target is calculated, a path deviation map is constructed, and a set of deviation nodes is output. Based on the historical backtracking trajectory of the navigation path segment interruption location, the corresponding local obstacle avoidance path segments are extracted from the 3D map. Combined with the point cloud change sequence and image residual map constructed during the obstacle avoidance process, the dynamic passability confidence of the local obstacle avoidance path segments is calculated. Obstacle avoidance semantic labels are generated based on the dynamic passability confidence, including temporary obstruction, continuous blockade and high-frequency fluctuation area, to provide semantic-level obstacle avoidance area constraints for subsequent path reconstruction. S3: Based on the set of deviated nodes and the remaining unfinished targets, execute a local reordering algorithm on the task graph to update the access order of the inspected targets, and generate a new optimal path by combining the obstacle avoidance semantic label and the current pose to guide subsequent inspections.

2. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 1, characterized in that, The graph search algorithm adopts the A* heuristic graph search algorithm, which uses the weighted sum of path distance and estimated travel cost as the cost function to perform a global path search between each inspection target and generate an initial navigation path set connecting each inspection target.

3. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 2, characterized in that, The construction of the task graph structure includes an undirected weighted graph with all inspection targets as nodes and the initial navigation path segment connecting any two targets as edges, which serves as the task graph structure. The weight of each edge is the actual passage cost of its corresponding navigation path segment. Based on the task graph structure, the minimum spanning tree algorithm is used to generate an optimal or near-optimal access sequence covering all nodes, thereby determining the initial inspection task execution order.

4. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 1, characterized in that, The robot's sensors include lidar, ultrasonic, or vision sensors. During the inspection process, the lidar, ultrasonic, or vision sensors on the robot body detect signals in real time. When an obstacle is detected to have entered the safe distance threshold of the current navigation path segment or the proportion of the path ahead being occupied exceeds a preset threshold, it is determined that the current navigation path segment has been interrupted. If the path is interrupted, the robot's current position is taken as the starting point, and all unfinished inspection targets in the task graph structure are taken as potential endpoints. On the environmental map subgraph after removing occupied edges, the shortest path algorithm is used to calculate the shortest path to each potential endpoint, and the path with the smallest total cost is selected. The inspection target corresponding to this path is recorded as the nearest inspection target. The shortest path is the shortest path segment.

5. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 4, characterized in that, The construction of the path deviation graph includes taking the most recently inspected target as the root node, taking subsequent inspected targets that could not be visited in the original order due to path interruption as deviation nodes, forming a directed weighted graph with a single source pointing to multiple unfinished targets, recording the topological relationship and cost increment between each deviation node and the original planned access sequence, and forming the deviation node set.

6. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 5, characterized in that, S2 further includes, for the location where the navigation path segment is interrupted, retracing the historical trajectory during the robot's obstacle avoidance behavior, extracting the actual driving trajectory used to bypass the obstacle from the pre-constructed 3D map, as a local obstacle avoidance path segment; associating the time period of the obstacle avoidance process, extracting continuous frame data of the LiDAR point cloud in this area to form a point cloud change sequence, and extracting the inter-frame image residual map of the visual sensor in the same area at different times; based on the point cloud change sequence, statistically analyzing the stability index of the obstacle contour; based on the image residual map, analyzing the dynamic change intensity of the texture area; fusing the stability index and the dynamic change intensity, and using a pre-trained neural network evaluation model, outputting the probability value that the local obstacle avoidance path segment can be traversed again in the future time window, as the dynamic traversal confidence.

7. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 6, characterized in that, Generate obstacle avoidance semantic labels based on the predefined interval to which the dynamic access confidence level belongs: When the confidence level is higher than the first threshold, it is marked as a temporary block. When the confidence level is below the second threshold, it is marked as a persistent block. When the confidence level is in the fluctuation range and the historical labeling frequency is high, it is marked as a high-frequency fluctuation region; the semantic label serves as a semantic-level obstacle avoidance region constraint condition for subsequent path reconstruction.

8. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 1, characterized in that, The execution of the local reordering algorithm includes: Based on the set of deviated nodes and all remaining unfinished inspection targets, a local optimization window is defined on the task graph structure. The definition of the local optimization window includes taking the robot's current position as the starting point and including the nodes in the set of deviated nodes and the nodes that were originally located after the most recently interrupted inspection target and have not yet been visited into the target subset to be reordered. With the goal of minimizing the estimated total cost of starting from the current position, visiting all nodes in the target subset, and finally returning to the preset destination, a tabu search algorithm with variable neighborhood search is used to solve the problem on the task graph structure, and the updated inspection target access order is output. Among them, the calculation of the estimated total cost of the path is combined with the obstacle avoidance semantic label to dynamically weight and adjust the passage cost of the relevant edges.

9. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 8, characterized in that, Based on the updated inspection target access order and the constraints provided by the obstacle avoidance semantic tags, a new optimal path is generated, including the following weight update rules: For areas marked as continuously blocked, the corresponding path segments are temporarily set to impassable status in the environment map, and are temporarily removed or given infinite weight in the task graph structure. For edges marked as temporary obstructions or high-frequency fluctuation regions, their weights in the task graph structure are updated to the product of the original passage cost and a dynamic penalty coefficient based on label type and confidence level.

10. The inspection path planning method based on the coordination of navigation path and obstacle avoidance strategy according to claim 9, characterized in that, S3 further includes taking the robot's current pose as the starting point of the path, running the graph search algorithm again on the task graph structure after applying the weight update rules, calculating the collision-free path of each inspection target point in the updated sequence segment by segment, and finally synthesizing a complete new global optimal path to guide the robot's subsequent inspection navigation.

Images