Dynamic path planning method and system for inspection robot

Abstract

The invention belongs to the technical field of data processing, and particularly discloses a dynamic path planning method and system for an inspection robot, and the method comprises the steps: collecting field data through a sensor, processing a laser scanning point cloud and a camera image through a positioning mapping algorithm, and fusing multi-source information to generate a real-time map. Obtaining a three-dimensional environment model containing a dynamic obstacle position; according to the three-dimensional environment model, obstacle features are extracted, a deep learning model is adopted to identify moving personnel and randomly placed goods, obstacle types and movement tracks are judged, and a classified obstacle data set is obtained; through the classified obstacle data set, the distance and the relative speed between the obstacle and the current position of the robot are calculated, and if the distance is smaller than a preset threshold value and the speed is larger than zero, high-priority interference is marked; according to the invention, the problem of inaccurate dynamic obstacle identification and obstacle avoidance in a complex scene in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a dynamic path planning method and system for inspection robots. Background Technology

[0002] In industrial inspection scenarios, robots need to achieve autonomous navigation and obstacle avoidance in complex and dynamic environments to ensure the efficiency and safety of inspection tasks.

[0003] However, the core technical challenge lies in how to construct a real-time updated 3D environment model in a dynamic and unpredictable environment, and how to achieve accurate path planning and dynamic obstacle avoidance based on this model.

[0004] This major problem encompasses challenges across the entire process from data collection to decision execution. Especially in scenarios such as industrial plants or warehouses, the environment often includes moving people, randomly placed goods, and other unpredictable dynamic obstacles. Robots need to accurately identify the type, location, and trajectory of these obstacles under limited computing resources and real-time requirements, while also assessing their potential threat level to their own path.

[0005] Focusing further on the details, the laser point cloud and camera image data collected by sensors may contain noise or be incomplete. How to integrate these multi-source heterogeneous data to generate a reliable real-time map is the primary technical challenge.

[0006] Furthermore, when identifying obstacles, the characteristics of moving people and goods differ significantly, and deep learning models may face the risk of misjudgment or omission, especially in situations with insufficient light or severe occlusion.

[0007] More importantly, even if a path plan is generated, predicting potential collision points and adjusting path parameters in a timely manner in a dynamic environment to cope with sudden changes remains a complex problem.

[0008] These seemingly minor issues all revolve around the core problem: how to enable robots to accurately perceive, quickly make decisions, and navigate safely in dynamic environments. This is especially important in industrial inspection scenarios, where any delay in perception or decision-making can lead to task failure or safety accidents. Therefore, solving this problem is of paramount importance for improving robot autonomy. Summary of the Invention

[0009] This invention provides a dynamic path planning method and system for inspection robots, aiming to solve the problem of inaccurate dynamic obstacle recognition and avoidance in complex scenarios in existing technologies.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The dynamic path planning method for inspection robots includes: collecting on-site data through sensors, processing laser scanning point clouds and camera images, fusing multi-source information to generate a real-time map, and obtaining a 3D environment model containing the locations of dynamic obstacles; extracting obstacle features based on the 3D environment model, using a deep learning model to identify several obstacles, determining the obstacle type and motion trajectory, and obtaining a classified obstacle dataset; calculating the distance and relative speed between the obstacle and the robot's current position using the classified obstacle dataset, marking obstacles as high-priority interference if the distance is less than a preset threshold and the speed is greater than zero, and obtaining a priority-ranked obstacle list; and inputting the priority-ranked obstacle list into the path planning algorithm. The system generates alternative paths based on the current map, evaluates the length and obstacle avoidance probability of each path, and obtains an optimal path sequence. Using this optimal path sequence, the robot's trajectory is simulated, and a particle filter method is used to predict potential collision points. If a predicted point overlaps with an obstacle, the sequence parameters are adjusted to obtain a corrected path sequence. Based on the corrected path sequence, a control command sequence is generated and transmitted to the robot's actuators. Execution deviations are monitored in real time, and it is determined whether the deviation exceeds a threshold. If it does, the commands are recalculated to obtain updated control commands. Using the updated control commands and fused with sensor feedback data, the 3D environment model is continuously optimized, and path planning is iterated to obtain the final inspection path adapted to the dynamic environment.

[0011] In one aspect of the invention, on-site data is collected by sensors, laser scanning point clouds and camera images are processed using a positioning mapping algorithm, and multi-source information is fused to generate a real-time map, thereby obtaining a three-dimensional environment model containing the locations of dynamic obstacles, including: By collecting data from the on-site environment through sensors, raw information including laser point clouds and camera images is obtained, and preliminary environmental perception data is acquired. Based on the collected laser point cloud, a positioning algorithm is used to spatially register the point cloud data, determine the position information in the point cloud, and form preliminary spatial structure data. By using camera images and combining mapping algorithms to extract features and spatially project the image content, visual mapping data corresponding to laser point clouds is obtained, and the basic content of real-time maps is constructed. By employing multi-source fusion technology, spatial structure data and visual mapping data are integrated and processed to obtain fused environmental data that includes the location of dynamic obstacles; If dynamic obstacles are detected in the fused environmental data, the obstacle locations are filtered by a preset threshold to determine whether the obstacles are in critical areas, and filtered obstacle distribution data is generated. Based on the filtered obstacle distribution data, a three-dimensional environment model containing the dynamic obstacle locations is constructed to obtain complete environmental description information; By continuously updating the 3D environment model and adjusting the movement trajectory of dynamic obstacles in real time, the final dynamic environmental information is determined.

[0012] In one aspect of the invention, obstacle features are extracted based on a three-dimensional environment model, and a deep learning model is used to identify moving people and randomly placed goods, determine the obstacle type and movement trajectory, and obtain a classified obstacle dataset, including: An initial environmental model is constructed by acquiring 3D environmental data, and multi-source information is integrated using sensor data fusion technology to obtain preliminary environmental model data. Based on the preliminary environmental model data, obstacle features are extracted, and a pre-trained deep learning model is applied to perform preliminary identification of objects in the environment to determine the set of potential obstacle objects. For a set of potential obstacle objects, analyze the distinguishing features between moving people and placed goods. If dynamic displacement is detected, the object is classified as a moving person, and a subset of dynamic objects is obtained after classification. Based on the categorized subset of dynamic objects and combined with historical location data, the continuity characteristics of the motion trajectory are calculated, and the trajectory analysis method is used to determine the motion trajectory pattern of the dynamic objects. By using the motion trajectory pattern, static objects are reconfirmed. If the position of a static object does not change and it meets the characteristics of cargo, it is classified as cargo and a subset of static obstacles is determined. Based on the dynamic object subset and the static obstacle subset, obstacle type information is integrated, and classification and organization techniques are applied to construct a complete obstacle dataset. By using a complete obstacle dataset, the 3D environment model is updated, and data storage technology is used to save the classification results, thereby obtaining real-time updated environmental information.

[0013] In one aspect of the invention, using a classified obstacle dataset, the distance and relative speed between the obstacle and the robot's current position are calculated. If the distance is less than a preset threshold and the speed is greater than zero, it is marked as a high-priority obstacle, thus obtaining a priority-ranked list of obstacles, including: By extracting the coordinate and motion information of each obstacle from the obstacle dataset and combining it with real-time data on the robot's position, the spatial distance and relative velocity between each obstacle and the robot are calculated to obtain preliminary distance and velocity results. Based on the calculated distance and speed results, condition judgment is made for each obstacle. If the distance is lower than a preset threshold and the relative speed is greater than zero, it is marked as a high-priority interference, and the marked obstacle classification set is obtained. Using the labeled obstacle classification set, all obstacles are sorted according to the identifier of high-priority interference to generate an initial priority sorting list; By combining the initial priority ranking list with dynamic updates of position and velocity data, the distance and velocity status of each obstacle are re-verified to determine whether the high priority conditions are still met, and the updated priority ranking results are obtained. Based on the updated priority ranking results, detailed motion trajectory data of obstacles with high priority interference are obtained, potential collision risks are analyzed, and the final priority ranking list is determined. By combining the final priority ranking list with the changes in the robot's position and obstacle dataset acquired in real time, the fluctuations in distance and speed are continuously monitored to determine whether the priority needs to be adjusted, and a dynamically adjusted obstacle priority list is obtained. A dynamically adjusted obstacle priority list is used to generate structured data output, which is then called by the subsequent path planning module to determine the execution order of the real-time obstacle avoidance strategy.

[0014] In one aspect of the invention, an obstacle list ordered by priority is input into a path planning algorithm, candidate paths are generated in conjunction with the current map, the length and obstacle avoidance probability of each path are evaluated, and a preferred path sequence is obtained, including: An initial obstacle distribution map is generated by retrieving sorted data from the obstacle priority list and combining it with map data sources in the current map library. Based on the obstacle distribution map and path planning method, a set of alternative paths is constructed using a pre-established path generation model; For each path in the candidate path set, a path evaluation method is used to calculate the path length value and record the length data of each path; If the path length exceeds the preset threshold range, the obstacle avoidance probability value is further analyzed by probability calculation to obtain the obstacle avoidance probability data for each path. Based on the obstacle avoidance probability value and path length value, the candidate path set is comprehensively sorted using the sequence sorting method to determine the preliminary preferred path order; By performing secondary verification on the path data in the preferred path sequence and combining it with the dynamic information after the obstacle distribution map is updated, it is determined whether the path sequence needs to be adjusted, and the final preferred path sequence is obtained. If the obstacle avoidance probability value in the final preferred path sequence is lower than the preset threshold, then the candidate path set is regenerated by combining the latest map data source in the current map library, and the path evaluation method and sequence sorting method are executed cyclically to determine a new preferred path sequence.

[0015] In one aspect of the invention, by optimizing the path sequence and simulating the robot's motion trajectory, a particle filtering method is used to predict potential collision points. If the predicted point overlaps with an obstacle, the sequence parameters are adjusted to obtain a corrected path sequence, including: By optimizing the path sequence, the initial simulation data of the robot's motion trajectory is obtained. The particle filtering method is used to estimate the distribution of trajectory points and obtain the prediction results of potential collision points. Based on the prediction results of potential collision points, determine whether the predicted points overlap with obstacles. If the predicted points overlap with obstacles, extract the key control points in the path parameters and determine the range of parameters that need to be adjusted. By extracting the range of path parameters, the parameters are optimized using preset adjustment rules to generate the first corrected path sequence; Based on the first corrected path sequence, the robot's motion trajectory is re-simulated to obtain the updated trajectory point distribution and determine whether there are any new potential collision points. If potential collision points still exist in the updated trajectory point distribution, the local weights of the path parameters are adjusted for the new collision point locations to obtain the second corrected path sequence. The final robot motion trajectory data is obtained through the second corrected path sequence. The particle filtering method is used to verify the final trajectory and determine whether the collision risk has been eliminated. If no collision risk is found during the final trajectory verification, the second corrected path sequence will be used as the final path output to generate control commands for the robot's movement.

[0016] In one aspect of the invention, a control command sequence is generated based on the corrected path sequence, transmitted to the robot actuator, and the execution deviation is monitored in real time to determine whether the deviation exceeds a threshold. If it does, the commands are recalculated to obtain updated control commands, including: Initial data is obtained from the path sequence, and the path sequence is transformed into a control command sequence using a pre-established mapping model to obtain a preliminary set of control commands. The control command set is sent to the robot execution module through the transmission channel, and the status data during the execution process is collected in real time to determine the dynamic changes during the execution process; For dynamically changing data, a real-time monitoring mechanism is used to compare the data with preset execution standards, calculate the execution deviation value, and determine whether the deviation exceeds the preset threshold. If the deviation exceeds the preset threshold, the instruction update process is triggered, and the adjusted control instruction sequence is regenerated by combining the execution feedback data and the path correction logic. By adjusting the control command sequence, the command input of the robot execution module is updated, new execution status data is obtained, and the execution effect is continuously monitored. Based on continuously monitored data, the deviation judgment process is executed cyclically. If the deviation still exceeds the preset threshold, the instruction update and adjustment steps are repeated to obtain stable execution results. After the execution results stabilize, the final path sequence and the correspondence between the control commands are recorded and stored in the historical database for reference in subsequent tasks.

[0017] In one aspect of the invention, by updating control commands, fusing sensor feedback data, continuously optimizing the three-dimensional environment model, and employing a path planning algorithm for iterative path planning, a final inspection path adapted to the dynamic environment is obtained, including: By collecting real-time environmental information through sensors and fusing feedback data from sensors and historical records, an initial dataset is constructed to obtain preliminary environmental perception results. Based on the preliminary environmental perception results and combined with the pre-established 3D model, a Kalman filter is used to perform data alignment and correction operations to determine the basis of the updated environmental model. Based on the updated environmental model, and integrating real-time adjustments of control commands, the A* algorithm is applied to perform preliminary path calculations and obtain an initial inspection path scheme. For the initial inspection path plan, combined with the dynamic environment change information, OpenCV is used to perform path feasibility analysis. If obstacles or environmental changes exceed the preset threshold, the path parameters are readjusted to obtain an optimized path plan. By continuously collecting sensor feedback data through the optimized path scheme, updating the dynamic elements in the environmental model, and determining whether the path execution conditions are met; If the path execution conditions are met, the final inspection path data will be generated based on the updated environment model and the optimized path scheme. If the conditions are not met, return to the OpenCV-based path feasibility analysis stage and iteratively calculate the path scheme. After obtaining the final inspection path data, the execution details of the inspection path are dynamically adjusted in conjunction with real-time control commands to determine the real-time adaptive solution for the inspection task.

[0018] In another aspect, the present invention also relates to a dynamic path planning system for an inspection robot, the system comprising: The data acquisition and modeling module is configured to collect on-site data through sensors, process laser scanning point clouds and camera images using a positioning mapping algorithm, fuse multi-source information to generate a real-time map, and obtain a three-dimensional environment model containing the location of dynamic obstacles. The obstacle recognition and classification module is configured to extract obstacle features based on a 3D environment model, use a deep learning model to identify moving people and randomly placed goods, determine the obstacle type and movement trajectory, and obtain a classified obstacle dataset. The priority evaluation module is configured to calculate the distance and relative speed between the obstacle and the robot's current position using the classified obstacle dataset. If the distance is less than a preset threshold and the speed is greater than zero, it is marked as a high-priority interference, thus obtaining a priority-ranked list of obstacles. The path planning module is configured to take a priority-sorted list of obstacles, which is then fed into the path planning algorithm. This algorithm generates alternative paths based on the current map, evaluates the length and obstacle avoidance probability of each path, and obtains the optimal path sequence. The collision prediction and correction module is configured to simulate the robot's motion trajectory by optimizing the path sequence, predict potential collision points using a particle filtering method, and adjust the sequence parameters to obtain a corrected path sequence if the predicted point overlaps with an obstacle. The control command generation and update module is configured to generate a control command sequence based on the corrected path sequence, transmit it to the robot actuator, monitor the execution deviation in real time, determine whether the deviation exceeds the threshold, and if it does, recalculate the command to obtain the updated control command. The environmental model optimization module is configured to continuously optimize the 3D environmental model by integrating sensor feedback data with updated control commands, and to obtain the final inspection path that adapts to the dynamic environment by iteratively planning the path using a path planning algorithm.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention collects data through sensors, fuses laser point cloud and image information to generate a real-time 3D environment model, and uses deep learning to accurately identify dynamic obstacle types and trajectories. Then, based on distance and speed, it assesses priorities, optimizes path planning algorithms to generate preferred paths with high obstacle avoidance probabilities, and uses particle filtering to predict collision risks, dynamically correcting path parameters. Finally, it generates control commands adapted to the environment, monitors execution deviations in real time, and continuously optimizes the model. This invention achieves high efficiency and safety for robot inspection in dynamic environments, significantly improving adaptability and obstacle avoidance performance in complex scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts for a dynamic path planning method for an inspection robot according to the present invention.

[0022] Figure 2This is the second flowchart of a dynamic path planning method for an inspection robot according to the present invention.

[0023] Figure 3 This is the third flowchart of a dynamic path planning method for an inspection robot according to the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments described. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the protection scope of the present invention.

[0025] Example 1 Please see Figures 1-3 As shown, this embodiment discloses a dynamic path planning method and system for an inspection robot, including: collecting on-site data through sensors, processing laser scanning point clouds and camera images using a positioning mapping algorithm, fusing multi-source information to generate a real-time map, and obtaining a three-dimensional environment model containing the locations of dynamic obstacles; extracting obstacle features based on the three-dimensional environment model, using a deep learning model to identify moving personnel and randomly placed goods, determining the obstacle type and movement trajectory, and obtaining a classified obstacle dataset; calculating the distance and relative speed between the obstacle and the robot's current position using the classified obstacle dataset, and marking obstacles as high-priority interference if the distance is less than a preset threshold and the speed is greater than zero, thus obtaining a priority-ranked obstacle list; and then, based on the priority-ranked obstacle list... The data is input into the path planning algorithm, which generates alternative paths based on the current map, evaluates the length and obstacle avoidance probability of each path, and obtains an optimal path sequence. Using this optimal path sequence, the robot's trajectory is simulated, and a particle filter method is used to predict potential collision points. If a predicted point overlaps with an obstacle, the sequence parameters are adjusted to obtain a corrected path sequence. Based on the corrected path sequence, a control command sequence is generated and transmitted to the robot's actuator. Execution deviations are monitored in real time, and it is determined whether the deviation exceeds a threshold. If it does, the commands are recalculated to obtain updated control commands. Using the updated control commands and sensor feedback data, the 3D environment model is continuously optimized, and the path planning algorithm iteratively plans the path to obtain the final inspection path adapted to the dynamic environment.

[0026] This invention collects data through sensors, fuses laser point cloud and image information to generate a real-time 3D environment model, and uses deep learning to accurately identify dynamic obstacle types and trajectories. Then, based on distance and speed, it assesses priorities, optimizes path planning algorithms to generate preferred paths with high obstacle avoidance probabilities, and uses particle filtering to predict collision risks, dynamically correcting path parameters. Finally, it generates control commands adapted to the environment, monitors execution deviations in real time, and continuously optimizes the model. This invention achieves high efficiency and safety for robot inspection in dynamic environments, significantly improving adaptability and obstacle avoidance performance in complex scenarios.

[0027] Example 2 Please see Figures 1-3 As shown, this embodiment is a further optimization based on Embodiment 1. In this embodiment, firstly, the method comprehensively perceives the on-site environment through multi-source sensors, acquires laser scanning point cloud and camera image data, and constructs a three-dimensional environment model including the location of dynamic obstacles. Specifically, step S1 involves collecting on-site data through sensors and performing preliminary processing on the data to form an environmental information foundation that can be used for subsequent analysis.

[0028] Step S1: Sensor Data Acquisition and Preliminary Processing. Multiple sensor devices deployed on the robot, such as LiDAR and cameras, are used to acquire real-time data about the environment at the inspection site. LiDAR generates high-precision point cloud data to capture the three-dimensional spatial structure information of the environment, while cameras acquire color image data to supplement the texture and color features of the environment. The acquired data undergoes preliminary noise reduction processing to remove anomalies caused by equipment vibration or environmental interference, ensuring the accuracy of subsequent processing.

[0029] Step S11: Spatial registration of laser point cloud data. For the point cloud data acquired by the lidar, a positioning algorithm is used to perform spatial registration processing on each frame of the point cloud. Specifically, by matching feature points between point clouds, point cloud data acquired at different times are aligned to the same coordinate system, forming continuous spatial structure information. This process first extracts key points in the point cloud, such as edge points or planar points, and then determines the optimal transformation matrix by iteratively calculating the distance between points, integrating the point cloud data into unified preliminary spatial structure data. It should be noted that the key to spatial registration is to reduce accumulated errors and ensure the consistency of point cloud data during long-term acquisition.

[0030] Step S12: Feature Extraction and Spatial Projection of Camera Images. The image data acquired by the camera is processed using a mapping algorithm. First, salient features such as edges, corners, or color regions are extracted from the image using image processing techniques. These features are then projected into three-dimensional space, establishing a correspondence with the laser point cloud data. Specifically, the projection process relies on the camera's intrinsic and extrinsic parameters. Triangulation methods are used to transform two-dimensional image features into positional information in three-dimensional space, constructing visual mapping data that matches the point cloud data. This process provides visual environmental information to supplement subsequent multi-source data fusion.

[0031] Step S2: Multi-source information fusion and real-time map generation. After acquiring preliminary spatial structure data and visual mapping data, multi-source fusion technology is used to integrate the two to generate a real-time map containing the locations of dynamic obstacles. Specifically, through timestamp alignment and spatial coordinate matching, the spatial structure information of the laser point cloud is unified with the visual feature information of the camera to construct fused environmental data. This process not only improves the comprehensiveness of environmental perception but also effectively captures dynamic changes in the environment.

[0032] Step S21: Integration and processing of spatial structure and visual data. A weighted fusion method is used to integrate the preliminary spatial structure data and visual mapping data. Specifically, based on the high precision of the laser point cloud data, it is assigned a higher weight to determine the basic geometry of the environment, while the visual mapping data is used to supplement any missing color or texture information in the point cloud data. The weights can be dynamically adjusted based on sensor confidence or environmental conditions (such as illumination or occlusion). Through point-by-point matching and region partitioning, the two data sources are superimposed in the same coordinate system to form fused environmental data. During the fusion process, inconsistencies in the data are corrected using RANSAC or ICP algorithms, such as smoothing to eliminate boundary errors between the point cloud and the image. It should be noted that inconsistencies in the data are corrected during the fusion process, such as smoothing to eliminate boundary errors between the point cloud and the image.

[0033] Step S22: Screening and Confirmation of Dynamic Obstacle Locations. In the fused environmental data, if objects with dynamic characteristics are detected, their spatial locations are screened using a preset threshold to determine if they are located in critical areas. Specifically, critical areas are typically defined as areas near the robot's inspection path or areas that have a potential impact on task execution. By calculating the distance between the object and the critical area, if the distance is less than a preset value, it is marked as a dynamic obstacle requiring focused attention, generating screened obstacle distribution data. This process effectively reduces the complexity of subsequent calculations, focusing on obstacles that directly affect the robot's movement.

[0034] Step S23: Construction and updating of the 3D environment model. Based on the filtered obstacle distribution data, a 3D environment model containing the locations of dynamic obstacles is constructed. Specifically, a complete environment description framework is formed by storing the static structure and dynamic obstacle information from the fused environment data in layers. The static structure part describes the fixed features of the environment, such as walls or fixed equipment, while the dynamic obstacle part records the location and preliminary feature information of moving objects. Furthermore, through continuous updating of the environment model, adjustments are made in real time to the movement trajectories of dynamic obstacles, ensuring that the model reflects the latest state of the environment. This process provides reliable data support for subsequent obstacle identification and path planning.

[0035] Step S3: Obstacle Feature Extraction and Classification. Based on the constructed 3D environment model, obstacle features are extracted. A deep learning model is used to classify and identify objects in the environment, determining obstacle types and movement trajectories, and generating a classified obstacle dataset. Specifically, this process aims to distinguish between moving people and randomly placed goods in the environment, providing a basis for the formulation of subsequent obstacle avoidance strategies.

[0036] Step S31: Obstacle Feature Extraction and Preliminary Identification. Based on the data in the 3D environment model, features such as the geometric shape, color distribution, and dynamic displacement of obstacles are extracted. Geometric features are extracted through point cloud clustering and principal component analysis; color distribution features are obtained through HSV or RGB space statistics of image pixels; and dynamic displacement features are obtained through the rate of change of the object's centroid between consecutive frames. The extracted features are input into a pre-trained multi-task convolutional neural network (Multi-Task CNN), which includes a branch for classification and a branch for bounding box regression, and is trained on large public datasets (such as COCO or KITTI) and specific industrial scene datasets. Preliminary identification of objects in the environment is performed to determine the set of potential obstacle objects.

[0037] For a set of potential obstacle objects, if an object shows significant positional changes in consecutive frames (e.g., its centroid displacement exceeds 10cm for more than three consecutive frames), it is classified as a moving person, forming a dynamic object subset. For objects with no significant positional changes, a secondary judgment is made based on their geometric shape and color characteristics; if they match the typical characteristics of goods... Specifically, geometric features are determined by analyzing the curvature and density distribution of point cloud data, color distribution features rely on pixel value statistics in image data, and dynamic displacement features are obtained by comparing positional changes in consecutive frames of data. The extracted features are input into a pre-trained deep learning model to perform preliminary object identification in the environment and determine a set of potential obstacles. The key to this process lies in the comprehensiveness of feature extraction, ensuring the accuracy of subsequent classification.

[0038] Step S32: Differentiating between moving personnel and placed goods. For the set of potential obstacle objects, the distinguishing features are analyzed, and the objects are divided into dynamic and static categories. Specifically, if an object shows a significant positional change in consecutive frames, it is classified as a moving person, forming the dynamic object subset. For objects with no significant positional change, a secondary judgment is made based on their geometric shape and color features. If they match typical characteristics of goods, such as a regular rectangular shape or a specific color distribution, they are classified as placed goods, forming the static obstacle subset. It should be noted that historical data is used to verify the classification results during the classification process to reduce the possibility of misclassification.

[0039] Step S33: Motion Trajectory Analysis and Secondary Confirmation. For the dynamic object subset, the continuity characteristics of its motion trajectory are analyzed by combining historical location data. Specifically, by calculating the displacement vector and velocity changes of the object over a continuous time period, the pattern of its motion trajectory is determined, such as whether it is linear or curvilinear motion. For the static obstacle subset, secondary confirmation is performed. If the object's position remains unchanged over a long period and matches the characteristics of cargo, it is finally confirmed as cargo placement. This process effectively improves the reliability of classification and provides accurate obstacle information for subsequent path planning.

[0040] Step S34: Integration and Update of the Obstacle Dataset. After classifying dynamic and static obstacles, the type information of the two types of obstacles is integrated to construct a complete obstacle dataset. Specifically, the dataset records the spatial location, type identifier, and trajectory information of each obstacle, and the classification results are saved using data storage technology. Furthermore, the latest obstacle information is continuously fed back into the 3D environment model to ensure the real-time nature of the environmental information. This process lays the foundation for subsequent priority ranking and path planning.

[0041] In one embodiment, the implementation of the matching algorithm for the spatial registration process of laser point cloud data in step S11 can be further refined. For example, when a robot inspects a factory workshop environment, the point cloud data may contain a large number of repetitive structures, such as neatly arranged equipment or pipes. To improve registration accuracy, edge feature points in the point cloud can be extracted first, and the relative transformation relationship between point cloud frames can be determined by calculating the distance and angle relationship between edge points. This method can effectively reduce matching errors and ensure the accuracy of spatial structure data when facing complex environments.

[0042] In another embodiment, for the camera image feature extraction and spatial projection in step S12, the image processing strategy can be adjusted according to different lighting conditions. For example, in a factory workshop, the lighting may vary over time or in different areas, affecting the image feature extraction effect. Therefore, adaptive brightness adjustment technology can be introduced in the image preprocessing stage to dynamically adjust the pixel value range according to the overall brightness distribution of the image, thereby improving the robustness of feature extraction. This adjustment not only improves the quality of the visual mapping data but also provides more reliable visual information for subsequent multi-source fusion.

[0043] In one possible implementation, the fusion weights for multi-source information fusion in step S21 can be adjusted according to different inspection scenarios. For example, in a factory workshop, some areas may have incomplete laser point cloud data due to equipment obstruction. In this case, the weight of visual mapping data can be increased to supplement the missing parts of the point cloud data with color and texture information from the image. In open areas, however, laser point cloud data is usually more complete, and its weight can be appropriately increased to ensure the accuracy of spatial structure information. This dynamic adjustment method can adapt to different environmental conditions and improve the comprehensiveness and accuracy of the fused environmental data.

[0044] In another embodiment, regarding the dynamic obstacle location screening in step S22, the definition range of the critical area can be adjusted according to the urgency of the inspection task. For example, in scenarios where the inspection task needs to be completed quickly, the critical area range can be narrowed, focusing only on the area closer to the robot, thereby reducing computation and improving processing speed. In scenarios with high safety requirements, the critical area range can be expanded to include more potential obstacles in the screening scope, ensuring the safety of robot movement. This flexible area definition method can optimize the obstacle screening effect according to task requirements.

[0045] In one embodiment, the feature extraction strategy for obstacle features in step S31 can be adjusted according to the diversity of obstacle types. For example, in a factory workshop, moving personnel may wear work clothes of different colors, and placed goods may have different shapes and sizes. Therefore, a multi-dimensional analysis method can be introduced during feature extraction, simultaneously considering multiple features such as geometric shape, color distribution, and dynamic displacement, ensuring that the extracted features comprehensively reflect the characteristics of the obstacle. This multi-dimensional analysis significantly improves the accuracy of initial identification and provides a solid foundation for subsequent classification.

[0046] In another embodiment, the classification logic for distinguishing between moving personnel and placed goods in step S32 can be optimized based on the specific business scenarios of the factory workshop. For example, in some workshops, moving personnel may frequently move around in specific areas, while placed goods are usually stacked in fixed locations. Therefore, a zone division strategy can be introduced to divide the workshop into areas with frequent personnel activity and areas for storing goods, and the classification results can be weighted and corrected based on the attributes of the areas where obstacles are located. This business scenario-based classification optimization method effectively reduces the occurrence of misjudgments and improves the reliability of classification.

[0047] In one possible implementation, the trajectory analysis method in step S33 can be adjusted based on the path layout of the factory workshop. For example, within the workshop, the movement trajectory of personnel may be constrained by the layout of passageways or equipment, typically exhibiting relatively regular path characteristics. Therefore, path constraints can be introduced into the trajectory analysis, and the rationality of the trajectory can be judged by comparing the deviation between the actual trajectory and the preset path. This scenario-constraint-based analysis method can more accurately identify the movement patterns of dynamic objects, providing more targeted information for the subsequent development of obstacle avoidance strategies.

[0048] It should be noted that the above embodiments are all described in the context of factory workshop inspection, aiming to achieve efficient perception and safe movement of robots in complex environments through multi-source data fusion and dynamic environment modeling. Subsequent steps will further develop priority ranking, path planning, and motion control based on the classified obstacle dataset to form a complete inspection solution.

[0049] In one embodiment, for the sensor data acquisition in step S1, different sensor configuration schemes can be selected according to the complexity of the factory workshop environment. For example, in densely populated areas, the scanning frequency of the LiDAR can be increased to capture more detailed spatial structure information, while a high-resolution camera can be equipped to acquire richer visual features. In relatively open areas, the scanning frequency can be appropriately reduced to reduce the data processing burden. This method of adjusting the sensor configuration according to environmental characteristics can optimize resource allocation while ensuring perception accuracy.

[0050] In another embodiment, for the real-time map generation in step S2, the map update strategy can be optimized by combining historical data from the factory workshop. For example, during long-term inspections, the static structure of some areas may remain unchanged, while the appearance of dynamic obstacles follows certain patterns. Therefore, the static structure information in historical data can be used as a basic template, and real-time updates can be performed only on the dynamically changing parts, thereby reducing the amount of computation and improving the efficiency of map generation. This optimization method based on historical data significantly improves the stability of the system during long-term operation.

[0051] In one possible implementation, a multi-model collaborative classification mechanism can be introduced for obstacle classification and identification in step S3 to improve classification accuracy. For example, in a factory workshop, the characteristics of moving personnel and placed goods may vary due to environmental interference, and a single model may be insufficient to comprehensively cover all situations. Therefore, multiple deep learning models can be deployed simultaneously to perform specialized identification of features such as geometric shape, color distribution, and dynamic displacement. Then, a voting mechanism can be used to integrate the classification results of each model to form the final classification decision. This multi-model collaborative approach effectively improves the robustness of classification and adapts to diverse needs in complex environments.

[0052] In another embodiment, the update frequency for the 3D environment model update in step S23 can be adjusted according to the factory workshop's inspection cycle. For example, during periods of intensive inspection tasks, the model update frequency can be increased to capture subtle changes in dynamic obstacles in real time, ensuring the timeliness of environmental information. Conversely, during periods of fewer inspection tasks, the update frequency can be appropriately reduced to minimize system resource consumption. This dynamic adjustment of the update frequency can achieve a balance between resources and effectiveness under different task loads.

[0053] In one embodiment, regarding the obstacle dataset integration in step S34, the data storage structure can be optimized according to the inspection task type in the factory workshop. For example, in tasks requiring frequent obstacle information queries, the dataset can be stored in blocks according to spatial regions, prioritizing the loading of obstacle data near the robot's current position, thereby improving query efficiency. In tasks requiring comprehensive analysis of obstacle distribution, the dataset can be categorized and stored by type, facilitating rapid statistics on the quantity and distribution of different types of obstacles. This flexible data storage method significantly improves data access efficiency and provides more convenient support for subsequent path planning.

[0054] In another embodiment, regarding the multi-source information fusion in step S21, the parameter settings of the fusion algorithm can be adjusted according to the equipment layout characteristics of the factory workshop. For example, in densely populated equipment areas, laser point cloud data may have significant gaps due to occlusion. In this case, the fusion ratio of visual data can be increased to fill in the blank areas of the point cloud data with image information. Conversely, in sparsely populated equipment areas, the fusion ratio of point cloud data can be increased to fully utilize its high-precision spatial information. This parameter adjustment method based on scene characteristics effectively improves the quality of the fused data, providing a more reliable foundation for subsequent environmental modeling.

[0055] In one possible implementation, the obstacle feature extraction in step S31 can be optimized by incorporating the environmental noise characteristics of the factory workshop. For example, in a workshop, the operation of machinery may generate vibrations or noise, affecting the stability of the point cloud data. Therefore, the point cloud data can be smoothed before feature extraction to reduce the impact of noise on geometric features. Simultaneously, filtering techniques can be introduced in color feature extraction to eliminate the influence of illumination changes on the image data. This optimization method targeting environmental noise significantly improves the stability of feature extraction, providing more accurate data support for subsequent recognition.

[0056] In one embodiment, the classification threshold in step S32 can be adjusted based on the activity patterns of personnel in the factory workshop. For example, during peak working hours, when there are more moving personnel and higher activity frequency, the classification threshold for dynamic displacement can be appropriately lowered to classify more objects with slight displacement as moving personnel, thereby improving sensitivity to potential dynamic obstacles. During off-peak hours, the classification threshold can be appropriately increased to reduce the possibility of misclassification. This time-based threshold adjustment method effectively adapts to the environmental characteristics of different time periods, improving the adaptability of classification.

[0057] In another embodiment, the trajectory analysis in step S33 can be optimized by combining the safety zone division of the factory workshop with the trajectory judgment logic. For example, certain areas in the workshop may be designated as safety passages, where personnel typically move around. Therefore, the path information of the safety passages can be used as a reference template, and the trajectory can be compared with the template to determine whether it meets expectations. This trajectory analysis method based on safety zones can more accurately identify the movement intentions of dynamic objects, providing more valuable information for the formulation of subsequent obstacle avoidance strategies.

[0058] In one possible implementation, the data update strategy in step S34 can be adjusted based on the complexity of the factory workshop inspection task. For example, in scenarios with high task complexity, the frequency of data updates can be increased to capture subtle changes in obstacle location and type in real time, ensuring the accuracy of the environmental model. Conversely, in scenarios with low task complexity, the update frequency can be appropriately reduced to decrease system resource consumption. This dynamic adjustment of the update strategy can achieve a balance between effectiveness and efficiency under different task requirements.

[0059] In one embodiment, for the sensor data acquisition in step S1, the sensor operating parameters can be optimized based on the ambient temperature and humidity conditions of the factory workshop. For example, in a high-temperature and high-humidity environment, the performance of the LiDAR and camera may be affected. In this case, the scanning power of the LiDAR can be adjusted to enhance the signal strength, while anti-fogging treatment can be applied to the camera lens to ensure the clarity of the image data. This parameter optimization method based on environmental conditions effectively improves the reliability of data acquisition and provides higher quality raw data for subsequent processing.

[0060] In another embodiment, for real-time map generation in step S2, the map storage structure can be optimized based on the characteristics of the factory workshop's inspection routes. For example, in scenarios where inspection routes are relatively fixed, map data can be stored hierarchically by route segments, prioritizing the loading of map information related to the current route segment, thereby improving data access speed. In scenarios where inspection routes are flexible and varied, map data can be stored as a whole by region, facilitating quick switching between map information in different regions. This route-based storage optimization significantly improves the efficiency of map generation and use.

[0061] In one possible implementation, the obstacle classification and identification in step S3 can be adjusted based on the equipment operating status in the factory workshop. For example, during peak equipment operation periods, noise and vibration in the workshop may interfere with sensor data. In this case, the intensity of data preprocessing can be increased, and the impact of noise can be reduced by averaging multiple frames of data. At the same time, an anti-noise mechanism can be introduced into the classification model to improve its fault tolerance to interference data. This strategy adjustment method based on equipment operating status effectively improves the stability of classification and identification, adapting to the diverse challenges in complex environments.

[0062] In one embodiment, the spatial registration of point cloud data in step S11 can be optimized by considering the flatness characteristics of the factory floor. For example, in areas with high flatness, ground points can be preferentially used as reference points, and the relative positional relationship between point cloud frames can be quickly determined by calculating the distribution consistency of ground points. In areas with low flatness, the reliance on vertical structural feature points can be increased, and the registration accuracy can be improved by matching feature points on walls or equipment edges. This algorithm optimization method based on ground characteristics significantly improves the adaptability of spatial registration and ensures the accuracy of data processing.

[0063] In another embodiment, the image feature extraction strategy in step S12 can be adjusted according to the lighting distribution characteristics of the factory workshop. For example, in areas with uneven lighting distribution, local contrast enhancement technology can be introduced into image processing to highlight the feature details in dark areas, while suppressing brightness in overly bright areas to ensure the balance of feature extraction. This strategy optimization method based on lighting distribution effectively improves the quality of visual mapping data and provides more reliable visual information support for subsequent fusion.

[0064] In one possible implementation, the multi-source fusion in step S21 can be optimized by considering the equipment density distribution in the factory workshop. For example, in areas with high equipment density, the fusion accuracy of local area data can be increased by refining the matching range between point clouds and images, thereby improving the local consistency of the fused data. In areas with low equipment density, the matching range can be appropriately widened to prioritize the overall coverage of the fused data. This equipment density-based fusion optimization effectively balances accuracy and efficiency, providing a higher-quality data foundation for subsequent environmental modeling.

[0065] In one embodiment, the dynamic obstacle screening in step S22 can be adjusted according to the inspection task objectives of the factory workshop. For example, in tasks where safety is the primary objective, the screening threshold can be set more strictly to include more potential dynamic obstacles in the focus, ensuring the safety of robot movement. Conversely, in tasks where efficiency is the primary objective, the screening threshold can be appropriately relaxed to reduce unnecessary computational burden and accelerate processing. This task-objective-based screening optimization method can achieve a balance between effectiveness and efficiency under different needs.

[0066] In another embodiment, the 3D environment model construction in step S23 can be optimized by considering the spatial layout characteristics of the factory workshop. For example, in a workshop with a relatively regular spatial layout, the model can be divided into multiple regular grid units, and environmental information can be stored according to the grid, thereby improving data query efficiency. In a workshop with a more complex spatial layout, an irregular division method can be used to store environmental information according to the actual structural features, ensuring the model's descriptive accuracy. This spatial layout-based model optimization method significantly improves the efficiency of using the environment model and provides more convenient data support for subsequent processing.

[0067] In one possible implementation, the feature extraction strategy in step S31 can be optimized based on the obstacle type distribution in the factory workshop. For example, in areas dominated by mobile personnel, dynamic displacement features can be extracted first, and potential dynamic obstacles can be quickly identified by analyzing position changes in consecutive frames. In areas dominated by goods, geometric shape features can be extracted first, and the contour information of static obstacles can be accurately identified by analyzing the curvature distribution of point cloud data. This type-distribution-based extraction optimization method effectively improves the targeting of feature extraction and provides more accurate data support for subsequent classification.

[0068] In one embodiment, the classification strategy in step S32 can be adjusted based on the characteristics of the inspection time period in the factory workshop. For example, during nighttime inspections, when the number of moving personnel is smaller, the classification model can appropriately increase its sensitivity to dynamic displacement, classifying more slightly moving objects as moving personnel to ensure timely identification of potential dynamic obstacles. During daytime inspections, the sensitivity can be appropriately reduced to decrease the possibility of misclassification. This time-based classification optimization method effectively adapts to the environmental characteristics of different time periods, improving the accuracy of classification.

[0069] In another embodiment, the trajectory analysis in step S33 can be optimized based on the characteristics of the factory workshop's passageway width. For example, in areas with narrow passageways, the trajectories of moving personnel are usually more concentrated, and the detection accuracy of trajectory anomalies can be improved by narrowing the tolerance range of the trajectory analysis. Conversely, in areas with wider passageways, the tolerance range can be appropriately widened to reduce misjudgments of normal trajectory fluctuations. This passageway width-based analysis optimization significantly improves the adaptability of trajectory analysis, providing more reliable information support for subsequent obstacle avoidance.

[0070] In one possible implementation, the data integration in step S34 can be optimized by considering the frequency of inspection tasks in the factory workshop. For example, in scenarios with high task frequency, obstacle data related to the current task can be integrated first, and the latest information can be quickly loaded through a temporary caching mechanism to improve data processing speed. In scenarios with low task frequency, the data can be fully integrated to form a complete long-term record, facilitating subsequent analysis. This task frequency-based integration optimization method effectively balances real-time performance and comprehensiveness, providing more flexible data support for subsequent path planning.

[0071] In one embodiment, the data acquisition strategy in step S1 can be optimized based on the dynamic characteristics of the factory workshop environment. For example, in areas with frequent dynamic changes, the frequency of data acquisition can be increased to capture subtle changes in the environment in real time and ensure data timeliness. In areas with less dynamic change, the acquisition frequency can be appropriately reduced to decrease data redundancy and optimize resource allocation. This dynamic acquisition optimization method significantly improves the efficiency of data acquisition and provides higher-quality raw data for subsequent processing.

[0072] In another embodiment, the map generation strategy in step S2 can be optimized based on the size of the inspection area in the factory workshop. For example, in scenarios with a small inspection area, a high-precision map generation method can be used, increasing the number of data sampling points to improve the map's detail representation. In scenarios with a large inspection area, the sampling density can be appropriately reduced to prioritize overall map coverage and reduce computational load. This area-size-based generation optimization method effectively balances accuracy and efficiency, providing more suitable data support for subsequent environmental modeling.

[0073] In one possible implementation, the classification strategy in step S3 can be optimized based on the complexity of the factory environment. For example, in areas with high environmental complexity, the diversity of training data for the classification model can be increased by introducing more sample data from complex scenarios, thereby improving the model's adaptability to diverse features. Conversely, in areas with low environmental complexity, the scale of training data can be simplified, reducing model training time and optimizing resource utilization efficiency. This complexity-based optimization approach significantly improves the adaptability of the classification, meeting diverse needs in different environments.

[0074] In one embodiment, the spatial registration in step S11 can be optimized by considering the mobility characteristics of equipment in the factory workshop. For example, in areas with high equipment mobility, the removal of dynamic feature points can be intensified. By analyzing the temporal changes in point cloud data, interference points caused by equipment movement can be eliminated, improving the stability of registration. In areas with low equipment mobility, the removal intensity can be reduced, retaining more feature point information to ensure comprehensive registration. This mobility-based registration optimization method effectively improves the reliability of spatial registration, providing a more accurate data foundation for subsequent processing.

[0075] In another embodiment, the feature extraction strategy in step S12 can be optimized based on the image noise characteristics of the factory workshop. For example, in areas with high image noise, the filtering intensity in the preprocessing stage can be increased, and the impact of noise on feature extraction can be reduced through multi-layer filtering technology. Simultaneously, a fault-tolerance mechanism can be introduced in feature matching to improve robustness to noisy data. This noise-based extraction optimization method significantly improves the quality of visual mapping data, providing more reliable visual information support for subsequent fusion.

[0076] In one possible implementation, the data fusion strategy in step S21 can be optimized by considering the environmental occlusion characteristics of the factory workshop. For example, in areas with severe occlusion, reliance on historical data can be increased, filling in missing parts of the current data by calling previous fusion results, thus ensuring the continuity of the fused data. In areas with less occlusion, current data can be used first for fusion, reducing reliance on historical data and improving data real-time performance. This occlusion-based fusion optimization method effectively balances continuity and real-time performance, providing higher-quality data support for subsequent modeling.

[0077] In one embodiment, the obstacle screening in step S22 can be optimized based on the inspection speed characteristics of the factory workshop. For example, in scenarios with faster inspection speeds, the scope of critical areas can be appropriately reduced to decrease the amount of screening calculations, improve processing speed, and ensure real-time performance. Conversely, in scenarios with slower inspection speeds, the scope of critical areas can be expanded to increase attention to potential obstacles and improve safety. This speed-based screening optimization method can achieve a balance between effectiveness and efficiency under different operating conditions.

[0078] In another embodiment, the model update in step S23 can be optimized by considering the frequency of environmental changes in the factory workshop. For example, in areas with high environmental change frequency, the model update frequency can be increased to capture the latest location information of dynamic obstacles in real time, ensuring the accuracy of the model. In areas with low environmental change frequency, the update frequency can be appropriately reduced to decrease resource consumption and optimize system performance. This frequency-based update optimization significantly improves the adaptability of model updates and provides more reliable data support for subsequent processing.

[0079] In one possible implementation, the feature extraction strategy in step S31 can be optimized based on the obstacle density characteristics of the factory workshop. For example, in areas with high obstacle density, the precision of feature extraction can be increased by improving the sampling density of point cloud data and the resolution of image data to capture more detailed features and ensure comprehensive extraction. Conversely, in areas with low obstacle density, the precision can be appropriately reduced to decrease computation and increase processing speed. This density-based extraction optimization effectively balances accuracy and efficiency, providing more suitable data support for subsequent classification.

[0080] In one embodiment, the classification strategy in step S32 can be adjusted based on the characteristics of the inspection task cycle in the factory workshop. For example, in scenarios with short task cycles, the classification process can be simplified, prioritizing the identification of dynamic obstacles and reducing the time for secondary confirmation of static obstacles, thereby improving processing speed. In scenarios with long task cycles, the comprehensiveness of the classification process can be increased, with detailed analysis of all obstacles to ensure classification accuracy. This classification optimization method based on cycle characteristics can achieve a balance between effectiveness and efficiency under different task conditions.

[0081] In another embodiment, the trajectory analysis in step S33 can be optimized based on the environmental constraints of the factory workshop. For example, in areas with more environmental constraints, the reliance on trajectory constraints can be increased, and the accuracy of trajectory analysis can be improved by analyzing the limiting effect of channel boundaries or equipment positions on the trajectory. In areas with fewer environmental constraints, the reliance on constraints can be reduced, and the focus on trajectory continuity can be increased to ensure the comprehensiveness of the analysis. This constraint-based optimization method significantly improves the adaptability of trajectory analysis and provides more reliable information support for subsequent obstacle avoidance.

[0082] In one possible implementation, the data integration in step S34 can be optimized by considering the scale and characteristics of the factory workshop's inspection tasks. For example, in scenarios with large task scales, the parallel processing capability of data integration can be increased, and the speed and efficiency of data integration can be improved through distributed storage and computing technologies. In scenarios with smaller task scales, a centralized integration approach can be adopted to reduce system complexity and ensure the stability of data processing. This scale-based optimization approach effectively balances speed and stability, providing more efficient data support for subsequent path planning.

[0083] This invention continues to use a multi-source data fusion-based robot environment perception and path planning method, detailing the specific implementation methods for subsequent steps. The following content will focus on obstacle priority ranking, path planning, motion trajectory simulation, and control command generation to ensure the completeness and feasibility of the technical solution.

[0084] Step S4: Obstacle Priority Ranking and List Generation. Based on the classified obstacle dataset, the distance and relative speed between each obstacle and the robot's current position are calculated to determine its potential interference with the robot's movement, generating a priority-ranked obstacle list. Specifically, this process aims to identify obstacles that have the greatest impact on the robot's movement safety and efficiency, providing a priority basis for subsequent path planning.

[0085] Step S41: Distance and Relative Velocity Calculation. For each obstacle in the obstacle dataset, its spatial coordinates and motion information are extracted. Combined with the robot's real-time position data, the spatial distance between each obstacle and the robot is calculated. Specifically, the spatial distance is obtained by calculating the Euclidean distance between the two points, while the relative velocity is determined by analyzing the displacement changes of the obstacle and the robot over a continuous time period. This process generates preliminary distance and velocity results, providing a data foundation for subsequent priority determination.

[0086] Step S42, marking high-priority interference. Based on the calculated distance and velocity results, each obstacle is conditionally judged. If the distance between an obstacle and the robot is less than a preset threshold and the relative velocity is greater than zero, it indicates that the obstacle may approach the robot in a short time and pose a potential threat, so it is marked as high-priority interference. Specifically, the preset threshold can be dynamically adjusted according to the safety requirements of the inspection environment and the robot's movement speed to ensure the rationality of the marking. This process generates a set of marked obstacle classifications, laying the foundation for subsequent sorting.

[0087] Step S43: Generation and updating of the priority ranking list. Based on the labeled obstacle classification set, all obstacles are sorted according to the identifiers of high-priority interference, generating an initial priority ranking list. Subsequently, by combining dynamic updates of position and velocity data, the distance and velocity status of each obstacle are re-verified to determine whether the high-priority conditions are still met, generating an updated priority ranking result. Furthermore, for obstacles with high-priority interference, detailed motion trajectory data are obtained to analyze potential collision risks and determine the final priority ranking list. This process ensures the real-time nature and accuracy of the priority list.

[0088] Step S44: Dynamic Adjustment and Structured Output. Based on the final priority ranking list, and combined with changes in the real-time acquired robot position and obstacle dataset, fluctuations in distance and speed are continuously monitored to determine if priority adjustments are needed, generating a dynamically adjusted obstacle priority list. Subsequently, the adjusted list is converted into structured data output for use by the subsequent path planning module, ensuring a high degree of match between the execution order of the obstacle avoidance strategy and the current environmental state. This process provides dynamically adaptive priority information for path planning.

[0089] Step S5: Path planning and optimal path sequence generation. Based on a priority-ranked obstacle list, data is input into the path planning algorithm. Alternative paths are generated using the current map, and the optimal path sequence is determined by evaluating path length and obstacle avoidance probability. Specifically, this process aims to provide the robot with a safe and efficient movement path, ensuring the successful completion of the inspection task.

[0090] Step S51: Obstacle distribution map and alternative path set generation. Sorting data is extracted from the obstacle priority list and combined with map data from the current map library to generate an initial obstacle distribution map, reflecting the spatial distribution of obstacles in the environment. Subsequently, based on the obstacle distribution map, a pre-established path generation model is used to construct an alternative path set. Specifically, the path generation model generates multiple feasible paths based on the robot's current position and target position, ensuring that each path avoids high-priority obstacles. This process provides diverse options for subsequent path evaluation.

[0091] Step S52: Path Length and Obstacle Avoidance Probability Assessment. For each path in the candidate path set, calculate its total length and record the length data for each path. If the path length exceeds a preset threshold, further analyze the obstacle avoidance probability of each path. Specifically, the obstacle avoidance probability is determined by analyzing the closest distance to high-priority obstacles on the path and the predicted trajectory of the obstacles, ensuring path safety. This process generates length and obstacle avoidance probability data for each path, providing a basis for subsequent ranking.

[0092] Step S53, Path Sequence Sorting and Verification. Based on obstacle avoidance probability values ​​and path length values, the candidate path set is comprehensively sorted to determine a preliminary preferred path sequence. Subsequently, the path data in the preferred path sequence is verified a second time, combined with the dynamic information after the obstacle distribution map is updated, to determine whether the path sequence needs to be adjusted, generating the final preferred path sequence. If the obstacle avoidance probability value in the final preferred path sequence is lower than a preset threshold, the candidate path set is regenerated based on the latest data in the current map database, and the evaluation and sorting steps are repeated to determine a new preferred path sequence. This process ensures the safety and efficiency of the path sequence.

[0093] Step S6: Motion Trajectory Simulation and Path Correction. Based on the optimized path sequence, the robot's motion trajectory is simulated, and potential collision points are predicted. If a predicted point overlaps with an obstacle, the path parameters are adjusted to generate a corrected path sequence. Specifically, this process aims to verify the feasibility of the path through simulation and ensure the safety of the robot's movement.

[0094] Step S61, Trajectory Simulation and Collision Point Prediction. Based on the preferred path sequence, initial simulation data of the robot's motion trajectory is obtained. Using the particle filtering method, the robot's state is defined as position and velocity. With the preferred path sequence as the control input, multi-step prediction of the robot's future motion state is performed. At the same time, the motion model of obstacles in the environment is considered. By calculating the intersection of each particle on the robot's predicted trajectory with the predicted position of the obstacle, the position of the potential collision point is predicted.

[0095] Step S62: Collision Point Determination and Parameter Adjustment. Based on the prediction results, it is determined whether the predicted point overlaps with an obstacle. If overlap exists, key control points in the path parameters are extracted to determine the range of parameters that need adjustment. Subsequently, the parameters are optimized using preset adjustment rules to generate a first corrected path sequence. Specifically, the adjustment rules locally adjust the curvature or direction of the path based on the specific location of the collision point to ensure obstacle avoidance. This process effectively reduces the risk of collision.

[0096] Step S63: Trajectory Resimulation and Secondary Correction. Based on the first corrected path sequence, the robot's motion trajectory is resimulated to obtain the updated trajectory point distribution and determine if any new potential collision points exist. If collision points still exist, the local weights of the path parameters are adjusted for the new collision point locations to generate a second corrected path sequence. Subsequently, based on the second corrected path sequence, the final robot motion trajectory data is obtained, and the final trajectory is checked using a verification method to determine whether the collision risk has been eliminated. If no collision risk is found, the second corrected path sequence is output as the final path, generating control commands for the robot's motion. This process ensures the safety and executability of the path.

[0097] Step S7: Control command generation and execution monitoring. Based on the corrected path sequence, a control command sequence is generated and transmitted to the robot actuator. Execution deviations are monitored in real time to ensure the accuracy of command execution. Specifically, this process aims to translate the planned path into actual motion, ensuring that the robot completes the inspection task according to the expected path.

[0098] Step S71, Path Sequence Conversion and Command Transmission. Initial data is extracted from the corrected path sequence. Using a pre-established mapping model, the path sequence is converted into a control command sequence, generating a preliminary control command set. Subsequently, the command set is sent to the robot execution module through the transmission channel to collect state data in real time during execution and determine dynamic changes. This process provides a data foundation for subsequent monitoring.

[0099] Step S72 involves deviation calculation and instruction update. For dynamically changing data, a real-time monitoring mechanism is used to compare the data with preset execution standards, calculate the execution deviation value, and determine if the deviation exceeds a preset threshold. If the deviation exceeds the threshold, an instruction update process is triggered. Combining execution feedback data and path correction logic, an adjusted control instruction sequence is regenerated. Subsequently, the adjusted instruction sequence is used to update the instruction input of the robot's execution module, obtain new execution status data, and continuously monitor the execution effect. This process ensures the accuracy of instruction execution.

[0100] Step S73, iterative adjustment and historical record. Based on continuously monitored data, the deviation judgment process is executed iteratively. If the deviation still exceeds the preset threshold, the instruction update and adjustment steps are repeated until a stable execution result is obtained. After the execution result stabilizes, the correspondence between the final path sequence and control instructions is recorded and stored in the historical database for reference by subsequent tasks. This process provides data accumulation for long-term operation.

[0101] Step S8, Environmental Model Optimization and Path Iteration. Based on the updated control commands and fused sensor feedback data, the 3D environmental model is continuously optimized. Path planning is iterated using a path planning algorithm to generate a final inspection path adapted to the dynamic environment. Specifically, this process aims to ensure that path planning keeps pace with environmental changes through continuous optimization.

[0102] Step S81, Sensor Feedback and Environmental Perception Update. Real-time environmental information is collected through sensors, and feedback data from sensors and historical records is fused to construct an initial dataset, generating preliminary environmental perception results. Subsequently, combined with a pre-established 3D model, data alignment and correction operations are performed to determine the basis of the updated environmental model. This process ensures the real-time nature of the environmental model.

[0103] Step S82, Preliminary Path Calculation and Feasibility Analysis. Based on the updated environmental model and incorporating real-time adjustments to control commands, a preliminary path calculation is performed to obtain an initial inspection path plan. Subsequently, combined with dynamic environmental change information, a path feasibility analysis is executed. If obstacles are detected or environmental changes exceed preset thresholds, the path parameters are readjusted to generate an optimized path plan. This process ensures the feasibility of the path.

[0104] Step S83: Path Execution Condition Judgment and Final Output. Based on the optimized path plan, sensor feedback data is continuously collected, dynamic elements in the environmental model are updated, and it is determined whether the path execution conditions are met. If the conditions are met, the final inspection path data is generated based on the updated environmental model and the optimized path plan. If not, the process returns to the feasibility analysis stage to iteratively calculate the path plan. Finally, combined with real-time control commands, the execution details of the inspection path are dynamically adjusted to determine the real-time adaptive solution for the inspection task. This process ensures the dynamic adaptability of path planning.

[0105] In one embodiment, the distance and relative velocity calculations in step S41 can be optimized based on the inspection speed characteristics of the factory workshop. For example, in scenarios where the robot's inspection speed is relatively fast, the real-time requirements for distance calculation can be increased. By simplifying the calculation steps, preliminary results can be obtained quickly, ensuring the timeliness of subsequent processing. In scenarios where the inspection speed is relatively slow, the accuracy of the calculation can be increased. By comprehensively analyzing multi-frame data, the accuracy of distance and velocity results can be improved. This speed-based optimization method effectively balances real-time performance and accuracy.

[0106] In another embodiment, the marking criteria for high-priority interference markings in step S42 can be adjusted by combining the division of safety zones in the factory workshop. For example, in scenarios with a small safety zone, the distance threshold can be appropriately lowered to mark more obstacles approaching the robot as high priority, ensuring safety. In scenarios with a large safety zone, the threshold can be appropriately increased to reduce unnecessary markings and optimize the allocation of computing resources. This marking optimization method based on region division significantly improves the adaptability of priority judgment.

[0107] In one possible implementation, the priority ranking in step S43 can be optimized based on the obstacle density characteristics of the factory workshop. For example, in areas with high obstacle density, the ranking weight of high-priority obstacles can be increased, ensuring timely obstacle avoidance by prioritizing the closest and fastest obstacles. In areas with low obstacle density, distance and speed factors can be balanced to generate a more comprehensive ranking result. This density-based ranking optimization method effectively improves the relevance of the priority list.

[0108] In one embodiment, the dynamic adjustment in step S44 can be optimized by considering the frequency of environmental changes in the factory workshop. For example, in scenarios with a high frequency of environmental changes, the frequency of priority adjustments can be increased to capture subtle changes in obstacle positions and speeds in real time, ensuring the accuracy of the list. Conversely, in scenarios with a low frequency of environmental changes, the adjustment frequency can be appropriately reduced to decrease resource consumption. This frequency-based adjustment optimization significantly improves the efficiency of dynamic adjustments.

[0109] In another embodiment, the generation model for the candidate path set in step S51 can be optimized based on the characteristics of the factory workshop's passageway layout. For example, in scenarios with regular passageway layouts, paths along the passageway direction can be prioritized to reduce path curvature changes and improve movement efficiency. In scenarios with complex passageway layouts, path diversity can be increased by generating more curved paths to ensure the feasibility of avoiding obstacles. This layout-based generation optimization method effectively improves the quality of the candidate path set.

[0110] In one possible implementation, the path evaluation in step S52 can be optimized by incorporating the inspection task objectives of the factory workshop. For example, in tasks where efficiency is the primary objective, the evaluation weight of path length can be increased to prioritize shorter paths and improve inspection speed. Conversely, in tasks where safety is the primary objective, the evaluation weight of obstacle avoidance probability can be increased to ensure path safety. This task objective-based evaluation optimization method significantly improves the rationality of path selection.

[0111] In one embodiment, the path sequence verification in step S53 can be optimized based on the dynamic characteristics of the factory workshop environment. For example, in areas with frequent dynamic changes, the verification frequency can be increased to capture changes in obstacle distribution in real time and ensure the adaptability of the path sequence. In areas with less dynamic change, the verification frequency can be appropriately reduced to decrease computation. This dynamic verification optimization method effectively balances real-time performance and efficiency.

[0112] In another embodiment, the trajectory simulation in step S61 can be optimized by incorporating the robot's motion characteristics in a factory setting. For example, in scenarios with a large robot turning radius, the simulation accuracy of the turning path can be increased by refining the distribution of trajectory points to predict the location of potential collision points. Conversely, in scenarios with a small turning radius, the simulation process can be simplified, reducing computational load. This motion characteristic-based simulation optimization significantly improves the accuracy of predictions.

[0113] In one possible implementation, the parameter adjustments in step S62 can be optimized based on the obstacle movement patterns in the factory workshop. For example, in scenarios where obstacle movement patterns are relatively regular, the path direction parameter can be adjusted first to avoid obstacles by changing the direction of movement. In scenarios with complex obstacle movement patterns, both direction and speed parameters can be adjusted comprehensively to ensure the flexibility of obstacle avoidance. This adjustment and optimization method based on movement patterns effectively improves the targeting of path correction.

[0114] In one embodiment, the secondary correction in step S63 can be optimized by considering the complexity of the factory workshop inspection task. For example, in scenarios with high task complexity, the number of iterations for the secondary correction can be increased to ensure the safety of the trajectory through multiple adjustments. Conversely, in scenarios with low task complexity, the number of iterations can be reduced to simplify the correction process. This complexity-based optimization method significantly improves the efficiency of path correction.

[0115] In another embodiment, the mapping model for instruction generation in step S71 can be optimized based on the characteristics of the actuators in the factory workshop. For example, in scenarios where the actuator response speed is fast, the frequency of instruction generation can be increased to update control instructions in real time and ensure timely execution. Conversely, in scenarios where the actuator response speed is slow, the generation frequency can be appropriately reduced to decrease instruction redundancy. This actuator-characteristic-based generation optimization method effectively improves the adaptability of instruction generation.

[0116] In one possible implementation, the monitoring mechanism for execution monitoring in step S72 can be optimized by considering the environmental interference characteristics of the factory workshop. For example, in scenarios with strong environmental interference, smoothing processing of monitoring data can be added, and the impact of interference can be reduced by averaging multiple frames of data, ensuring the accuracy of deviation calculation. In scenarios with weak environmental interference, the processing can be simplified to improve monitoring speed. This monitoring optimization method based on interference characteristics significantly improves the reliability of execution monitoring.

[0117] In one embodiment, the adjustment strategy for the cyclic adjustment in step S73 can be optimized based on the task execution cycle of the factory workshop. For example, in scenarios with short execution cycles, the cyclic adjustment process can be simplified to quickly generate update instructions and ensure the continuity of execution. In scenarios with long execution cycles, the comprehensiveness of the adjustment process can be increased to ensure the accuracy of instruction updates. This adjustment optimization method based on cycle characteristics effectively balances speed and accuracy.

[0118] In another embodiment, the environmental perception update in step S81 can be optimized by considering the sensor configuration characteristics of the factory workshop. For example, in scenarios with a large number of sensors, the data fusion dimension can be increased to improve the comprehensiveness of the perception results by integrating more data sources. In scenarios with fewer sensors, the fusion process can be simplified to reduce computational load. This configuration-based update optimization significantly improves the efficiency of environmental perception.

[0119] In one possible implementation, the path feasibility analysis in step S82 can be optimized based on the environmental change patterns of the factory workshop. For example, in scenarios with relatively regular environmental change patterns, the analysis of major change areas can be prioritized, reducing the computational load on secondary areas and improving analysis speed. In scenarios with complex environmental change patterns, all areas can be analyzed comprehensively to ensure the full extent of feasibility. This change-pattern-based analysis and optimization method effectively improves the targeting of path analysis.

[0120] In one embodiment, the path execution condition judgment in step S83 can be optimized by combining the urgency of the factory workshop inspection tasks. For example, in scenarios with high task urgency, the execution conditions can be appropriately relaxed to prioritize the timeliness of path generation and ensure task progress. Conversely, in scenarios with low task urgency, the conditions can be strictly enforced to ensure the safety and reliability of the path. This urgency-based judgment optimization method significantly improves the adaptability of path output.

[0121] In another embodiment, the priority ranking in step S4 can be optimized based on the size of the inspection area in the factory workshop. For example, in scenarios with a large inspection area, more attention can be paid to distant obstacles, and the priority ranking logic can be adjusted by comprehensively considering the trends in distance and speed changes. In scenarios with a small inspection area, priority can be given to nearby obstacles, simplifying the computation of distant obstacles. This scale-based ranking optimization method effectively balances comprehensiveness and efficiency.

[0122] In one possible implementation, the path planning in step S5 can be optimized by incorporating the equipment operating status of the factory workshop into the planning algorithm. For example, during peak equipment operation periods, there may be more dynamic obstacles in the workshop. In this case, the obstacle avoidance weight in path planning can be increased, generating more backup paths to ensure flexibility in obstacle avoidance. Conversely, during off-peak equipment operation periods, the number of backup paths can be reduced, simplifying the planning process and improving computational efficiency. This operation status-based planning optimization method significantly improves the adaptability of path planning.

[0123] In one embodiment, the trajectory correction in step S6 can be optimized based on the width characteristics of the factory workshop's passageways. For example, in scenarios with narrow passageways, the precision of local path adjustments can be increased by refining the distribution of control points to ensure obstacle avoidance accuracy. Conversely, in scenarios with wide passageways, the adjustment precision can be appropriately relaxed to reduce computational load and increase correction speed. This width-based optimization method effectively improves the targeting of trajectory correction.

[0124] In another embodiment, the execution strategy for the instructions in step S7 can be optimized based on the task execution environment of the factory floor. For example, in scenarios with a more complex execution environment, the feedback frequency of instruction execution can be increased, and the accuracy of execution can be ensured by collecting status data in real time. Conversely, in scenarios with a simpler execution environment, the feedback frequency can be appropriately reduced to decrease resource consumption. This execution optimization method based on environmental characteristics significantly improves the efficiency of instruction execution.

[0125] In one possible implementation, the iteration strategy for the path iteration in step S8 can be optimized based on the frequency of inspection tasks in the factory workshop. For example, in scenarios with high task frequency, the frequency of path iteration can be increased to capture environmental changes in real time and ensure the adaptability of the path. In scenarios with low task frequency, the iteration frequency can be appropriately reduced to decrease computation and optimize resource allocation. This frequency-based iterative optimization method effectively balances real-time performance and efficiency.

[0126] In one embodiment, the distance calculation in step S41 can be optimized by considering the flatness characteristics of the factory workshop ground. For example, in scenarios with high ground flatness, the distance calculation process can be simplified by directly calculating the straight-line distance between two points, thus improving calculation speed. In scenarios with low ground flatness, consideration of terrain undulations can be added, and the path distance can be calculated in segments to improve the accuracy of the results. This flatness-based optimization method significantly improves the adaptability of distance calculation.

[0127] In another embodiment, the priority marking in step S42 can be optimized based on the activity patterns of personnel in the factory workshop. For example, during periods of high personnel activity, the distance threshold can be lowered to mark more potential dynamic obstacles as high priority, ensuring safety. Conversely, during periods of low personnel activity, the threshold can be appropriately increased to reduce the number of marks and optimize computational resources. This activity-based marking optimization method effectively improves the targeting of priority marking.

[0128] In one possible implementation, the path generation in step S51 can be optimized by incorporating the distribution of safety passages in the factory workshop. For example, in scenarios with densely distributed safety passages, paths along the safety passages can be prioritized to ensure safe movement. Conversely, in scenarios with sparsely distributed safety passages, the diversity of paths can be increased, enhancing obstacle avoidance flexibility by generating more alternative paths. This passage distribution-based optimization method significantly improves the quality of the candidate path set.

[0129] In one embodiment, the collision point prediction in step S61 can be optimized based on the obstacle movement speed characteristics of the factory workshop. For example, in scenarios where obstacles move at high speeds, the prediction lead time can be increased, and the timeliness of the prediction can be ensured by analyzing trajectory data over a longer period. Conversely, in scenarios where obstacles move at low speeds, the lead time can be appropriately reduced to simplify the prediction process. This speed-based prediction optimization method effectively improves the accuracy of collision point prediction.

[0130] In another embodiment, the transmission strategy for the instruction transmission in step S71 can be optimized based on the characteristics of the factory workshop's communication environment. For example, in scenarios with poor communication, a redundancy mechanism for instruction transmission can be added to ensure instruction integrity through multiple transmissions. Conversely, in scenarios with good communication, the transmission process can be simplified, redundant data reduced, and transmission speed increased. This communication environment-based transmission optimization significantly improves the reliability of instruction transmission.

[0131] In one possible implementation, the data fusion strategy in step S81 can be optimized based on the sensor data quality characteristics of the factory workshop. For example, in scenarios with high data quality, the reliance on current data can be increased, and the weight of historical data can be reduced to improve the real-time performance of the fusion result. Conversely, in scenarios with low data quality, the weight of historical data can be increased, and the stability of the fusion result can be ensured by integrating more data sources. This data quality-based fusion optimization method effectively balances real-time performance and reliability.

[0132] In one embodiment, the feasibility analysis in step S82 can be combined with an environmental update frequency optimization analysis method for the factory workshop. For example, in scenarios with a high environmental update frequency, the analysis frequency can be increased to capture environmental changes in real time and ensure the feasibility of the path. Conversely, in scenarios with a low environmental update frequency, the analysis frequency can be appropriately reduced to decrease computation and optimize resource allocation. This update frequency-based analysis optimization method significantly improves the efficiency of feasibility analysis.

[0133] In another embodiment, the output strategy for the final path output in step S83 can be optimized based on the inspection task objectives of the factory workshop. For example, in a task where safety is the primary objective, the path with the highest obstacle avoidance probability can be output first to ensure safe movement. In a task where efficiency is the primary objective, the shortest path can be output first to improve inspection speed. This task-objective-based output optimization method effectively improves the rationality of the path output. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic path planning method for inspection robots, characterized in that, include: Data is collected from the field by sensors, and laser scanning point clouds and camera images are processed. Multi-source information is fused to generate a real-time map and obtain a three-dimensional environment model that includes the location of dynamic obstacles. Based on the 3D environment model, obstacle features are extracted, and a deep learning model is used to identify several current obstacles, determine the obstacle type and movement trajectory, and obtain a classified obstacle dataset. Using the categorized obstacle dataset, the distance and relative speed between the obstacle and the robot's current position are calculated. If the distance is less than a preset threshold and the speed is greater than zero, it is marked as a high-priority obstacle, and a priority-ranked list of obstacles is obtained. The obstacle list, sorted by priority, is input into the path planning algorithm. The algorithm generates alternative paths based on the current map, evaluates the length and obstacle avoidance probability of each path, and obtains the optimal path sequence. By optimizing the path sequence, simulating the robot's motion trajectory, and using the particle filtering method to predict potential collision points, if the predicted point overlaps with the obstacle, the sequence parameters are adjusted to obtain a corrected path sequence. Based on the corrected path sequence, a control command sequence is generated and transmitted to the robot actuator. The execution deviation is monitored in real time to determine whether the deviation exceeds the threshold. If it does, the command is recalculated to obtain the updated control command. By updating control commands, integrating sensor feedback data, continuously optimizing the 3D environment model, and iterating path planning, an inspection path that adapts to the dynamic environment is finally obtained.

2. The dynamic path planning method for inspection robots according to claim 1, characterized in that: Data is collected from the field using sensors, and laser scanning point clouds and camera images are processed using a positioning and mapping algorithm. Multi-source information is fused to generate a real-time map, resulting in a 3D environment model that includes the locations of dynamic obstacles. By collecting data from the on-site environment through sensors, raw information including laser point clouds and camera images is obtained, and preliminary environmental perception data is acquired. Based on the collected laser point cloud, a positioning algorithm is used to spatially register the point cloud data, determine the position information in the point cloud, and form preliminary spatial structure data. By using camera images and combining mapping algorithms to extract features and spatially project the image content, visual mapping data corresponding to laser point clouds is obtained, and the basic content of real-time maps is constructed. By employing multi-source fusion technology, spatial structure data and visual mapping data are integrated and processed to obtain fused environmental data that includes the location of dynamic obstacles; If dynamic obstacles are detected in the fused environmental data, the obstacle locations are filtered by a preset threshold to determine whether the obstacles are in critical areas, and filtered obstacle distribution data is generated. Based on the filtered obstacle distribution data, a three-dimensional environment model containing the dynamic obstacle locations is constructed to obtain complete environmental description information; By continuously updating the 3D environment model and adjusting the movement trajectory of dynamic obstacles in real time, the final dynamic environmental information is determined.

3. The dynamic path planning method for inspection robots according to claim 1, characterized in that, Based on the 3D environment model, obstacle features are extracted, and a deep learning model is used to identify moving people and randomly placed goods, determine the obstacle type and movement trajectory, and obtain a classified obstacle dataset, including: An initial environmental model is constructed by acquiring 3D environmental data, and multi-source information is integrated using sensor data fusion technology to obtain preliminary environmental model data. Based on the preliminary environmental model data, obstacle features are extracted, and a pre-trained deep learning model is applied to perform preliminary identification of objects in the environment to determine the set of potential obstacle objects. For a set of potential obstacle objects, analyze the distinguishing features between moving people and placed goods. If dynamic displacement is detected, the object is classified as a moving person, and a subset of dynamic objects is obtained after classification. Based on the categorized subset of dynamic objects and combined with historical location data, the continuity characteristics of the motion trajectory are calculated, and the trajectory analysis method is used to determine the motion trajectory pattern of the dynamic objects. By using the motion trajectory pattern, static objects are reconfirmed. If the position of a static object does not change and it meets the characteristics of cargo, it is classified as cargo and a subset of static obstacles is determined. Based on the dynamic object subset and the static obstacle subset, obstacle type information is integrated, and classification and organization techniques are applied to construct a complete obstacle dataset. By using a complete obstacle dataset, the 3D environment model is updated, and data storage technology is used to save the classification results, thereby obtaining real-time updated environmental information.

4. The dynamic path planning method for inspection robots according to claim 1, characterized in that, Using the categorized obstacle dataset, the distance and relative velocity between the obstacle and the robot's current position are calculated. If the distance is less than a preset threshold and the velocity is greater than zero, it is marked as a high-priority obstacle, resulting in a priority-ranked list of obstacles, including: By extracting the coordinate and motion information of each obstacle from the obstacle dataset and combining it with real-time data on the robot's position, the spatial distance and relative velocity between each obstacle and the robot are calculated to obtain preliminary distance and velocity results. Based on the calculated distance and speed results, condition judgment is performed for each obstacle. If the distance is lower than a preset threshold and the relative speed is greater than zero, it is marked as a high-priority interference, and a set of marked obstacle classifications is obtained. Using the labeled obstacle classification set, all obstacles are sorted according to the identifier of high-priority interference to generate an initial priority sorting list; By combining the initial priority ranking list with dynamic updates of position and velocity data, the distance and velocity status of each obstacle are re-verified to determine whether the high priority conditions are still met, and the updated priority ranking results are obtained. Based on the updated priority ranking results, detailed motion trajectory data of obstacles with high priority interference are obtained, potential collision risks are analyzed, and the final priority ranking list is determined. By combining the final priority ranking list with the changes in the robot's position and obstacle dataset acquired in real time, the fluctuations in distance and speed are continuously monitored to determine whether the priority needs to be adjusted, and a dynamically adjusted obstacle priority list is obtained. A dynamically adjusted obstacle priority list is used to generate structured data output, which is then called by the subsequent path planning module to determine the execution order of the real-time obstacle avoidance strategy.

5. The dynamic path planning method for inspection robots according to claim 1, characterized in that, The priority list of obstacles is input into the path planning algorithm. Combined with the current map, alternative paths are generated. The length and obstacle avoidance probability of each path are evaluated to obtain the optimal path sequence, including: An initial obstacle distribution map is generated by retrieving sorted data from the obstacle priority list and combining it with map data sources in the current map library. Based on the obstacle distribution map and path planning method, a set of alternative paths is constructed using a pre-established path generation model; For each path in the candidate path set, a path evaluation method is used to calculate the path length value and record the length data of each path; If the path length exceeds the preset threshold range, the obstacle avoidance probability value is further analyzed by probability calculation method to obtain the obstacle avoidance probability data of each path; Based on the obstacle avoidance probability value and path length value, the candidate path set is comprehensively sorted using the sequence sorting method to determine the preliminary preferred path order; By performing secondary verification on the path data in the preferred path sequence and combining it with the dynamic information after the obstacle distribution map is updated, it is determined whether the path sequence needs to be adjusted, and the final preferred path sequence is obtained. If the obstacle avoidance probability value in the final preferred path sequence is lower than the preset threshold, then the candidate path set is regenerated by combining the latest map data source in the current map library, and the path evaluation method and sequence sorting method are executed cyclically to determine a new preferred path sequence.

6. The dynamic path planning method for inspection robots according to claim 1, characterized in that, By optimizing the path sequence and simulating the robot's trajectory, a particle filter method is used to predict potential collision points. If the predicted point overlaps with an obstacle, the sequence parameters are adjusted to obtain a corrected path sequence, including: By optimizing the path sequence, the initial simulation data of the robot's motion trajectory is obtained. The particle filtering method is used to estimate the distribution of trajectory points and obtain the prediction results of potential collision points. Based on the prediction results of potential collision points, determine whether the predicted points overlap with obstacles. If the predicted points overlap with obstacles, extract the key control points in the path parameters and determine the range of parameters that need to be adjusted. By extracting the range of path parameters, the parameters are optimized using preset adjustment rules to generate the first corrected path sequence; Based on the first corrected path sequence, the robot's motion trajectory is re-simulated to obtain the updated trajectory point distribution and determine whether there are any new potential collision points. If potential collision points still exist in the updated trajectory point distribution, the local weights of the path parameters are adjusted for the new collision point locations to obtain the second corrected path sequence. The final robot motion trajectory data is obtained through the second corrected path sequence. The particle filtering method is used to verify the final trajectory and determine whether the collision risk has been eliminated. If no collision risk is found during the final trajectory verification, the second corrected path sequence will be used as the final path output to generate control commands for the robot's movement.

7. The dynamic path planning method for inspection robots according to claim 1, characterized in that, Based on the corrected path sequence, a control command sequence is generated and transmitted to the robot actuator. Execution deviations are monitored in real time, and it is determined whether the deviation exceeds a threshold. If it does, the commands are recalculated to obtain updated control commands, including: Initial data is obtained from the path sequence, and the path sequence is transformed into a control command sequence using a pre-established mapping model to obtain a preliminary set of control commands. The control command set is sent to the robot execution module through the transmission channel, and the status data during the execution process is collected in real time to determine the dynamic changes during the execution process; For dynamically changing data, a real-time monitoring mechanism is used to compare the data with preset execution standards, calculate the execution deviation value, and determine whether the deviation exceeds the preset threshold. If the deviation exceeds the preset threshold, the instruction update process is triggered, and the adjusted control instruction sequence is regenerated by combining the execution feedback data and the path correction logic. By adjusting the control command sequence, the command input of the robot execution module is updated, new execution status data is obtained, and the execution effect is continuously monitored. Based on continuously monitored data, the deviation judgment process is executed cyclically. If the deviation still exceeds the preset threshold, the instruction update and adjustment steps are repeated to obtain stable execution results. After the execution results stabilize, the final path sequence and the correspondence between the control commands are recorded and stored in the historical database for reference in subsequent tasks.

8. The dynamic path planning method for inspection robots according to claim 1, characterized in that, By updating control commands and fusing sensor feedback data, the 3D environment model is continuously optimized. A path planning algorithm is used for iterative path planning to obtain the final inspection path adapted to the dynamic environment, including: By collecting real-time environmental information through sensors and fusing feedback data from sensors and historical records, an initial dataset is constructed to obtain preliminary environmental perception results. Based on the preliminary environmental perception results and combined with the pre-established 3D model, a Kalman filter is used to perform data alignment and correction operations to determine the basis of the updated environmental model. Based on the updated environmental model, and integrating real-time adjustments of control commands, the A* algorithm is applied to perform preliminary path calculations and obtain an initial inspection path scheme. For the initial inspection path plan, combined with the dynamic environment change information, OpenCV is used to perform path feasibility analysis. If obstacles or environmental changes exceed the preset threshold, the path parameters are readjusted to obtain an optimized path plan. By continuously collecting sensor feedback data through the optimized path scheme, updating the dynamic elements in the environmental model, and determining whether the path execution conditions are met; If the path execution conditions are met, the final inspection path data will be generated based on the updated environment model and the optimized path scheme. If the conditions are not met, return to the OpenCV-based path feasibility analysis stage and iteratively calculate the path scheme. After obtaining the final inspection path data, the execution details of the inspection path are dynamically adjusted in conjunction with real-time control commands to determine the real-time adaptive solution for the inspection task.

9. A dynamic path planning system for an inspection robot, characterized in that, The system includes: The data acquisition and modeling module is configured to collect on-site data through sensors, process laser scanning point clouds and camera images using a positioning mapping algorithm, fuse multi-source information to generate a real-time map, and obtain a three-dimensional environment model containing the location of dynamic obstacles. The obstacle recognition and classification module is configured to extract obstacle features based on a 3D environment model, use a deep learning model to identify moving people and randomly placed goods, determine the obstacle type and movement trajectory, and obtain a classified obstacle dataset. The priority evaluation module is configured to calculate the distance and relative speed between the obstacle and the robot's current position using the classified obstacle dataset. If the distance is less than a preset threshold and the speed is greater than zero, it is marked as a high-priority interference, and a priority-ranked list of obstacles is obtained. The path planning module is configured to take a priority list of obstacles and input it into the path planning algorithm. It then combines the current map to generate alternative paths, evaluates the length and obstacle avoidance probability of each path, and obtains the optimal path sequence. The collision prediction and correction module is configured to simulate the robot's motion trajectory by optimizing the path sequence, predict potential collision points using a particle filtering method, and adjust the sequence parameters to obtain a corrected path sequence if the predicted point overlaps with an obstacle. The control command generation and update module is configured to generate a control command sequence based on the corrected path sequence, transmit it to the robot actuator, monitor the execution deviation in real time, determine whether the deviation exceeds the threshold, and if it does, recalculate the command to obtain the updated control command. The environmental model optimization module is configured to continuously optimize the 3D environmental model by integrating sensor feedback data with updated control commands, and to obtain the final inspection path that adapts to the dynamic environment by iteratively planning the path using a path planning algorithm.

Images