A method and system for scheduling robot inspection in a park

Abstract

The application provides a kind of park robot inspection scheduling method and system, comprising: obtaining real-time environmental perception data and inputting obstacle identification model, determine temporary structured obstacle not contained in pre-constructed static map;Extract the closed boundary coordinates of the obstacle to generate independent forbidden area data;The forbidden area data is mounted as an independent top layer temporary layer to the topmost layer of the layered map in an incremental manner, forming a top layer temporary forbidden area layer, and setting its coverage area as a non-passable core obstacle area;According to the obstacle type query mapping table, obtain the inflation distance, combine the inspection task safety level weighted adjustment to obtain the target inflation distance, and expand outward based on the core obstacle area to define the feasible area for detouring;Generate multiple obstacle avoidance detour trajectories in the area, evaluate and select the optimal trajectory through the multi-index cost function, and control the robot to perform inspection scheduling.The application can accurately avoid obstacles when temporary obstacles appear.

Description

Technical Field

[0001] This application relates to the field of robot scheduling technology, and in particular to a robot inspection and scheduling method and system for a park. Background Technology

[0002] When the park security patrol robot performs its pre-set routine patrol tasks, the underlying grid map must continuously reflect the actual traffic conditions of the park's physical environment to ensure the feasibility of the planned path and the continuity of the patrol task. When large-scale structured temporary obstacles such as fences, scaffolding, building material stacks, or road excavations are added to the park, the area changes from a passable to an impassable area. If the robot continues to travel along the original planned path, it will directly face the risk of path obstruction, partial obstacle avoidance failure, or even shutdown and paralysis, resulting in security blind spots in the patrol area. Safety hazards such as blocked fire lanes and unauthorized personnel climbing over them cannot be detected and dealt with in a timely manner. Therefore, when temporary structured obstacles appear in the park that exceed the coverage of the static map, the robot's patrol path must be adaptively adjusted to bypass the impassable area, ensuring the continuity of the patrol task and the integrity of security coverage.

[0003] Existing path planning technologies for robot obstacle avoidance are mainly divided into two categories. One category is real-time obstacle avoidance methods based on local perception. The robot relies on its onboard LiDAR or visual sensors to detect obstacles in front of it in real time and generate detour actions within a local area. However, this type of method lacks awareness of the global spatial range of obstacles. When faced with large-area structured obstacles such as fences, it is prone to repeatedly triggering local obstacle avoidance or even getting stuck, and is unable to generate a large-scale detour trajectory that crosses the entire obstacle area, resulting in broken inspection paths. Another type of global planning method based on static maps involves manually marking obstacle areas or writing them into a raster map in real time before performing global path replanning. However, existing methods typically only binarize the perceived obstacle areas into impassable raster cells without performing fine semantic differentiation of obstacle types, boundary morphology, and risk levels. This results in the generated inflated no-go zones being either too conservative or too aggressive. At the same time, existing methods mostly use fixed inflation parameters, which cannot dynamically adjust the detour safety margin according to the differences in spatial hazard boundaries of different types of temporary obstacles and the safety requirements of the current inspection task. Therefore, existing obstacle avoidance path planning methods still have significant deficiencies in terms of detour accuracy, safety, and engineering usability for large structured temporary obstacles in the park. Summary of the Invention

[0004] The main objective of this application is to propose a robot inspection and scheduling method and system for a park, which can accurately avoid temporary obstacles.

[0005] To achieve the above objectives, a first aspect of this application proposes a robot inspection and scheduling method for a park, comprising: Acquire real-time environmental perception data during the robot's inspection process; The real-time environmental perception data is input into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in the pre-built static map. Obtain the closed boundary coordinates of the temporary structured obstacle, map the closed boundary coordinates to the map coordinate system and convert them into grid cells to generate independent no-patrol area data; The data of the restricted patrol area is incrementally overlaid and mounted onto the top layer of the pre-constructed layered map as an independent top-level temporary layer to form a top-level temporary restricted patrol layer. The area covered by the top-level temporary no-patrol layer is designated as an impassable core obstacle zone; Based on the type of the temporary structured obstacle, a preset mapping table is queried to obtain the corresponding expansion distance. The expansion distance is then weighted and adjusted according to the safety level of the robot's current inspection task to obtain the target expansion distance. The target expansion distance is then expanded outward from the core obstacle area to obtain an impassable expansion restricted area that includes the core obstacle area. The passable area outside the expansion restricted area is defined as a detourable area. Within the detourable area, several obstacle avoidance and detour inspection candidate trajectories are generated. A multi-index cost function is used to evaluate the cost of each obstacle avoidance and detour inspection candidate trajectory in order to select the optimal target detour inspection trajectory. The robot is controlled to perform inspections and scheduling within the park by following the target's inspection trajectory.

[0006] In some embodiments, inputting the real-time environmental perception data into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in a pre-built static map includes: Retrieve environmental images captured by fixed cameras in the park and fuse these images with real-time environmental perception data to obtain fused perception data. The fused perception data is input into a pre-trained obstacle recognition model to identify candidate obstacles in the environment and the spatial boundaries corresponding to the candidate obstacles. The spatial boundaries of each candidate obstacle are matched with the pre-built static map, and candidate obstacles that overlap with existing obstacles in the pre-built static map are eliminated to obtain the first temporary obstacle. Based on the spatial boundaries of each of the first temporary obstacles, the transient probability is calculated, and obstacles with a transient probability lower than a first preset threshold are retained as second temporary obstacles. Predict the duration of each of the second temporary obstacles, and filter out short-term road-occupying obstacles whose duration is lower than a second preset threshold to obtain the third temporary obstacle; When the spatial boundary offset of the third temporary obstacle in consecutive frames of perception data is less than a third preset threshold and the morphological overlap is greater than a fourth preset threshold, and it has a closed and fitable rigid boundary, it is determined to be a temporary structured obstacle.

[0007] In some embodiments, mapping the coordinates of the closed boundary to a map coordinate system and converting them into raster cells to generate independent no-patrol area data specifically involves: The coordinates of the closed boundary of the temporary structured obstacle are transformed from the sensor coordinate system used for real-time perception to the global map coordinate system of the layered map to obtain the internal region of the closed boundary in the global coordinate system. Based on the preset grid resolution of the layered map, the closed boundary of the inner region of the closed boundary is discretized to obtain a boundary grid sequence. Based on the boundary grid sequence, determine all internal grid cells located within the region inside the closed boundary. All internal grid cells are marked as occupied, thereby generating the no-patrol area data.

[0008] In some embodiments, the step of incrementally overlaying and mounting the restricted patrol area data as an independent top-level temporary layer onto the outermost layer of a pre-constructed layered map to form a top-level temporary restricted patrol layer includes: Create an independent temporary raster layer based on the restricted patrol area data, and mark the raster cells corresponding to the restricted patrol area data as occupied in the temporary raster layer; The temporary raster layer is used as the top overlay layer and incrementally overlaid on the bottom permanent map of the layered map to obtain the overlay result; In the overlay result, the temporary raster layer is set to have a higher priority than the underlying permanent map to obtain the target overlay result; The top-level temporary no-patrol layer is determined based on the target overlay result.

[0009] In some embodiments, querying a preset mapping table to obtain the corresponding expansion distance based on the type of the temporary structured obstacle includes: The risk classification label is determined according to the type of the temporary structured obstacle, wherein the risk classification label includes at least the high-risk prohibited patrol category and the boundary patrol category; Based on the risk classification label, a matching process is performed in the mapping table to determine the corresponding inflation distance.

[0010] In some embodiments, the step of weighting and adjusting the expansion distance based on the safety level of the robot's current inspection task to obtain the target expansion distance specifically involves: Obtain the task safety level identifier of the robot's current inspection task, and determine the task weighting coefficient based on the task safety level identifier; The target expansion distance is obtained by multiplying the expansion distance by the task weighting coefficient.

[0011] In some embodiments, expanding the target expansion distance outward from the core obstacle zone to obtain an impassable expanded restricted zone including the core obstacle zone specifically involves: Convert the target inflation distance into the number of inflation grids on the grid map; Using the grid set of the core obstacle area as the initial region and the number of expanded grids as the expansion radius, a morphological expansion operation is performed outward to obtain the expanded region. The entire expanded region is marked as an impassable expanded restricted area, including the core obstacle zone.

[0012] In some embodiments, generating a plurality of obstacle avoidance and detour inspection candidate trajectories within the feasible detour area includes: Based on the robot's current direction of travel, determine the left and right bypass boundaries of the expansion restricted area; Within the feasible detour area, several path key control points are selected along the left detour boundary and / or the right detour boundary according to a preset sampling step size, and multiple initial obstacle avoidance detour paths are generated based on the path key control points. Each of the initial obstacle avoidance detour paths starts from the original inspection path upstream of the obstacle, bypasses the obstacle, and returns to the original inspection path downstream. The detour direction or the position of the path key control points of each of the initial obstacle avoidance detour paths are not exactly the same. Each of the initial obstacle avoidance paths is smoothed and optimized to generate a smoothed candidate trajectory corresponding to each initial obstacle avoidance path; For each of the smooth candidate trajectories, a curvature check is performed, and smooth candidate trajectories with a curvature radius smaller than the robot's minimum turning radius are eliminated, while the remaining smooth candidate trajectories are retained as the obstacle avoidance and detour inspection candidate trajectories.

[0013] In some embodiments, the step of evaluating the cost of each of the obstacle avoidance and detour inspection candidate trajectories using a multi-index cost function to select the optimal target detour inspection trajectory includes: For each of the obstacle avoidance and detour inspection candidate trajectories, calculate the minimum safe distance between each of the obstacle avoidance and detour inspection candidate trajectories and the boundary of the expansion restricted area, and determine the safety cost component based on the minimum safe distance; Calculate the curvature value of each of the aforementioned obstacle avoidance and bypass inspection candidate trajectories at each path point, and determine the smoothness cost component based on the curvature value; Calculate the total path length of each of the obstacle avoidance and detour inspection candidate trajectories, and determine the efficiency cost component based on the total path length; Obtain the preset passage cost weights of each grid cell traversed by each of the proposed obstacle avoidance and detour inspection candidate trajectories, and sum the preset passage cost weights of each grid cell to obtain the total passage cost component; The comprehensive cost of each candidate trajectory is obtained by weighted summing of the security cost component, the smoothness cost component, the efficiency cost component, and the total passage cost component. The target bypass inspection trajectory is determined based on the comprehensive cost value.

[0014] To achieve the above objectives, a second aspect of this application proposes a robot inspection and scheduling system for a park, the system comprising: The first module is used to acquire real-time environmental perception data during the robot's inspection process; The second module is used to input the real-time environmental perception data into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in the pre-built static map. The third module is used to obtain the closed boundary coordinates of the temporary structured obstacle, map the closed boundary coordinates to the map coordinate system and convert them into grid cells to generate independent no-patrol area data; The fourth module is used to incrementally overlay and mount the restricted patrol area data as an independent top-level temporary layer onto the outermost layer of the pre-constructed layered map to form a top-level temporary restricted patrol layer. The fifth module is used to set the area covered by the top-level temporary no-patrol layer as an impassable core obstacle zone; The sixth module is used to query a preset mapping table based on the type of the temporary structured obstacle to obtain the corresponding expansion distance, and to adjust the expansion distance by weighting it in combination with the safety level of the robot's current inspection task to obtain the target expansion distance. The target expansion distance is then expanded outward from the core obstacle area to obtain an impassable expansion restricted area that includes the core obstacle area. The passable area outside the expansion restricted area is defined as a detourable area. Within the detourable area, several obstacle avoidance and detour inspection candidate trajectories are generated. A multi-index cost function is used to evaluate the cost of each obstacle avoidance and detour inspection candidate trajectory in order to select the optimal target detour inspection trajectory. The seventh module is used to control the robot to perform inspections and scheduling within the park according to the target's inspection trajectory.

[0015] This application acquires real-time environmental perception data and inputs it into an obstacle recognition model, enabling the system to accurately identify temporary structured obstacles and their closed boundaries not recorded in static maps within complex park environments. This provides accurate spatial information for subsequent obstacle avoidance operations. The closed boundary coordinates are then mapped and converted into grid cells to generate independent no-go zone data, which is incrementally mounted as a top-level temporary layer onto the outermost layer of the layered map. This allows temporary obstacle information to be quickly injected into the map system without affecting the stability of the underlying permanent map, while avoiding destructive modifications to the global map and ensuring the accuracy of map data operations. Furthermore, the area covered by the top-level temporary no-go layer is set as an impassable core obstacle zone, clearly defining the absolute no-entry range and establishing rigid spatial constraints for trajectory generation. Finally, the basic expansion distance is obtained by querying a mapping table based on the type of temporary structured obstacle and combined with the robot's current inspection data. The safety level of the task is used to weight and adjust the expansion distance to obtain the target expansion distance. Then, based on the core obstacle area, the expansion extends outward to form an impassable expansion no-go zone. This mechanism allows the size of the expansion no-go zone to adaptively adjust according to the degree of danger reflected by the obstacle type and the dynamic changes in task safety requirements. This avoids the problems of overly conservative or under-aggressive detours caused by fixed expansion parameters, ensuring accurate adaptation of obstacle avoidance safety margin. Subsequently, a feasible detour area is delineated outside the expansion no-go zone, and several candidate trajectories are generated. A multi-index cost function is used to comprehensively evaluate the safety, smoothness, and passage efficiency of each candidate trajectory, and the optimal target detour trajectory is selected from them. This ensures that the final detour path is not only safe and away from obstacles and kinematically feasible, but also has the optimal path cost. Finally, the robot is controlled to perform inspections along the target trajectory, realizing a closed-loop process from perception, mapping, no-go zone delineation, expansion optimization to trajectory generation and execution. Attached Figure Description

[0016] Figure 1 This is a flowchart of a robot inspection and scheduling method for a park provided in an embodiment of this application; Figure 2 This is a flowchart of steps S201 to S206 provided in this application; Figure 3 This is a flowchart of steps S301 to S304 provided in this application; Figure 4 This is a flowchart of steps S401 to S404 provided in this application; Figure 5 This is a flowchart of steps S501 to S506 provided in this application; Figure 6 This is a schematic diagram of the structure of the robot inspection and scheduling system for the park provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] This application provides a method and system for scheduling robot inspections in a park, which aims to accurately avoid temporary obstacles.

[0021] The robot inspection and scheduling method and system for parks provided in this application are specifically described through the following embodiments. First, the robot inspection and scheduling method for parks in this application embodiment is described.

[0022] The robot inspection and scheduling method for a park provided in this application relates to the field of equipment control technology. The robot inspection and scheduling system for a park provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can implement the robot inspection and scheduling method for the park, but is not limited to the above forms.

[0023] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0024] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user will be obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent will the necessary user-related data for the normal operation of the embodiments of this application be obtained.

[0025] Figure 1 This is an optional flowchart of the robot inspection and scheduling method for a park provided in the embodiments of this application, which may include, but is not limited to, steps S101 to S107.

[0026] Step S101: Obtain real-time environmental perception data during the robot's inspection process; In some embodiments, as the inspection robot travels along a preset route, the lidar mounted on the top of the vehicle emits laser beams to the surrounding environment at a scanning frequency of 10Hz. By measuring the time difference between the laser beam's emission and return, the distance between each scanning point and the center of the lidar is calculated. Combined with the current scanning angle recorded by the encoder inside the lidar, an original three-dimensional spatial point cloud frame covering a horizontal 360° and a vertical field of view of 30° is generated.

[0027] After receiving the raw point cloud frame, the airborne edge computing module first converts the polar coordinates (distance, angle) of each scan point into Cartesian coordinates (x, y, z) in the sensor coordinate system according to the factory calibration parameters of the LiDAR, thus obtaining the raw point cloud set. Then, the three-dimensional space is divided into a cubic voxel grid with a side length of 0.1 meters. Only the point closest to the voxel center is retained in each grid, and other redundant points are removed. While preserving the main geometric structure of the environment, the point cloud data volume is compressed to 30%~50% of the original data. The downsampled point cloud then enters the statistical outlier filtering stage: the average Euclidean distance between each point and its 50 neighboring points is calculated. If this distance exceeds twice the standard deviation of the average distance of all points, it is identified as an outlier and removed.

[0028] The effective environmental point cloud data output after the above preprocessing, together with the timestamp information of the corresponding frame, constitutes the real-time environmental perception data of the current scanning cycle.

[0029] Step S102: Input real-time environmental perception data into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in the pre-built static map; Please refer to Figure 2 In some embodiments, step S102 includes steps S201 to S206; Step S201: Retrieve environmental images captured by fixed cameras in the park, and fuse the environmental images with real-time environmental perception data to obtain fused perception data; In some embodiments, the edge computing module uses the timestamp of the robot's current global positioning pose as a reference and sends image retrieval commands to fixed cameras within a 50-meter radius of the robot's current location via the park's local area network. This retrieves environmental images of the park roads captured by each camera within a time window of ±1 second before and after the current moment. Then, time alignment is performed based on the timestamp, discarding data frames with timestamp deviations exceeding ±50ms. The environmental images from multiple fixed cameras are then fused with real-time environmental perception data to form fused perception data that simultaneously contains color and texture information (from the cameras) and precise distance information (from the LiDAR). Each fused point carries three-dimensional coordinates, RGB color values, source sensor identifiers, and timestamp information.

[0030] It should be noted that each fusion point in the fused sensing data carries three-dimensional coordinates, RGB color values, source sensor identifiers, and timestamps.

[0031] Step S202: Input the fused perception data into the pre-trained obstacle recognition model to identify candidate obstacles in the environment and the spatial boundaries corresponding to the candidate obstacles. In some embodiments, fused sensing data is input into an obstacle recognition model pre-deployed in an edge computing module to complete the recognition of candidate obstacles and boundary output. The obstacle recognition model adopts an encoder-decoder architecture and is pre-trained based on labeled samples collected from historical inspections of the park: the encoder part is based on an improved PointNet++ network, extracting geometric distribution features (including normal vector, curvature, and local density) and color statistical features (including RGB mean and variance) within a local neighborhood of 0.3 meters for each point in the fused sensing data, and encoding them into a 512-dimensional feature vector; the decoder part upsamples the feature vector to the original point cloud resolution through three layers of deconvolution operation, outputs the probability distribution of each point belonging to a preset category such as "color steel fence", "scaffolding", "excavated road surface", "building material stockpile", "pedestrian", "vehicle", "tree", "wall", etc., and regresses the offset of each point relative to its target center.

[0032] The obstacle recognition model performs Euclidean distance clustering based on the category prediction results of each point, with a threshold of 0.2 meters. Points that belong to the same target and are spatially adjacent are aggregated into a candidate obstacle instance. The minimum bounding cube of the instance is output as its three-dimensional spatial boundary. At the same time, the semantic recognition confidence score in the range of 0 to 1 is output. Low-confidence candidate targets with scores below 0.6 are filtered out. Finally, a list of high-confidence candidate obstacles is output, with each item in the list corresponding to a candidate obstacle and its spatial boundary.

[0033] It should be noted that the obstacle recognition model training phase uses laser-image fusion point cloud samples accumulated from historical inspections of the park to construct the dataset. The samples cover all preset categories, including corrugated steel fences, scaffolding, excavated roads, and building material piles, with no less than 2000 effective samples per category. Point-by-point semantic annotation is used to label each point with its category label, and the center coordinates of each obstacle instance are also labeled as the ground truth for offset regression. The model training employs a multi-task joint loss function, with the total loss consisting of a weighted average of the classification loss and the center offset regression loss: the classification loss uses FocalLoss to alleviate the class imbalance between foreground and background in the point cloud samples; the center offset regression loss uses a smoothed L1 loss to constrain the accuracy of the point's offset prediction from the target center. The weights for the two types of losses are 0.7 and 0.3, respectively. The training optimizer uses AdamW, with an initial learning rate of 0.001, a batch size of 16, and a total of 80 training rounds. The learning rate decays to 0.5 times its original value every 20 rounds. When the classification accuracy on the validation set improves by less than 0.1% for 10 consecutive rounds, an early stopping mechanism is triggered to terminate training and save the optimal weights.

[0034] Step S203: Match the spatial boundaries of each candidate obstacle with the pre-built static map, and remove candidate obstacles that overlap with existing obstacles in the pre-built static map to obtain the first temporary obstacle; In some embodiments, a pre-built static map, pre-built offline, is loaded from local storage. This map is a raster-based representation of the projected outlines of all permanent buildings, walls, lampposts, fire hydrants, curbs along fixed inspection routes, and evergreen trees within the park.

[0035] For each candidate obstacle, its 3D spatial boundary is projected onto a horizontal plane as a polygon. All grid cells within the polygon's coverage area are traversed, and each grid cell is checked to see if it has been marked as occupied in the pre-built static map. The proportion of occupied grid cells to the total number of grid cells in the candidate obstacle's projection area is counted. If the proportion exceeds 20%, the candidate obstacle is determined to have a high degree of overlap with existing obstacles in the static map, is marked as an existing static object, and is removed from the candidate list. If the proportion does not exceed 20%, it is determined to be a newly added object not recorded in the static map and is retained.

[0036] After being compared and eliminated one by one, the remaining candidate obstacles constitute the first temporary obstacles. Each first temporary obstacle still carries the original spatial boundary coordinates and semantic category label.

[0037] It should be clarified that the elimination operation here only applies to the global map update candidate queue. Eliminated candidate obstacles will continue to participate in the robot's local real-time obstacle avoidance planning and dynamic detour, but will no longer be included as temporary structured obstacles in the subsequent map update process.

[0038] Step S204: Calculate the transient probability based on the spatial boundary of each first temporary obstacle, and retain obstacles with transient probabilities lower than the first preset threshold as second temporary obstacles; In some embodiments, for each first temporary obstacle, the transient probability calculation edge computing module deployed on the edge gateway is invoked to calculate the transient probability value e in the range of 0 to 1 for the obstacle.

[0039] The calculation of the instantaneous probability value e requires the integration of four types of features: first, spatial location stability features; second, morphological contour consistency features; third, semantic category prior features; and fourth, environmental interference compensation features. These four types of features are then linearly weighted and summed (the weights are considered configurable parameters, with default values ​​of 0.35 for spatial location, 0.30 for morphological consistency, 0.25 for semantic prior, and 0.10 for environmental compensation) to obtain the instantaneous probability e of the obstacle.

[0040] Next, the edge computing module compares the calculated e-value with a preset first threshold (0.3 in this embodiment): if e ≥ 0.3, the obstacle is determined to have obvious instantaneous dynamic attributes (such as pedestrians or temporarily parked vehicles), and is classified as an unstable object and filtered out; if e < 0.3, the obstacle is determined to have stable static attributes and is retained to proceed to the next step. All objects retained after filtering constitute the "second temporary obstacle" list.

[0041] It should be noted that the spatial position stability feature is obtained by taking the position coordinates of the obstacle's core in the global coordinate system from 50 frames (corresponding to approximately 5-second time windows) of perception data continuously collected at the current 10Hz sampling frequency, and calculating its standard deviation. The larger the standard deviation, the more unstable the position, and the higher the instantaneous probability weight. The morphological contour consistency feature is obtained by calculating the IoU overlap of the obstacle's 3D point cloud and the overlap area ratio of the 2D projected contour in the current frame and the previous frame, and taking the weighted average of the two. The semantic category prior feature is obtained by reading the corresponding prior transient probability from the preset prior table based on the semantic category output in step S202 (e.g., 0.9 for "pedestrian", 0.6 for "vehicle", and 0.1 for "fence"). The environmental interference compensation feature is obtained by multiplying the calculated original transient probability by a compensation coefficient of 1.5 when the pose covariance output by the robot's localization edge calculation module exceeds 0.3 meters (if the value exceeds 1 after compensation, it is truncated to 1).

[0042] It should be noted that among the four types of features, spatial location stability, morphological contour consistency, and semantic category prior are additive features. These are all first normalized to the [0,1] interval using a preset mapping rule, and then weighted and summed according to their corresponding weights to obtain the basic instantaneous probability. Environmental interference compensation is a multiplicative adjustment coefficient and does not participate in the weighted summation. Specifically, the normalization rules are as follows: the spatial location stability feature uses the location standard deviation σ as input and employs piecewise linear mapping. When σ≤0.05 meters, the corresponding component value is 0; when σ≥1.0 meters, the corresponding component value is 1; and linear interpolation is used in the intermediate interval. The morphological contour consistency feature directly uses the complement of the three-dimensional IoU overlap (1-IoU) as the component value; the higher the overlap, the lower the component value. The semantic category prior feature directly reads the 0~1 prior values ​​from the preset prior table as component values. After the three types of components are weighted and summed according to the default weights to obtain the basic instantaneous probability, if the positioning pose covariance exceeds 0.3 meters, it is multiplied by 1.5 times the environmental compensation coefficient. If the value exceeds 1 after compensation, it is truncated to 1, and finally the instantaneous probability e is obtained.

[0043] Step S205: Predict the duration of each second temporary obstacle and filter out short-term road-occupying obstacles with a duration lower than the second preset threshold to obtain the third temporary obstacle; In some embodiments, for each second temporary obstacle, a pre-built offline categorized survival curve library is invoked according to its semantic category, and the survival duration is predicted using the survival function S(t)=P(T>t) of the corresponding category; where T is the total survival duration of the obstacle from its first observation to its disappearance, and S(t) is the probability that the total survival duration exceeds time t. The subsequent survival duration of the obstacle is calculated by substituting the time point when the obstacle is first continuously and stably observed into the current time.

[0044] Next, the predicted duration is compared with a second preset threshold (4 hours in this embodiment): if the predicted duration is < 4 hours, it is determined to be a short-term temporary road occupancy event (such as temporary unloading of express delivery vehicles or short-term equipment stacking), which only triggers local dynamic obstacle avoidance and is not included in the global map management process, and is therefore filtered out; if the predicted duration is ≥ 4 hours, it is determined to be a temporary obstacle with long-term impact and is retained. The objects retained after filtering constitute the third temporary obstacle list.

[0045] It should be noted that the survival curve library collects historical inspection records and construction records of the park over the past 12 months through offline phases. For each type of obstacle, the survival time distribution parameters are pre-constructed by fitting the full cycle of appearance and disappearance using maximum likelihood estimation. For example, after fitting the generalized Gamma distribution for the color steel fence type, the median survival time is about 30 days, and the median survival time for the scaffolding type is about 60 days.

[0046] It should be noted that the obstacle has been observed for a period of time. Based on the conditional survival formula S(t|T> )=S(t) / S( Solve for S(t|T>). The time corresponding to )=0.5 ,Will The difference between the current time and the predicted duration of the obstacle is used as the predicted duration of its existence. This value represents the duration of the obstacle's existence as observed in the past. Given time constraints, the obstacle has a 50% probability of remaining in existence for a certain period of time, demonstrating engineering robustness and effectively mitigating the interference of extreme values ​​on the prediction results.

[0047] Step S206: When the spatial boundary offset of the third temporary obstacle in consecutive multi-frame perception data is less than the third preset threshold and the morphological overlap is greater than the fourth preset threshold, and it has a closed and fitable rigid boundary, it is determined to be a temporary structured obstacle.

[0048] In some embodiments, for each third temporary obstacle, the edge computing module continues to perform continuous multi-cycle stability verification. Specifically, it retrieves the perception records of the third temporary obstacle in the current inspection cycle and the previous two inspection cycles, and extracts the spatial boundary center coordinates and point cloud set of each record. First, it calculates the maximum Euclidean distance offset between the spatial boundary centers in the three records. If the maximum Euclidean distance offset exceeds a third preset threshold (0.2 meters in this embodiment), it indicates that the obstacle position is not yet stable and is not included in the map update. If the maximum Euclidean distance offset does not exceed 0.2 meters, it calculates the 3D IoU overlap between the obstacle point cloud sets of two adjacent records, and takes the minimum value of the IoU between adjacent frames in the three records. If the minimum value does not exceed a fourth preset threshold (0.8 in this embodiment), it indicates that the obstacle shape is still changing and is not included in the map update. If the minimum value is greater than 0.8, it indicates that the obstacle's spatial boundary and shape are stable. Subsequently, the module performs RANSAC-based planar fitting and corner detection on the point cloud boundary of the obstacle, extracts the projected contour of the point cloud on the horizontal plane, searches for corners along the contour direction and connects them to form edges. If the closed contour has no less than 3 edges, all edges are connected end to end, and the overall fitting residual is less than 0.05 meters, it is determined to have a "closed and fitable rigid boundary". The third temporary obstacle that meets all three conditions (boundary offset ≤ 0.2 meters, morphological IoU > 0.8, and has a closed rigid boundary) is finally identified as a "temporary structured obstacle".

[0049] It should be noted that an inspection cycle is defined as the time required for the robot to complete one full preset route, which is approximately 30 minutes in this embodiment.

[0050] Step S103: Obtain the coordinates of the closed boundary of the temporary structured obstacle, map the coordinates of the closed boundary to the map coordinate system and convert them into grid cells to generate independent no-patrol area data; In some embodiments, the closed boundary coordinates of temporary structured obstacles are obtained. Specifically, when the obstacle recognition model outputs a candidate obstacle instance, it simultaneously generates a minimum bounding box for that instance. For regularly shaped obstacles, the model outputs its minimum bounding box, described by the coordinates of eight vertices in the sensor coordinate system, with each vertex containing positional components in the x, y, and z directions. For irregular polygonal obstacles (such as irregularly stacked material areas), the model outputs an ordered closed point sequence describing its contour, with adjacent points connected sequentially to form a closed polyline. The aforementioned vertex or point sequence data together serve as the original closed boundary coordinates of the obstacle.

[0051] In some embodiments, the coordinates of the closed boundary are mapped to the map coordinate system and converted into grid cells to generate independent no-go zone data. Specifically, the coordinates of the closed boundary of the temporary structured obstacle are converted from the sensor coordinate system used for real-time perception to the global map coordinate system of the layered map to obtain the internal region of the closed boundary in the global coordinate system. According to the grid resolution preset by the layered map, the closed boundary of the internal region of the closed boundary is discretized to obtain a boundary grid sequence. Based on the boundary grid sequence, all internal grid cells located in the internal region of the closed boundary are determined. All internal grid cells are marked as occupied to generate no-go zone data. Specifically, firstly, the coordinates of the closed boundary in the sensor coordinate system are converted to the global map coordinate system of the layered map, combined with the robot's current global pose, to obtain the closed boundary and its internal region in the global coordinate system. Then, according to the grid resolution preset by the layered map (in this embodiment, it is consistent with the static base map, taking 0.1 meters), the closed boundary in the global coordinate system is discretized, and points are sampled along the boundary contour according to the grid cell size to generate a boundary grid sequence with the beginning and end connected. Then, a scan-line fill algorithm is used to determine and fill all raster cells inside the closed boundary based on the boundary raster sequence, forming a complete internal raster set. For regular boundaries with simple shapes, a seed fill algorithm can also be used to improve computational efficiency. Finally, all raster cells in the internal raster set are uniformly marked as occupied and encapsulated as independent no-patrol area data for subsequent layer overlay and path planning.

[0052] Step S104: The restricted patrol area data is incrementally overlaid and mounted onto the top layer of the pre-built layered map as an independent top-level temporary layer to form a top-level temporary restricted patrol layer. Please refer to Figure 3 In some embodiments, step S104 includes: Step S301: Create an independent temporary raster layer based on the restricted patrol area data, and mark the raster cells corresponding to the restricted patrol area data as occupied in the temporary raster layer; In some embodiments, the edge computing module creates an independent temporary raster layer based on the generated restricted patrol area data, serving as an independent storage and operation unit. In this temporary raster layer, all raster cells corresponding to the restricted patrol area data are marked as occupied. This layer only records the raster occupancy status, does not modify the underlying data of the layered map, is independent of the underlying permanent map storage of the layered map, and does not alter the original basic raster information such as roads and security points within the park.

[0053] Step S302: The temporary raster layer is used as the top overlay layer and incrementally overlaid on the bottom permanent map of the layered map to obtain the overlay result. In some embodiments, the edge computing module uses the marked temporary grid layer as the top overlay layer and incrementally overlays it onto the outermost layer of the pre-built layered map to obtain the map overlay result. This incremental overlay mechanism only processes grid cells within the coverage area of ​​temporary obstacles and does not modify the fixed road, wall, building, fire station, and permanent static obstacle grids in the underlying permanent map; the entire overlay process does not perform any operations on the underlying permanent map, the underlying base map remains intact, and the stability of the map reference on which the robot's global localization depends is ensured.

[0054] Step S303: In the overlay result, set the priority of the temporary raster layer to be higher than that of the underlying permanent map to obtain the target overlay result; In some embodiments, based on the above overlay results, the system sets the query and reference priority of the temporary raster layer to be higher than that of the underlying permanent map. When the robot navigation module reads the map raster status, it prioritizes retrieving the raster attributes of the corresponding coordinates in the top-level temporary raster layer; if the coordinates are within the coverage area of ​​the temporary no-go zone, the occupancy marker in the temporary layer is used as the standard, and the historical passage markers at the same location in the underlying permanent map are masked. Only when the target coordinates are not within the coverage area of ​​any top-level temporary raster layer will the system fall back to the underlying permanent map to read the raster attributes, thereby achieving the covering constraint of the temporary obstacle area on the underlying map.

[0055] Step S304: Determine the top-level temporary no-patrol layer based on the target overlay result.

[0056] In some embodiments, the system determines the final top-level temporary no-patrol layer for path planning based on the above target overlay results. This layer is composed of all currently overlaid independent temporary raster layers merged according to priority, providing the path planning module with complete spatial constraints of the no-patrol area. The map system simultaneously and automatically stores a snapshot of the differential version of this layer overlay, recording the current inspection path topology, raster passage weights, and differential data before and after the change, providing a basis for restoration after subsequent layer destruction and map rollback following obstacle removal. The cloud-based intelligent management platform simultaneously receives map change logs, providing real-time visual display of the temporary obstacle no-patrol area, achieving closed-loop visual management of map hot updates.

[0057] Step S105: Set the area covered by the top temporary no-patrol layer as an impassable core obstacle zone; In some embodiments, when the path planning module reads the map raster status, it first queries the top-level temporary restricted patrol layer through the priority retrieval mechanism set in step S303. If the target coordinates fall within the coverage area of ​​this layer, the occupancy marker in this layer is used as the standard, without backtracking to the bottom-level permanent map to read attributes. Thus, the entire raster area covered by this layer is equivalent to an impassable constraint space at the path planning level. Based on this, the system traverses all raster units in the top-level temporary restricted patrol layer, reads the index coordinates and occupancy status marker of each raster unit, and filters out all raster coordinates marked as occupied, forming an occupied raster coordinate list. The raster coordinates in this list are clustered according to the eight-connected neighborhood rule, dividing the occupied raster units into the same obstacle area instance. Each independent connected component corresponds to a core obstacle area. Each core obstacle area is stored in the core obstacle area management list of the layered map in the form of a raster index set. This list is an independent data structure, isolated from the storage area of ​​fixed obstacles in the bottom-level permanent map. Each entry in the list contains an obstacle area identifier, an occupied raster coordinate set, closed boundary coordinates, obstacle semantic type, and a timestamp. During the generation of subsequent obstacle avoidance and detour inspection candidate trajectories, the path planning module directly reads the grid coordinate set in the core obstacle area management list and performs an expansion operation based on this set to generate an inaccessible expanded restricted area containing the core obstacle area. When the top-level temporary restricted patrol layer is unloaded due to the removal of temporary obstacles, the corresponding entry in the core obstacle area management list is simultaneously deleted, and the map system performs a rollback based on the differential version snapshot stored in step S304, restoring the passability state of the underlying permanent map.

[0058] Step S106: Based on the type of temporary structured obstacle, query the preset mapping table to obtain the corresponding expansion distance. Combine the safety level of the robot's current inspection task to adjust the expansion distance with weights to obtain the target expansion distance. Expand the target expansion distance outward from the core obstacle area to obtain the impassable expansion restricted area including the core obstacle area. Define the passable area outside the expansion restricted area as the detour feasible area. In the detour feasible area, generate several obstacle avoidance detour inspection candidate trajectories. Use a multi-index cost function to evaluate the cost of each obstacle avoidance detour inspection candidate trajectory in order to select the optimal target detour inspection trajectory. In some embodiments, the system queries a preset mapping table to obtain the corresponding expansion distance based on the type of temporary structured obstacles. This includes: determining the corresponding risk classification label based on the type of temporary structured obstacles, wherein the risk classification label includes at least high-risk prohibited patrol categories and boundary inspection categories; and matching the risk classification label in the mapping table to determine the corresponding expansion distance. Specifically, when determining the expansion distance based on the type of temporary structured obstacles, the edge computing module has a preset risk classification mapping table and an expansion distance mapping table. After determining the risk classification label, the system queries the preset expansion distance mapping table using the label as an index. This mapping table configures a larger basic expansion distance (e.g., 1.0 meter) for high-risk prohibited patrol categories to ensure a sufficient safe distance between the robot and obstacles when navigating around them, while configuring a smaller expansion distance (e.g., 0.5 meters) for boundary inspection categories to preserve as much passable space as possible while ensuring safety.

[0059] It should be noted that the risk classification mapping table uses the semantic category labels output by the obstacle recognition model as the first-level index. Each record contains a semantic category field and a corresponding risk classification label field, with a one-to-one mapping between the two. The inflation distance mapping table uses the risk classification label as the second-level index. Each record contains a risk classification label field and a basic inflation distance field. When looking up the tables, the system first performs an exact string match in the risk classification mapping table based on the obstacle's semantic category to obtain the corresponding risk classification label. If the semantic category is not predefined in the risk classification mapping table, the system defaults to classifying it as a boundary inspection category and logs it, prompting administrators to supplement the mapping configuration. After obtaining the risk classification label, the system performs an equal-value match in the inflation distance mapping table using the label as the key to read the corresponding basic inflation distance value. If the label match fails, the system reverts to the default safe distance (1.0 meter in this embodiment) and triggers an alarm to prevent subsequent detours from failing due to empty inflation parameters caused by missing configuration.

[0060] In some embodiments, the expansion distance is weighted and adjusted based on the safety level of the robot's current inspection task to obtain the target expansion distance. Specifically, this involves: obtaining the task safety level identifier of the robot's current inspection task; determining the task weighting coefficient based on the task safety level identifier; and multiplying the expansion distance by the task weighting coefficient to obtain the target expansion distance. Specifically, after obtaining the basic expansion distance, the system reads the task safety level identifier from the task configuration file of the current inspection task. This identifier is pre-set according to the risk level of the area traversed by the task. For example, inspection tasks involving high-voltage power rooms, hazardous chemical storage areas, or densely populated areas are marked as high safety level; routine road inspections are marked as standard level; and closed, uninhabited area inspections are marked as low safety level. The edge computing module pre-stores a mapping table between safety levels and task weighting coefficients. High safety levels correspond to weighting coefficients greater than 1 (e.g., 1.3), standard levels correspond to coefficients of 1.0, and low safety levels correspond to coefficients less than 1 (e.g., 0.8). This mapping table can be remotely adjusted via a cloud management platform. Multiplying the base expansion distance obtained in step S103 with the weighting coefficient of the task yields the target expansion distance for the current inspection task. This weighting adjustment mechanism automatically increases the safe distance for the robot to bypass temporary obstacles when performing high-risk tasks, while appropriately reducing the distance in low-risk tasks to improve inspection efficiency, thus achieving a dynamic balance between safety and efficiency.

[0061] In some embodiments, the target expansion distance is expanded outward from the core obstacle area to obtain an impassable expanded restricted area containing the core obstacle area. Specifically, the target expansion distance is converted into the number of expanded graticules on the raster map; the raster set of the core obstacle area is used as the initial region, and the number of expanded graticules is used as the expansion radius to perform morphological expansion operations outward to obtain the expanded region; the entire expanded region is marked as an impassable expanded restricted area containing the core obstacle area. Specifically, after obtaining the target expansion distance, the system divides the distance value by the resolution and rounds up according to the preset raster resolution of the layered map (0.1 meters in this embodiment) to obtain the number of expanded graticules on the raster map. Subsequently, the raster set of the set core obstacle area is used as the initial seed region, and the number of expanded graticules is used as the expansion radius. A morphological expansion algorithm (e.g., using a circular structuring element whose radius is equal to the number of expanded graticules) is used to expand the core obstacle area outward layer by layer: each iteration incorporates the eight-neighbor or four-neighbor graticules of the current region boundary until the number of expansion layers reaches the number of expanded graticules. The final expanded region is the expanded region. The expanded region completely covers the core obstacle area and its surrounding safety buffer zone. The system marks all grid cells within the expanded region as impassable, thereby generating an impassable expanded restricted area that includes the core obstacle area.

[0062] Please refer to Figure 4In some embodiments, within the feasible detour area, several obstacle avoidance and detour inspection candidate trajectories are generated, including steps S401 to S404: Step S401: Based on the robot's current direction of travel, determine the left and right bypass boundaries of the expansion restricted area; In some embodiments, the robot's current position is projected onto the outline of the expansion restricted area along the direction of travel to obtain the entry reference point upstream of the obstacle and the return reference point downstream. Taking the robot's current heading angle in the global coordinate system as the reference direction, the azimuth angle of each point on the expansion restricted area outline relative to the robot's current position is calculated. The outline segment with the azimuth angle to the left of the reference direction is extracted as the left detour boundary, and the outline segment with the azimuth angle to the right of the reference direction is extracted as the right detour boundary. The left and right detour boundaries extend from the entry reference point upstream of the obstacle to the return reference point downstream, forming two basic path corridors that the robot can choose to detour through.

[0063] Step S402: Within the feasible detour area, select several path key control points along the left detour boundary and / or the right detour boundary according to the preset sampling step size, and generate multiple initial obstacle avoidance detour paths based on the path key control points. Each initial obstacle avoidance detour path starts from the original inspection path upstream of the obstacle, bypasses the obstacle, and returns to the original inspection path downstream. The detour direction or the location of the path key control points of each initial obstacle avoidance detour path is not exactly the same. In some embodiments, the robot determines the entry position on the original inspection path upstream of the obstacle and the return position on the original inspection path downstream of the obstacle based on the robot's current pose. Then, control points are sequentially selected along the left and right detour boundaries, starting from the entry position using an equal arc length sampling method, until the return position is reached. Each initial obstacle avoidance detour path leaves the original inspection path from the entry position upstream of the obstacle, bypasses the expansion restricted area along the selected side detour boundary, and rejoins the original inspection path at the return position downstream of the obstacle. To increase the diversity of candidate trajectories, the system can set an offset parameter during sampling, causing the control point positions of each path to make a small lateral offset near the detour boundary, or alternately use the left and right detour boundaries to generate paths with different detour directions, thereby forming multiple initial obstacle avoidance detour paths with different control point positions or detour directions.

[0064] It should be noted that the direction of the aforementioned lateral offset is defined as the normal direction at each point on the detour boundary, and the magnitude of the offset is taken from the preset offset set {-0.3 m, -0.15 m, 0 m, 0.15 m, 0.3 m}. For the same detour side, the system selects different offset magnitudes to generate multiple sets of control point sequences. The larger the magnitude, the further the corresponding path is from the detour boundary and the deeper it is into the feasible detour area. To cover both left and right detour modes, the system sequentially switches the selection of the detour side when generating the path: the path generated with an offset of 0 meters travels along the original detour boundary as the base path, the path generated with a positive offset is offset outward, and the path generated with a negative offset is offset inward (but the offset is still constrained to not intrude into the expansion restricted area). At the same time, the above offset set is applied to the left and right boundaries respectively, forming multiple initial obstacle avoidance detour paths with different detour directions or control point positions.

[0065] Step S403: Perform smoothing optimization on each initial obstacle avoidance path to generate smooth candidate trajectories corresponding to each initial obstacle avoidance path; In some embodiments, the system further performs smoothing optimization on each initial obstacle avoidance path to eliminate possible sharp turns and abrupt curvature changes in the path. The smoothing optimization employs a fitting method based on B-spline curves or cubic spline interpolation, using a discrete sequence of control points as the fitting control vertices. Spline interpolation generates a continuous smooth curve that passes through all control points or approximates them in the least-squares sense. During the optimization process, the objective function is to minimize the rate of change of path curvature, while simultaneously constraining the optimized path from intruding into impassable expansion zones. Finally, smooth candidate trajectories corresponding to each initial obstacle avoidance path are generated.

[0066] This smoothing process makes the candidate trajectory geometrically continuous and its first and second derivatives continuous, which is beneficial to the stability of the robot's motion during execution.

[0067] Step S404: Perform curvature verification on each smooth candidate trajectory, remove smooth candidate trajectories with curvature radii smaller than the robot's minimum turning radius, and retain the remaining smooth candidate trajectories as obstacle avoidance and detour inspection candidate trajectories.

[0068] In some embodiments, multiple path points are sampled at equal intervals along each smooth candidate trajectory. The radius of curvature of the trajectory is calculated at each path point (obtained by calculating the first and second derivatives of the position coordinates with respect to the arc length). The minimum radius of curvature value on the entire trajectory is taken and compared with the robot chassis's inherent minimum turning radius parameter. If the radius of curvature at any point in the trajectory is less than the robot's minimum turning radius, the trajectory is determined to exceed the robot's kinematic constraints and is discarded. All smooth candidate trajectories with radii of curvature greater than or equal to the minimum turning radius are retained and used as the final obstacle avoidance and bypass inspection candidate trajectory set for subsequent cost evaluation and optimal trajectory selection.

[0069] Please refer to Figure 5 In some embodiments, a multi-index cost function is used to evaluate the cost of each obstacle avoidance and detour inspection candidate trajectory in order to select the optimal target detour inspection trajectory, including steps S501 to S50: Step S501: For each obstacle avoidance and detour inspection candidate trajectory, calculate the minimum safe distance between each obstacle avoidance and detour inspection candidate trajectory and the boundary of the expansion restricted area, and determine the safety cost component based on the minimum safe distance. In some embodiments, multiple path points are sampled at preset intervals along the candidate obstacle avoidance and bypass inspection trajectory. For each path point, the minimum Euclidean distance between it and the boundary of the impassable expanded restricted area is calculated. The minimum distance value of all path points on the entire trajectory is taken as the clearance distance of the trajectory; safety cost component The safety cost is calculated by subtracting the reciprocal of the clearance distance from a preset safety threshold. A larger clearance distance results in a lower safety cost, and vice versa. The calculation formula is as follows: ,in The minimum distance between the candidate trajectory and the boundary of the expansion exclusion zone. A small constant is preset to prevent the denominator from being zero. In addition, if any position in the trajectory is less than the boundary of the restricted area by a preset safety threshold (0.3 meters by default), a penalty is added to the safety cost to strengthen the punishment for behavior that deviates from the boundary of the obstacle.

[0070] Step S502: Calculate the curvature value of each obstacle avoidance and bypass inspection candidate trajectory at each path point, and determine the smoothness cost component based on the curvature value. In some embodiments, for the smoothness cost component, the system samples multiple path points along each candidate trajectory at equal arc length intervals, and calculates the curvature value at each path point using the first and second derivatives of a cubic B-spline curve, thus obtaining the curvature sequence of the entire trajectory. Smoothness Cost Component The cost is calculated by taking a weighted combination of the sum of squares of curvature and the sum of squares of the rate of change of curvature at each point on the trajectory. The greater the curvature or the more drastic the change in curvature, the higher the cost. The formula is as follows: ,in The total number of sampling points. For the first The curvature value at each sampling point This represents the rate of change of curvature relative to arc length at that point. and Preset weighting coefficients (default value) =0.7、 =0.3).

[0071] Step S503: Calculate the total path length of each obstacle avoidance and detour inspection candidate trajectory, and determine the efficiency cost component based on the total path length. In some embodiments, for the efficiency cost component, the system performs numerical integration on the geometric curve of each candidate trajectory to calculate its total path length in the global map coordinate system (including detour segments and entry and return segments overlapping with the original inspection path), and compares this total length with the original inspection path length. Efficiency Cost Component Directly taking the total path length value and normalizing it results in lower efficiency and higher cost for longer paths. The calculation formula is as follows: ,in The total path length of the candidate trajectories. This represents the original length of the inspection path between the upstream and downstream sides of the obstacle. This is the efficiency cost weighting coefficient (default value is 1.0); this normalization process makes the efficiency costs between different candidate trajectories comparable and is not affected by changes in map scale.

[0072] Step S504: Obtain the preset passage cost weight of each grid passed by each obstacle avoidance and detour inspection candidate trajectory, and sum the preset passage cost weights of each grid to obtain the total passage cost component. In some embodiments, for the total passage cost component, the system iterates through all the grid cells traversed by each candidate trajectory, reads the preset passage cost weight of each grid cell from the layered map, and sums up the passage cost weights of all the grid cells traversed along the trajectory to obtain the total, which is the total passage cost component. The preset passage cost weights are pre-defined during the underlying permanent map construction phase. For example, grid cells near the edge of green belts are assigned higher passage weights (e.g., 1.5) due to the potential risk of being scratched by branches, while grid cells in the center of smooth roads are assigned lower passage weights (e.g., 1.0). This cumulative weight value represents the overall passage cost incurred in executing the trajectory; the smaller the cumulative value, the lower the degree of intrusion of the trajectory into high-passage-cost areas. The calculation formula is: ,in, The set of all raster cells traversed by the candidate trajectory. This is the preset access cost weight for grid cell g. The total access cost component represents the comprehensive access cost required to execute the trajectory; the smaller the value, the lower the degree of intrusion of the trajectory into high access cost areas.

[0073] Step S505: The safety cost component, smoothness cost component, efficiency cost component, and total passage cost component are weighted and summed to obtain the comprehensive cost value of each candidate trajectory. In some embodiments, after obtaining the four cost components, the system performs a weighted summation of the components to obtain the comprehensive cost value of each candidate trajectory. The calculation formula is as follows: ,in , , , These are the security cost component, the smoothness cost component, the efficiency cost component, and the sum of the passage costs, respectively. , , , The preset weighting coefficients for the corresponding components (the default value in this embodiment is 1). =0.35、 =0.20、 =0.25、 =0.20), each weight coefficient can be remotely configured through the cloud management platform to adapt to the strategy preferences in different scenarios.

[0074] Step S506: Determine the target detour inspection trajectory based on the comprehensive cost value.

[0075] In some embodiments, after curvature verification is completed, the system first generates a "feasible candidate trajectory set". This set contains only smooth candidate trajectories with a curvature radius not less than the robot's minimum turning radius. Trajectories marked as infeasible are not included in this set and do not participate in subsequent selection. After calculating the comprehensive cost value of all feasible candidate trajectories, the system sorts them in the feasible set according to their comprehensive cost value from smallest to largest, and determines the feasible candidate trajectory with the lowest comprehensive cost value as the optimal target detour inspection trajectory.

[0076] It should be noted that if multiple feasible candidate trajectories have the same and lowest comprehensive cost, the system will perform a secondary screening based on the priority of each cost component: the trajectory with the lower safety cost component will be selected first to ensure detour safety; if the safety cost components are also the same, the trajectory with the lower smoothness cost component will be selected to improve the stability and tracking accuracy of the robot's actual driving.

[0077] Step S107: Control the robot to perform inspection and scheduling in the park according to the target bypass inspection trajectory.

[0078] In some embodiments, the system sends the target circumvention and inspection trajectory to the robot's motion control module. The sent content includes a sequence of trajectory points (each trajectory point contains x and y coordinates and a heading angle θ in the global coordinate system) and the corresponding desired velocity information. Upon receiving the trajectory point sequence, the motion control module first converts the global coordinates into a relative pose in the robot's local coordinate system. Then, using a model predictive control-based tracking algorithm, with the robot's current pose as the state variable and linear and angular velocities as control variables, it solves the following optimization problem within each control cycle (50ms). The state variable is defined as... The control quantity is defined as ,in For linear acceleration, Let angular acceleration be the value. The prediction time domain is taken as... Steps, time interval between each step The total prediction time is 1.0s. The objective function is: The first term is the trajectory tracking error penalty term. For the candidate trajectory, the first The reference state sequence of the step, This is the state weight matrix; the second term is the control magnitude penalty term. This is the control weight matrix; the third term is the penalty term for the rate of change of the control variable. This is the rate-of-change weight matrix, used to suppress high-frequency jitter in the control output. Constraints include: kinematic constraints (linear velocity under the differential driving model). angular velocity acceleration angular acceleration Safety constraints (the minimum distance between all points on the predicted trajectory and the boundary of the expansion exclusion zone is not less than 0.1m). After solving this optimization problem, output the optimal linear velocity command. and angular velocity command After receiving the command, the chassis controller sends it to the drive motor and steering motor via the CAN bus, driving the robot to move along the target trajectory.

[0079] This application acquires real-time environmental perception data and inputs it into an obstacle recognition model, enabling the system to accurately identify temporary structured obstacles and their closed boundaries not recorded in static maps within complex park environments. This provides accurate spatial information for subsequent obstacle avoidance operations. The closed boundary coordinates are then mapped and converted into grid cells to generate independent no-go zone data, which is incrementally mounted as a top-level temporary layer onto the outermost layer of the layered map. This allows temporary obstacle information to be quickly injected into the map system without affecting the stability of the underlying permanent map, while avoiding destructive modifications to the global map and ensuring the accuracy of map data operations. Furthermore, the area covered by the top-level temporary no-go layer is set as an impassable core obstacle zone, clearly defining the absolute no-entry range and establishing rigid spatial constraints for trajectory generation. Finally, the basic expansion distance is obtained by querying a mapping table based on the type of temporary structured obstacle and combined with the robot's current inspection data. The safety level of the task is used to weight and adjust the expansion distance to obtain the target expansion distance. Then, based on the core obstacle area, the expansion extends outward to form an impassable expansion no-go zone. This mechanism allows the size of the expansion no-go zone to adaptively adjust according to the degree of danger reflected by the obstacle type and the dynamic changes in task safety requirements. This avoids the problems of overly conservative or under-aggressive detours caused by fixed expansion parameters, ensuring accurate adaptation of obstacle avoidance safety margin. Subsequently, a feasible detour area is delineated outside the expansion no-go zone, and several candidate trajectories are generated. A multi-index cost function is used to comprehensively evaluate the safety, smoothness, and passage efficiency of each candidate trajectory, and the optimal target detour trajectory is selected from them. This ensures that the final detour path is not only safe and away from obstacles and kinematically feasible, but also has the optimal path cost. Finally, the robot is controlled to perform inspections along the target trajectory, realizing a closed-loop process from perception, mapping, no-go zone delineation, expansion optimization to trajectory generation and execution.

[0080] Please see Figure 6 This application also provides a robot inspection and scheduling system for a park, which can realize the above-mentioned robot inspection and scheduling system for a park. The system includes: The first module 100 is used to acquire real-time environmental perception data during the robot's inspection process; The second module 200 is used to input real-time environmental perception data into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in the pre-built static map. The third module 300 is used to obtain the coordinates of the closed boundary of temporary structured obstacles, map the coordinates of the closed boundary to the map coordinate system and convert them into grid cells to generate independent no-patrol area data; The fourth module 400 is used to incrementally overlay and mount the restricted patrol area data as an independent top-level temporary layer onto the top layer of the pre-built layered map to form a top-level temporary restricted patrol layer. Module 500 is used to set the area covered by the top temporary no-patrol layer as an impassable core obstacle zone; The sixth module 600 is used to query a preset mapping table based on the type of temporary structured obstacles to obtain the corresponding expansion distance. The expansion distance is then weighted and adjusted according to the safety level of the robot's current inspection task to obtain the target expansion distance. The target expansion distance is then expanded outward from the core obstacle area to obtain the impassable expansion restricted area that includes the core obstacle area. The passable area outside the expansion restricted area is defined as the detour feasible area. Within the detour feasible area, several obstacle avoidance and detour inspection candidate trajectories are generated. The cost of each obstacle avoidance and detour inspection candidate trajectory is evaluated using a multi-index cost function to select the optimal target detour inspection trajectory. The seventh module 700 is used to control the robot to perform inspections and scheduling in the park according to the target around the inspection trajectory.

[0081] The specific implementation method of the robot inspection and scheduling method in this park is basically the same as the specific implementation of the robot inspection and scheduling system in the above-mentioned park, and will not be repeated here.

[0082] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned robot inspection and scheduling method for the park. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0083] Another embodiment of this application also provides a hardware structure for an electronic device, which includes: The processor can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to achieve the technical solutions provided in the embodiments of this application. The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory and called by the processor to execute the robot inspection and scheduling method for the park according to the embodiments of this application. Input / output interfaces are used to implement information input and output; The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). A bus is used to transfer information between various components of a device, such as processors, memory, input / output interfaces, and communication interfaces. The processor, memory, input / output interfaces, and communication interfaces communicate with each other within the device via a bus.

[0084] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-mentioned robot inspection and scheduling method for the park.

[0085] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0086] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0087] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0088] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0089] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0090] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0091] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for scheduling robot inspections in a park, characterized in that, The method includes the following steps: Acquire real-time environmental perception data during the robot's inspection process; The real-time environmental perception data is input into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in a pre-built static map. This includes: retrieving environmental images captured by fixed cameras in the park and fusing the environmental images with the real-time environmental perception data to obtain fused perception data; inputting the fused perception data into a pre-trained obstacle recognition model to identify candidate obstacles in the environment and their corresponding spatial boundaries; and matching the spatial boundaries of each candidate obstacle with the pre-built static map to eliminate obstacles that already exist in the pre-built static map. Overlapping candidate obstacles are used to obtain first temporary obstacles; based on the spatial boundaries of each first temporary obstacle, the transient probability is calculated, and obstacles with transient probabilities lower than a first preset threshold are retained as second temporary obstacles; the duration of each second temporary obstacle is predicted, and short-term road-occupying obstacles with durations lower than a second preset threshold are filtered out to obtain third temporary obstacles; when the spatial boundary offset of the third temporary obstacle in consecutive multi-frame perception data is less than a third preset threshold and the morphological overlap is greater than a fourth preset threshold, and it has a closed and fitable rigid boundary, it is determined to be a temporary structured obstacle; Obtain the closed boundary coordinates of the temporary structured obstacle, map the closed boundary coordinates to the map coordinate system and convert them into grid cells to generate independent no-patrol area data; The data of the restricted patrol area is incrementally overlaid and mounted onto the top layer of the pre-constructed layered map as an independent top-level temporary layer to form a top-level temporary restricted patrol layer. The area covered by the top-level temporary no-patrol layer is designated as an impassable core obstacle zone; Based on the type of the temporary structured obstacle, a preset mapping table is queried to obtain the corresponding expansion distance, including: determining the corresponding risk classification label based on the type of the temporary structured obstacle, wherein the risk classification label includes at least high-risk prohibited patrol category and boundary patrol category; and matching the risk classification label in the mapping table to determine the corresponding expansion distance. The expansion distance is weighted and adjusted based on the safety level of the robot's current inspection task to obtain the target expansion distance. Specifically, the task safety level identifier of the robot's current inspection task is obtained, and a task weighting coefficient is determined based on the task safety level identifier. The expansion distance is then multiplied by the task weighting coefficient to obtain the target expansion distance. Based on the core obstacle area, the target expansion distance is expanded outward to obtain an impassable expansion restricted area including the core obstacle area. The passable range outside the expansion restricted area is defined as a detourable area. Within the detourable area, several obstacle avoidance and detour inspection candidate trajectories are generated. A multi-index cost function is used to evaluate the cost of each obstacle avoidance and detour inspection candidate trajectory in order to select the optimal target detour inspection trajectory. The robot is controlled to perform inspections and scheduling within the park by following the target's inspection trajectory.

2. The method according to claim 1, characterized in that, The process of mapping the coordinates of the closed boundary to a map coordinate system and converting them into raster cells to generate independent no-patrol area data specifically involves: The coordinates of the closed boundary of the temporary structured obstacle are transformed from the sensor coordinate system used for real-time perception to the global map coordinate system of the layered map to obtain the internal region of the closed boundary in the global coordinate system. Based on the preset grid resolution of the layered map, the closed boundary of the inner region of the closed boundary is discretized to obtain a boundary grid sequence. Based on the boundary grid sequence, determine all internal grid cells located within the region inside the closed boundary. All internal grid cells are marked as occupied, thereby generating the no-patrol area data.

3. The method according to claim 1, characterized in that, The step of incrementally overlaying and mounting the restricted patrol area data as an independent top-level temporary layer onto the outermost layer of the pre-constructed layered map to form a top-level temporary restricted patrol layer includes: Create an independent temporary raster layer based on the restricted patrol area data, and mark the raster cells corresponding to the restricted patrol area data as occupied in the temporary raster layer; The temporary raster layer is used as the top overlay layer and incrementally overlaid on the bottom permanent map of the layered map to obtain the overlay result; In the overlay result, the temporary raster layer is set to have a higher priority than the underlying permanent map to obtain the target overlay result; The top-level temporary no-patrol layer is determined based on the target overlay result.

4. The method according to claim 1, characterized in that, The expansion distance outward from the core obstacle zone to obtain an impassable expanded restricted zone including the core obstacle zone is as follows: Convert the target inflation distance into the number of inflation grids on the grid map; Using the grid set of the core obstacle area as the initial region and the number of expanded grids as the expansion radius, a morphological expansion operation is performed outward to obtain the expanded region. The entire expanded region is marked as an impassable expanded restricted area, including the core obstacle zone.

5. The method according to claim 1, characterized in that, Within the feasible detour area, several obstacle avoidance and detour inspection candidate trajectories are generated, including: Based on the robot's current direction of travel, determine the left and right bypass boundaries of the expansion restricted area; Within the feasible detour area, several path key control points are selected along the left detour boundary and / or the right detour boundary according to a preset sampling step size, and multiple initial obstacle avoidance detour paths are generated based on the path key control points. Each of the initial obstacle avoidance detour paths starts from the original inspection path upstream of the obstacle, bypasses the obstacle, and returns to the original inspection path downstream. The detour direction or the position of the path key control points of each of the initial obstacle avoidance detour paths are not exactly the same. Each of the initial obstacle avoidance paths is smoothed and optimized to generate a smoothed candidate trajectory corresponding to each initial obstacle avoidance path; For each of the smooth candidate trajectories, a curvature check is performed, and smooth candidate trajectories with a curvature radius smaller than the robot's minimum turning radius are eliminated, while the remaining smooth candidate trajectories are retained as the obstacle avoidance and detour inspection candidate trajectories.

6. The method according to claim 1, characterized in that, The step of evaluating the cost of each of the candidate obstacle avoidance and detour inspection trajectories using a multi-index cost function to select the optimal target detour inspection trajectory includes: For each of the obstacle avoidance and detour inspection candidate trajectories, calculate the minimum safe distance between each of the obstacle avoidance and detour inspection candidate trajectories and the boundary of the expansion restricted area, and determine the safety cost component based on the minimum safe distance; Calculate the curvature value of each of the aforementioned obstacle avoidance and bypass inspection candidate trajectories at each path point, and determine the smoothness cost component based on the curvature value; Calculate the total path length of each of the obstacle avoidance and detour inspection candidate trajectories, and determine the efficiency cost component based on the total path length; Obtain the preset passage cost weights of each grid cell traversed by each of the proposed obstacle avoidance and detour inspection candidate trajectories, and sum the preset passage cost weights of each grid cell to obtain the total passage cost component; The comprehensive cost of each candidate trajectory is obtained by weighted summing of the security cost component, the smoothness cost component, the efficiency cost component, and the total passage cost component. The target bypass inspection trajectory is determined based on the comprehensive cost value.

7. A robot inspection and scheduling system for a park, used to execute the robot inspection and scheduling method for a park as described in any one of claims 1-6, characterized in that, The system includes: The first module is used to acquire real-time environmental perception data during the robot's inspection process; The second module is used to input the real-time environmental perception data into a preset obstacle recognition model to identify temporary structured obstacles in the park that are not included in the pre-built static map. The third module is used to obtain the closed boundary coordinates of the temporary structured obstacle, map the closed boundary coordinates to the map coordinate system and convert them into grid cells to generate independent no-patrol area data; The fourth module is used to incrementally overlay and mount the restricted patrol area data as an independent top-level temporary layer onto the outermost layer of the pre-constructed layered map to form a top-level temporary restricted patrol layer. The fifth module is used to set the area covered by the top-level temporary no-patrol layer as an impassable core obstacle zone; The sixth module is used to query a preset mapping table based on the type of the temporary structured obstacle to obtain the corresponding expansion distance, and to adjust the expansion distance by weighting it in combination with the safety level of the robot's current inspection task to obtain the target expansion distance. The target expansion distance is then expanded outward from the core obstacle area to obtain an impassable expansion restricted area that includes the core obstacle area. The passable area outside the expansion restricted area is defined as a detourable area. Within the detourable area, several obstacle avoidance and detour inspection candidate trajectories are generated. A multi-index cost function is used to evaluate the cost of each obstacle avoidance and detour inspection candidate trajectory in order to select the optimal target detour inspection trajectory. The seventh module is used to control the robot to perform inspections and scheduling within the park according to the target's inspection trajectory.

Images