AGV intelligent obstacle avoidance and path optimization method and system

Abstract

The invention relates to the technical field of AGV path planning, in particular to an intelligent obstacle avoidance and path optimization method and system for an AGV. The method comprises the following steps: obtaining an omnibearing image of an AGV activity area; performing static obstacle recognition on the omnibearing image to construct each static obstacle model; performing scene three-dimensional structure perception on the omni-directional image, and performing three-dimensional mapping based on each static obstacle model so as to obtain a global perception scene model; obtaining a cargo transportation planning log; performing cargo global available path planning on the global perception scene model according to a cargo transportation planning log, and constructing a global available planning path; and carrying out dynamic path optimization based on the global available planning path, and carrying out dynamic obstacle potential collision point mining according to the dynamically optimized transportation path so as to generate a collision time point and a collision position point. According to the invention, through dynamic and efficient path obstacle avoidance planning, the accuracy and stability of AGV driving are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of AGV path planning, and in particular to an intelligent obstacle avoidance and path optimization method and system for AGV. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, automated guided vehicles (AGVs), as an efficient logistics transportation tool, have been widely used in warehousing, production lines, logistics distribution and other fields. Through autonomous navigation and automatic control, AGVs can efficiently complete cargo handling tasks in complex working environments, significantly improving production efficiency. However, in actual applications, AGVs often face various complex and uncertain environmental challenges when performing tasks, especially the emergence of dynamic obstacles and path planning problems, which makes AGV's intelligent obstacle avoidance and path optimization a key technical issue that needs to be solved urgently.

[0003] Traditional AGV path planning and obstacle avoidance methods mainly rely on preset trajectories and simple environmental perception technologies, and usually rely on static maps and simple obstacle avoidance algorithms to achieve path selection. These methods can complete tasks well in relatively stable and simple environments, but they often have great limitations when facing dynamic environments, sudden obstacles and complex paths. Traditional methods are difficult to effectively respond to real-time changes in dynamic obstacles and cannot flexibly adjust paths according to changes in the environment, resulting in low efficiency and even collision accidents. In addition, planning methods based on fixed paths are also unable to adapt to sudden demands in complex environments, resulting in inflexible path planning and even affecting the overall work efficiency of AGVs.

[0004] In order to meet the requirements of modern intelligent warehousing and production lines for AGV efficiency and safety, traditional path planning and obstacle avoidance methods can no longer meet the complexity of practical applications. Currently, there is an urgent need for an intelligent obstacle avoidance and path optimization method that can perceive the environment in real time, accurately predict the behavior of dynamic obstacles, and optimize the path based on the dynamic environment. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes an intelligent obstacle avoidance and path optimization method and system for AGV to solve at least one of the above technical problems.

[0006] To achieve the above object, the present invention provides an intelligent obstacle avoidance and path optimization method for AGV, comprising the following steps: Step S1: Obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct a model of each static obstacle; Step S2: performing three-dimensional structure perception of the scene on the omnidirectional image, and performing three-dimensional mapping based on each static obstacle model, thereby obtaining a global perception scene model; Step S3: Obtain cargo transportation planning logs; perform cargo global available path planning on the global perception scene model according to the cargo transportation planning logs, and construct a globally available planning path; Step S4: Perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; Step S5: mining potential collision points of dynamic obstacles according to the dynamically optimized transport path to generate collision time points and location points; Step S6: Adaptive obstacle avoidance decision is made according to the collision time point and location point, and local path fine-tuning of the collision point is performed to construct a dynamic obstacle avoidance path optimization model.

[0007] By acquiring all-round images of the AGV activity area, the system can obtain environmental information from multiple angles and perspectives, providing comprehensive basic data for subsequent path planning and obstacle avoidance decisions. This image data acquisition method ensures that obstacles in every direction can be clearly captured. Through image processing technology and deep learning algorithms (such as convolutional neural networks (CNNs)), static obstacles in the image can be accurately identified, greatly improving the recognition accuracy. The identified static obstacles can generate a detailed model of each obstacle through information such as material and shape, providing a basis for subsequent path planning and obstacle avoidance algorithm precise positioning, accurately identifying and modeling static obstacles, allowing AGV to focus on path avoidance and physical property analysis of static obstacles, and at the same time providing dynamic obstacle detection. By sensing the three-dimensional structure of the scene, AGV can more accurately understand the complexity of the environment, and generate a real three-dimensional scene map by using depth sensors (such as lidar and depth cameras) and visual information fusion, so that AGV can plan paths and avoid obstacles in a more three-dimensional perspective. By mapping the static obstacle model into three-dimensional space, it ensures that AGV can accurately identify and avoid collisions with obstacles, and taking into account the impact of obstacles at different heights or positions, this model can help AGV better perform global path planning, not only limited to planar paths, but covering all angles of three-dimensional space. By obtaining the cargo transportation planning log, AGV can plan paths according to real-time task requirements, and different The transportation requirements of goods (such as destination, priority, size, etc.) will affect the path selection. To ensure that the path planning is task-oriented and customized, combined with the global perception scene model, AGV not only performs local path planning, but can find the optimal path in the entire area. This global path planning takes into account factors such as static obstacles, cargo demand and transportation efficiency, ensuring the maximization of transportation efficiency. Through global path planning, AGV can avoid blind spots and areas with concentrated obstacles, reduce the risk of collision, and ensure the safety of AGV during transportation. When AGV is performing tasks, the environment or obstacles will change. Through dynamic path optimization, AGV adjusts the path in real time to cope with environmental changes, such as new obstacles or obstacles in the site. In case of emergencies, compared with static path planning, dynamic path optimization ensures that AGV can adapt flexibly in complex environments. Through dynamic optimization, AGV selects a more efficient path according to the actual situation of the current path (such as traffic flow, obstacle distribution, etc.), reducing transportation time and energy consumption. The path optimization process not only optimizes the time cost, but also optimizes the space cost, making AGV more efficient during driving. Dynamic path optimization effectively avoids bottlenecks and congested areas in the environment, improves overall transportation efficiency, and avoids conflicts between multiple AGVs at the same location. By analyzing the behavior of dynamic obstacles in real time, the system can predict potential collision points in advance and generate collision time and location. This early warning mechanism provides AGV with sufficient time and space to make adjustments.Avoid collision accidents. The potential collision point mining of dynamic obstacles is based on real-time data and environmental feedback, which can accurately calculate the time and location of the collision. Through real-time tracking of AGV motion trajectory and dynamic obstacles, the system can respond quickly. This link increases the reliability and robustness of the AGV system in a dynamic environment, making it more stable in complex practical application scenarios. Based on the collision time and location, the system can make adaptive obstacle avoidance decisions in real time, and calculate the best obstacle avoidance path to ensure the smooth operation of the AGV. The AGV no longer relies on a preset fixed path, but can flexibly adjust the driving route according to real-time environmental changes. Compared with traditional obstacle avoidance methods, local path fine-tuning can fine-tune the path, avoid unnecessary path changes, and improve the accuracy and stability of AGV driving. This fine-tuning avoids large-scale path re-planning, saving computing resources and time. As the AGV continues to travel and avoid obstacles, the system optimizes the path planning algorithm through feedback learning, making it more adaptable in different environments. The constantly adjusted obstacle avoidance model allows the AGV to avoid obstacles more efficiently and intelligently in each operation, improving the overall system intelligence level.

[0008] In this specification, an intelligent obstacle avoidance and path optimization system for AGV is provided, which is used to execute the intelligent obstacle avoidance and path optimization method for AGV as described above, including: The obstacle recognition module is used to obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct each static obstacle model; A global perception module, used to perceive the three-dimensional structure of the scene in the omnidirectional image, and perform three-dimensional mapping based on each static obstacle model, so as to obtain a global perception scene model; The available path planning module is used to obtain the cargo transportation planning log; perform global available path planning for cargo on the global perception scene model according to the cargo transportation planning log, and construct a global available planning path; The path optimization module is used to perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; The collision prediction module is used to mine potential collision points of dynamic obstacles according to the dynamic optimized transportation path to generate collision time and location points; The local path fine-tuning module is used to make adaptive obstacle avoidance decisions based on the collision time and location points, and to fine-tune the local path at the collision point, thereby building a dynamic obstacle avoidance path optimization model.

[0009] Through omnidirectional image acquisition, the present invention enables AGV to obtain environmental data from different angles in real time, avoiding the blind spots of traditional perception systems, thus ensuring comprehensive perception of the entire activity area. The recognition of static obstacles is based on computer vision and deep learning algorithms (such as convolutional neural networks (CNNs)), which can efficiently and accurately distinguish various static obstacles from images, such as walls, equipment, shelves, etc. The recognition accuracy improves the quality of path planning and obstacle avoidance decisions, and generates a detailed geometric model for each static obstacle, including information such as the shape, size, and material of the obstacle, providing important data support for subsequent three-dimensional mapping and path planning. Through the three-dimensional structure perception of the scene, AGV can more accurately understand the spatial layout of the environment, including the height of the obstacle. , distribution and distance information. Through stereoscopic perception, AGV can avoid not only planar obstacles, but also uneven objects. Combined with the static obstacle model, AGV can build a detailed environmental map in three-dimensional space. This process not only includes the location of obstacles, but also can represent the spatial structure of the site, so that path planning can make efficient decisions in complex environments. The global perception scene model provides richer and more accurate environmental information for path planning, helping AGV to better identify road capacity, optimal paths and potential obstacles, and improve the accuracy and reliability of path planning. By obtaining the cargo transportation planning log, the system can customize path planning for AGV according to specific transportation tasks and goals, the type, weight and destination of the cargo. Factors such as the environment and the speed of the vehicle will affect the selection and optimization of the path. Combined with the global perception scenario model, the path planning module can take into account all obstacles in the environment, road conditions, traffic flow and other factors to develop the most suitable global path. This not only reduces the redundant part of the path, but also ensures the efficiency and safety of the transportation process. During the operation of the AGV, the environment changes (the emergence of new obstacles or temporary blockages on the path). The path optimization module can dynamically adjust the global planning path to cope with these changes and ensure the efficient completion of the transportation task. Through dynamic optimization, the system can select the optimal path based on real-time traffic conditions, cargo demand, energy consumption and other factors, reduce unnecessary detours and time waste, and reduce transportation costs. By analyzing dynamic The position, speed and movement trajectory of obstacles, the collision prediction module can evaluate the potential collision points with the AGV path in real time, which provides timely warning for the AGV and helps it make avoidance decisions in advance. Based on collision prediction, the AGV avoids high-risk areas in advance to avoid accidental collisions. Avoiding collisions not only ensures the safety of equipment and goods, but also reduces the maintenance cost of the system. Through accurate collision point prediction, the AGV can maintain a high stability during real-time transportation and avoid path interruptions or unexpected downtime. Based on the real-time predicted collision point information, the local path fine-tuning module quickly makes adaptive obstacle avoidance decisions according to the current position and motion state of the AGV. The AGV can flexibly choose strategies such as stopping, detouring or adjusting speed to avoid collisions.Local path fine-tuning can minimize unnecessary path adjustments and only make fine-tuning when necessary, thus avoiding large-scale path planning modifications and ensuring the continuity and stability of transportation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A schematic diagram of the steps of an intelligent obstacle avoidance and path optimization method for AGV of the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Detailed implementation flow chart of step S3. DETAILED DESCRIPTION

[0011] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0012] The present application example provides an intelligent obstacle avoidance and path optimization method and system for AGV. The execution subject of the intelligent obstacle avoidance and path optimization method and system for AGV includes but is not limited to: mechanical equipment, data processing platform, cloud server node, network upload device, etc. equipped with the system can be regarded as the general computing node of the present application, and the data processing platform includes but is not limited to: at least one of audio and image management system, information management system, and cloud data management system.

[0013] See also Figures 1 to 4 The present invention provides an intelligent obstacle avoidance and path optimization method for AGV, and the intelligent obstacle avoidance and path optimization method for AGV comprises the following steps: Step S1: Obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct a model of each static obstacle; Step S2: performing three-dimensional structure perception of the scene on the omnidirectional image, and performing three-dimensional mapping based on each static obstacle model, thereby obtaining a global perception scene model; Step S3: Obtain cargo transportation planning logs; perform cargo global available path planning on the global perception scene model according to the cargo transportation planning logs, and construct a globally available planning path; Step S4: Perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; Step S5: mining potential collision points of dynamic obstacles according to the dynamically optimized transport path to generate collision time points and location points; Step S6: Adaptive obstacle avoidance decision is made according to the collision time point and location point, and local path fine-tuning of the collision point is performed to construct a dynamic obstacle avoidance path optimization model.

[0014] By acquiring all-round images of the AGV activity area, the system can obtain environmental information from multiple angles and perspectives, providing comprehensive basic data for subsequent path planning and obstacle avoidance decisions. This image data acquisition method ensures that obstacles in every direction can be clearly captured. Through image processing technology and deep learning algorithms (such as convolutional neural networks (CNNs)), static obstacles in the image can be accurately identified, greatly improving the recognition accuracy. The identified static obstacles can generate a detailed model of each obstacle through information such as material and shape, providing a basis for subsequent path planning and obstacle avoidance algorithm precise positioning, accurately identifying and modeling static obstacles, allowing AGV to focus on path avoidance and physical property analysis of static obstacles, and at the same time providing dynamic obstacle detection. By sensing the three-dimensional structure of the scene, AGV can more accurately understand the complexity of the environment, and generate a real three-dimensional scene map by using depth sensors (such as lidar and depth cameras) and visual information fusion, so that AGV can plan paths and avoid obstacles in a more three-dimensional perspective. By mapping the static obstacle model into three-dimensional space, it ensures that AGV can accurately identify and avoid collisions with obstacles, and taking into account the impact of obstacles at different heights or positions, this model can help AGV better perform global path planning, not only limited to planar paths, but covering all angles of three-dimensional space. By obtaining the cargo transportation planning log, AGV can plan paths according to real-time task requirements, and different The transportation requirements of goods (such as destination, priority, size, etc.) will affect the path selection. To ensure that the path planning is task-oriented and customized, combined with the global perception scene model, AGV not only performs local path planning, but can find the optimal path in the entire area. This global path planning takes into account factors such as static obstacles, cargo demand and transportation efficiency, ensuring the maximization of transportation efficiency. Through global path planning, AGV can avoid blind spots and areas with concentrated obstacles, reduce the risk of collision, and ensure the safety of AGV during transportation. When AGV is performing tasks, the environment or obstacles will change. Through dynamic path optimization, AGV adjusts the path in real time to cope with environmental changes, such as new obstacles or obstacles in the site. In case of emergencies, compared with static path planning, dynamic path optimization ensures that AGV can adapt flexibly in complex environments. Through dynamic optimization, AGV selects a more efficient path according to the actual situation of the current path (such as traffic flow, obstacle distribution, etc.), reducing transportation time and energy consumption. The path optimization process not only optimizes the time cost, but also optimizes the space cost, making AGV more efficient during driving. Dynamic path optimization effectively avoids bottlenecks and congested areas in the environment, improves overall transportation efficiency, and avoids conflicts between multiple AGVs at the same location. By analyzing the behavior of dynamic obstacles in real time, the system can predict potential collision points in advance and generate collision time and location. This early warning mechanism provides AGV with sufficient time and space to make adjustments.Avoid collision accidents. The potential collision point mining of dynamic obstacles is based on real-time data and environmental feedback, which can accurately calculate the time and location of the collision. Through real-time tracking of AGV motion trajectory and dynamic obstacles, the system can respond quickly. This link increases the reliability and robustness of the AGV system in a dynamic environment, making it more stable in complex practical application scenarios. Based on the collision time and location, the system can make adaptive obstacle avoidance decisions in real time, and calculate the best obstacle avoidance path to ensure the smooth operation of the AGV. The AGV no longer relies on a preset fixed path, but can flexibly adjust the driving route according to real-time environmental changes. Compared with traditional obstacle avoidance methods, local path fine-tuning can fine-tune the path, avoid unnecessary path changes, and improve the accuracy and stability of AGV driving. This fine-tuning avoids large-scale path re-planning, saving computing resources and time. As the AGV continues to travel and avoid obstacles, the system optimizes the path planning algorithm through feedback learning, making it more adaptable in different environments. The constantly adjusted obstacle avoidance model allows the AGV to avoid obstacles more efficiently and intelligently in each operation, improving the overall system intelligence level.

[0015] In the embodiment of the present invention, refer to Figure 1 , is a schematic flow chart of a method of the present invention. In this example, the steps of the method include: Step S1: Obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct a model of each static obstacle; In this embodiment, a suitable omnidirectional camera or 360-degree camera (such as Ricoh Theta V or Insta360 One X) is selected. These devices can provide a panoramic view and help capture details of the AGV activity area. Ensure that the camera has sufficient resolution (at least 4K) to obtain clear image data during subsequent processing. Set it up in the AGV activity area and ensure that the ambient light is sufficient and there are no strong shadows to reduce image noise. It is necessary to consider dynamic factors in the environment, such as people and other moving objects, and try to collect images during off-peak hours. Determine the installation position and angle of the camera, usually in the center or a higher position of the AGV to obtain a more comprehensive perspective. Start the omnidirectional camera and start image acquisition. Set the image acquisition frequency, such as capturing an image every 1 second, so that enough data is collected. The image should contain all static features of the AGV activity area. Record the time and location of image acquisition to ensure that subsequent processing can perform data analysis based on the timestamp. Preprocess the collected omnidirectional images, including denoising, enhancement, and cropping. Use image processing libraries (such as OpenCV) to perform image denoising and contrast enhancement to improve the accuracy of subsequent recognition. Convert the omnidirectional image to a format suitable for model input, such as scaling the image to a specified size (e.g., 640x480 pixels). Select a suitable static obstacle recognition model, such as YOLO (You Only Look Once) or Faster R-CNN. Both models can detect and recognize objects in the image in real time. Train the model using annotated datasets (e.g., COCO or custom datasets) to improve its recognition of specific static obstacles (e.g., walls, pillars, equipment, etc.). During training, hyperparameters such as learning rate (0.001) and batch size (16) should be set to improve model performance. Input the preprocessed omnidirectional image into the trained static obstacle recognition model for forward inference to obtain recognition results. The model will output the bounding box, category label, and confidence score of each detected static obstacle. Set a confidence threshold (e.g., 0.5) and only keep detection results above this threshold to reduce false detections. For each identified static obstacle, build a corresponding static obstacle model based on its bounding box and category information. Each model should contain information such as the location, size, and category of the obstacle. Store the constructed static obstacle model in the database for subsequent query and analysis. Make sure to record the recognition time, location coordinates and category of each obstacle for subsequent processing.

[0016] Step S2: performing three-dimensional structure perception of the scene on the omnidirectional image, and performing three-dimensional mapping based on each static obstacle model, thereby obtaining a global perception scene model; In this embodiment, omnidirectional images are collected. These images should be high-resolution and minimize blur and distortion during the acquisition process. Ensure that the image coverage is wide enough for subsequent three-dimensional reconstruction. Organize the omnidirectional images into a format suitable for processing (such as JPEG or PNG), and include timestamps and location identifiers in the file name for subsequent management. Select a suitable three-dimensional reconstruction algorithm, such as Structure from Motion (SfM) or Multi-View Stereo (MVS). SfM recovers the three-dimensional structure of the scene from multi-angle images, while MVS is used to refine the reconstructed surface details. Set algorithm parameters, such as feature point extraction method (such as SIFT or ORB), feature matching threshold, and reconstruction accuracy (such as 1 cm). Use the selected feature point extraction algorithm (such as SIFT or ORB) to process the omnidirectional images and extract key points and descriptors. Ensure that the extracted feature points are robust to scene changes. Set extraction parameters, such as the number of feature points extracted per image (such as 500), to ensure feature diversity and coverage. Match the extracted feature points and use a feature matching algorithm (such as FLANN or BFMatcher) to calculate the similarity of feature points between adjacent images. Set a matching threshold (such as 0.75) to ensure the accuracy of the matching results. Filter out false matches and use the RANSAC algorithm to improve the matching accuracy and eliminate outliers. Generate a preliminary 3D point cloud using triangulation based on the matched feature points and camera parameters. Ensure that the density and fineness of the point cloud meet the requirements of subsequent processing (such as 10,000 points / square meter). Store the generated point cloud data in an appropriate format (such as PCL or PLY) for subsequent visualization and analysis. Integrate the data (such as position, size, and category) of each static obstacle model constructed in step S1. These models should be aligned with the 3D point cloud data through a coordinate system for mapping. Record the relevant information of each obstacle model, including its relative position and orientation in 3D space. Map the static obstacle model to the generated 3D point cloud. Use the transformation matrix to translate and rotate each obstacle model to ensure its accurate position and posture in the point cloud. Set the accuracy and tolerance of the mapping to ensure that no significant errors (such as less than 5 cm) are introduced during the mapping process. Combine all 3D point cloud data and static obstacle models to generate a global perception scene model. Use 3D visualization software (such as Blender or MeshLab) for final model rendering and optimization. Ensure that the global perception scene model can be viewed from different perspectives and accurately reflects the spatial layout of the AGV activity area.

[0017] Step S3: Obtain cargo transportation planning logs; perform cargo global available path planning on the global perception scene model according to the cargo transportation planning logs, and construct a globally available planning path; In this embodiment, the data source of the cargo transportation planning log is determined, which is usually the transportation management system (TMS) or logistics management system (WMS) within the enterprise. These systems record the detailed information of all transportation tasks, including cargo type, quantity, transportation time and route, etc. Data is obtained through API interface or database query to ensure the real-time and accuracy of data. Use SQL query or API call to extract cargo transportation planning log. Set extraction conditions, such as time range (for example, the past week) and transportation status (pending transportation, transportation). The important fields extracted include: cargo ID, cargo type, starting location, target location, estimated transportation time and priority, etc. Ensure that the previously constructed global perception scene model (step S2) can be combined with the location information in the cargo transportation planning log. The global perception scene model should contain the location and size information of static obstacles. Convert the scene model into a format suitable for path planning algorithm processing (such as a grid map or a graphic model). Select a path planning algorithm, commonly used algorithms include A* algorithm, Dijkstra algorithm or rapidly expanding random tree (RRT). The selected algorithm should be able to perform efficient path planning based on the scene model and cargo transportation needs. Set the parameters of the algorithm, such as heuristic functions (such as Manhattan distance or Euclidean distance) and the accuracy of the path search (such as accurate to 0.1 meters). Input the starting and target locations of the goods into the selected path planning algorithm, and perform path search in combination with the global perception scene model. The algorithm will generate an optimal path based on obstacles and environmental conditions. During the planning process, consider the characteristics of the goods (such as size, weight) and transportation requirements (such as priority, transportation time). Set weight parameters to balance the influence of different factors in path planning. The generated path should be a series of coordinate points that represent the driving trajectory of the AGV in the global perception scene. Post-process the generated path to ensure that the path is smooth and meets the kinematic characteristics of the AGV (such as maximum turning radius and acceleration). Set path smoothing parameters and use methods such as B-splines or Bezier curves to optimize the path to eliminate unnecessary sharp angles or sharp turns. Store the generated globally available planned paths in a database or file, and record the detailed information of each path, including the starting position, target position, path point coordinates, and path length. Assign a unique identification ID to each path for subsequent query and management. Use visualization tools (such as Matplotlib or 3D visualization software) to display the generated globally available planning path. Make sure the path is clearly visible in the global perception scene model for easy verification and analysis.

[0018] Step S4: Perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; In this embodiment, a variety of sensors (such as laser radar, camera and ultrasonic sensor) are installed on the AGV to monitor the changes in the surrounding environment in real time. Ensure that the installation position of the sensor can cover the driving path of the AGV, and the data collection frequency is set to 10Hz to achieve real-time monitoring. The collected data includes the location, speed, direction and environmental conditions (such as road conditions, weather, etc.) of dynamic obstacles, which will provide a basis for subsequent dynamic path optimization. Ensure that the previously generated global available planning path data is available for dynamic optimization. The path data should contain information such as the coordinates of each path point, the estimated driving time and the path length. Import the path data into the dynamic path optimization system for further processing. Select a suitable dynamic path optimization algorithm, such as the dynamic A algorithm, RRT or model-based predictive control (MPC). These algorithms can dynamically adjust the path of the AGV based on real-time sensor data to cope with environmental changes. Determine algorithm parameters, such as the heuristic function in the dynamic A algorithm, the expansion step size of the RRT, and the control time step size of the MPC (such as 0.5 seconds). Set the goals of dynamic path optimization, such as minimizing driving time, avoiding collisions, and reducing energy consumption. Assign different weights to each goal so that the impact of each goal is balanced during the optimization process. Define the AGV's kinematic constraints (such as maximum speed, acceleration, and turning radius) to ensure that the optimized path is feasible in actual operation. Also set environmental constraints, such as the minimum safe distance of dynamic obstacles. During the AGV's driving process, obtain sensor data in real time to identify dynamic obstacles and their movement trajectories. Input this information into the dynamic path optimization algorithm to calculate the new optimal path. The dynamic A* algorithm will update the path in each time step, taking into account the new obstacle position and expected movement, and generate new path points. Smooth the generated new path to ensure the continuity and feasibility of the path. Use methods such as B-splines or Bezier curves for smoothing to eliminate sharp corners and unnecessary turns in the path.

[0019] Step S5: mining potential collision points of dynamic obstacles according to the dynamically optimized transport path to generate collision time points and location points; In this embodiment, sensors installed on the AGV (such as laser radar, camera, and ultrasonic sensor) are used to monitor dynamic obstacles in the surrounding environment in real time. Ensure that the acquisition frequency of the sensor is 10Hz in order to obtain sufficiently accurate obstacle position and speed data. The collected data should include the current position, speed, direction of movement, and size information of the obstacle to provide a basis for subsequent collision point mining. Format the collected dynamic obstacle data into a structured format (such as JSON or CSV) for subsequent analysis. The data should include information such as obstacle ID, position coordinates (x, y, z), velocity vector (vx, vy, vz), and timestamp. Set a time window (such as 1 second) to collect the state changes of the obstacle within the time window for dynamic analysis. Select a suitable collision prediction algorithm, such as a model based on time and distance, or use a machine learning method (such as a support vector machine or a neural network) to train a collision detection model. Determine the parameters of the algorithm, such as the time prediction range (such as 3 seconds) and the distance threshold (such as 1 meter), so as to make effective collision predictions on the path. If machine learning methods are used, historical data sets need to be prepared, including the relative position, speed, and collision cases of dynamic obstacles and AGVs. These data are used to train the model to improve the accuracy of collision prediction. Hyperparameters such as learning rate (0.001), training batch size (32), and number of iterations (1000) are set to optimize model performance. The real-time monitored dynamic obstacle data is compared with the dynamically optimized transportation path of the AGV to calculate the relative position and motion trajectory of the AGV and each obstacle. Within a set time window (such as 3 seconds), the relative motion between the AGV and the dynamic obstacle is analyzed frame by frame to identify potential collision time points and location points. If the calculated collision time is less than the set threshold (such as 2 seconds), the time point and the corresponding location coordinates are recorded. The identified collision time points and location points are recorded in the database, ensuring that each record contains the timestamp, location coordinates, and the relevant obstacle ID. A unique identifier is assigned to each potential collision point for subsequent query and analysis. The data format should be structured to facilitate subsequent data processing and visualization.

[0020] Step S6: Adaptive obstacle avoidance decision is made according to the collision time point and location point, and local path fine-tuning of the collision point is performed to construct a dynamic obstacle avoidance path optimization model.

[0021] In this embodiment, the collision time point and location point data generated in step S5 are integrated to create a data structure (such as a dictionary or database table), which contains detailed information of each potential collision point, including timestamp, location coordinates and related obstacle information. The priority of each collision point is set so that a higher risk collision point can be quickly responded to during adaptive obstacle avoidance decision-making. A threshold is set. If the collision probability exceeds 0.7, the collision point is processed first. A suitable obstacle avoidance decision model is selected, such as a rule-based decision system or a reinforcement learning model. The rule-based system is simple and effective and suitable for real-time decision-making, while the reinforcement learning model can perform better in complex environments. The input parameters of the decision model are set, including the AGV when The AGV uses the previous state (such as position, speed), environmental information (such as obstacle position, path) and collision point information. During the AGV's driving process, the current state and environmental information are monitored in real time and input into the adaptive obstacle avoidance decision model. The model calculates the obstacle avoidance strategy based on the current state and collision point information and sets the decision threshold. When the distance to the collision point is less than the set value (such as 1.5 meters), the AGV must take action immediately and adjust the driving speed or direction of the AGV according to the obstacle avoidance strategy. According to the obstacle avoidance strategy output by the decision model, a new temporary target point is generated, and a new driving path is calculated. The path planning algorithm (such as A* or Dijkstra) is used to recalculate the path after obstacle avoidance and ensure that the new path avoids obstacles. The collision-free area is smoothed for the generated new path. B-spline or Bezier curve is used to eliminate sharp angles and unnecessary turns in the path. According to the current position of the AGV and the position information of the collision point, the range of the local path is defined. Usually a radius (such as 2 meters) centered on the collision point is selected, and the path is fine-tuned within this range. In order to improve efficiency, the accuracy standard of the local path can be set, such as the path point spacing not exceeding 0.5 meters. A suitable path fine-tuning algorithm is selected, such as the local A algorithm or RRT. These algorithms can quickly find feasible paths in the local area. Algorithm parameters are set, such as the maximum number of steps for local path search (such as 100 steps) and the accuracy of path search (such as 0. 1 meter), execute the path fine-tuning algorithm within the defined local range, generate a new local path, avoid collision points and ensure that the AGV can pass smoothly, combine the new local path with the global path to form an updated dynamic obstacle avoidance path, verify the fine-tuned path to ensure that it meets the kinematic constraints of the AGV, such as the maximum turning radius and the minimum turning radius, integrate the results of the adaptive obstacle avoidance decision-making process and the local path fine-tuning, and build a dynamic obstacle avoidance path optimization model. This model should be able to respond to environmental changes in real time and automatically adjust the path according to real-time data. Set the model's learning mechanism to ensure that the model accumulates experience in each obstacle avoidance process, thereby improving future decision-making efficiency and accuracy.

[0022] In this embodiment, refer to Figure 2, is a flowchart of detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Step S11: Acquire a full-range image of the AGV activity area based on multiple high-definition cameras; Step S12: performing panoramic image stitching processing on the omnidirectional images of the AGV activity area to construct a panoramic image of the activity area; Step S13: performing image brightness enhancement processing on the panoramic image of the active area to generate a brightness enhanced panoramic image; Step S14: static obstacle recognition is performed on the brightness enhanced panoramic image to construct each static obstacle model.

[0023] In this embodiment, multiple high-definition cameras are selected. It is recommended to use cameras with a resolution of at least 1080p to ensure that the clarity of the acquired images meets the subsequent processing requirements. Consider using wide-angle cameras to increase the field of view and reduce blind spots. Determine the model and technical specifications of the camera to ensure that it has good low-light performance and dynamic range to adapt to different lighting conditions. Arrange the cameras reasonably in the AGV activity area to ensure that the perspectives of each camera overlap to form a seamless coverage area. The running path and activity range of the AGV should be taken into consideration during the layout to avoid blind spots. If the activity area is a 10m × 10m space, install a camera in each corner and in the middle to ensure that the image of the activity area is captured in all directions. Start the camera and perform real-time image acquisition through image acquisition software. Use OpenCV or other image processing libraries to acquire and process image data. Set the acquisition frequency, such as acquiring 30 frames per second, to ensure that the movement of the AGV in dynamic scenes can be clearly recorded. Record, store the collected image data in a local server or cloud storage for subsequent processing, ensure that the storage format is a lossless compression format (such as PNG or TIFF) to preserve image quality, record the timestamp and camera ID of each frame of the image to facilitate subsequent image stitching and processing, and correct the images collected by each camera before stitching to eliminate lens distortion. Use camera calibration technology to obtain internal and external parameters, and perform image correction through OpenCV's calibrateCamera function to ensure the geometric consistency of the image to improve the stitching effect. Use feature point detection algorithms (such as SIFT, SURF, or ORB) to extract key points and descriptors in each image. These feature points will be used for subsequent image matching and stitching. Select enough feature points in each image to ensure the continuity and consistency of the stitched panoramic image. Perform image matching through feature point descriptors, calculate the matching point pairs between each pair of adjacent images, and use FLANN (Fast Library for Approximate Nearest Neighbors) or BFMatcher for matching, filter matching points, use the RANSAC algorithm to remove false matches, ensure the accuracy of matching points, use matching points to calculate the homography matrix (Homography), transform and stitch adjacent images, use OpenCV's warpPerspective function to transform images, use multi-level fusion algorithms (such as weighted average) to smooth the stitching area, reduce the visual impact of gaps and seams, and generate seamless panoramic images. Use image processing algorithms to analyze the brightness distribution of panoramic images, calculate the image histogram to determine the uniformity and distribution of brightness, identify dark and bright areas in the image, provide a basis for subsequent brightness enhancement processing, and select a suitable brightness enhancement algorithm, such as histogram equalization or adaptive histogram equalization (CLAHE).CLAHE can enhance the overall brightness of an image while maintaining local contrast. Algorithm parameters, such as contrast limit and grid size, are set to control the enhancement effect. The selected brightness enhancement algorithm is implemented through OpenCV. The panoramic image is processed using the functions cv2.createCLAHE() and apply() methods for adaptive histogram equalization. During the processing, the color balance of the image is ensured to avoid color distortion caused by brightness enhancement. The quality of the enhanced panoramic image is evaluated to ensure that the brightness and contrast meet expectations. The image is evaluated through visual inspection and calculation of indicators such as the mean square error (MSE). The panoramic image with brightness enhancement is saved in a high-quality format for subsequent obstacle recognition. The panoramic image with brightness enhancement is preprocessed, including color space conversion (such as from BGR to HSV) and denoising to improve. To improve the accuracy of subsequent recognition, use Gaussian blur or median filtering to remove noise from the image and improve image quality. Use edge detection algorithms (such as Canny edge detection) to extract edge features in the image to identify potential static obstacles. Use contour detection (such as OpenCV's findContours function) to identify objects in the image and extract contour information. Use machine learning or deep learning algorithms to classify the extracted features. Apply convolutional neural networks (CNNs) to train the model to identify different types of static obstacles (such as shelves, columns, etc.). Generate training data sets and train the model to ensure that it can accurately identify various static obstacles. Build a model for each identified static obstacle, extract its location, shape, size and other feature information, and store it in the database for subsequent use. Ensure that each model has a unique identifier so that it can be managed and called in the system.

[0024] In this embodiment, the specific steps of step S14 are: Perform image visual recognition on the brightness-enhanced panoramic image and mark all object nodes in the area; Perform deep semantic segmentation on all object nodes in the area and extract static obstacles; Detect edge contours of static obstacles; Perform boundary frame division processing according to the edge contour line to obtain multiple static obstacle image frames; Calculate the geometric shape of each obstacle for multiple static obstacle image frames one by one to generate the geometric shape features of each obstacle; Performing ambient light reflection feature recognition on multiple static obstacle image frames to generate obstacle material characteristics; Obstacle morphology fitting is performed based on the geometric features and material properties of each obstacle to construct each static obstacle model.

[0025] In this embodiment, a deep learning model suitable for image visual recognition is selected, such as YOLO (You Only LookOnce) or Faster R-CNN. These models have good real-time detection capabilities and can handle complex scenes. Prepare training sets and validation sets to ensure that the model can recognize various objects in the activity area. These data sets should include image samples under different lighting, angles and distances. Use the labeled image data set to train the selected model to improve its recognition accuracy. Hyperparameters such as learning rate and batch size should be adjusted during training, and it is recommended to use the Adam optimizer. During the training process, monitor the loss function and accuracy of the model and make appropriate adjustments to ensure that the model achieves a high recognition accuracy (such as ≥90%). Input the panoramic image with enhanced brightness into the trained deep learning model. Forward propagation is performed through the model to obtain the detection results of all objects in the area, including the bounding box and category label of each object. Set a confidence threshold (such as 0.5) and only retain the detection results above the threshold to reduce false detection. Draw the bounding box and category label on the original image to generate a labeled image. The location information of each object node (including coordinates and category) will be stored in the database for subsequent processing. Select a deep semantic segmentation algorithm, such as U-Net or DeepLabV3. These algorithms can classify images at the pixel level and are suitable for extracting static obstacles. Prepare training data to ensure that the dataset contains diverse scenes and different types of static obstacles. Train the deep semantic segmentation model and adjust hyperparameters to optimize model performance. Monitor the performance of the model on the validation set to ensure high accuracy and low intersection over union (IoU). Input the labeled panoramic image into the deep semantic segmentation model for pixel-level classification. The model will output a segmented image of the same size as the original image, where each pixel is labeled with the corresponding category. Extract the area belonging to the static obstacle from the segmented image and generate a binary mask image (mask). The mask image will be used for subsequent contour detection and bounding box segmentation. Select an edge detection algorithm, such as Canny edge detection or Sobel operator. These algorithms can effectively identify edge information in images. Set the parameters of edge detection, such as the high and low thresholds in the Canny algorithm. It is recommended to set them to 100 and 200 initially to adapt to most environments. Input the binary mask image obtained by segmentation into the edge detection algorithm to obtain the edge contour of the static obstacle. Optimize the edge detection effect by adjusting the parameters to ensure clear edges. Use the contour extraction function of image processing libraries such as OpenCV to obtain the coordinate information of the edge contour. Store the extracted contour information in a data structure for subsequent processing. Use the boundingRect() function of OpenCV to divide the bounding box for the extracted contour. This function can calculate the minimum enclosing rectangle based on the coordinates of the contour.For each detected contour, call the bounding box partitioning function to calculate the minimum enclosing rectangle of each obstacle and record the coordinates and size of the bounding box. Store the bounding box information in the database for subsequent geometric calculation and analysis. Define the geometric features of each obstacle, such as area, perimeter, aspect ratio, tilt angle, etc. These features are crucial for subsequent obstacle identification and classification. For each bounding box, use the OpenCV library to calculate its geometric features. The area calculation is implemented by the cv2.contourArea() function, and the perimeter is calculated by the cv2.arcLength() function. Store the geometric features of each obstacle in a data structure to form an information table containing all obstacle features. Select a material recognition algorithm based on spectral characteristics, such as reflectivity analysis. By analyzing the reflection characteristics of light in the image, the material properties of the obstacle are inferred. Perform light reflection analysis on each obstacle image frame to obtain reflectivity features. Use color information (RGB values) of different regions in the image to estimate material properties. Combine the material properties of each obstacle with the geometric features to form a comprehensive feature data set for subsequent model fitting. Select an appropriate morphological fitting algorithm, such as the least squares method or RANSAC (random sampling consensus algorithm), to generate a model based on the geometric features of the obstacles. Fit the model based on the geometric morphological features and material properties of each obstacle. Use the selected algorithm to parametrically model the shape of the obstacle. After fitting, use the test set to verify the accuracy and robustness of the model. Evaluate the performance of the model through cross-validation to ensure its applicability in different scenarios.

[0026] In this embodiment, refer to Figure 3 , is a flowchart of detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Step S21: performing architectural structural element analysis on the brightness enhanced panoramic image to obtain a plurality of architectural structural elements; Step S22: performing scene three-dimensional structure perception according to multiple building structural elements to generate a three-dimensional scene building map; Step S23: locating the three-dimensional position of each static obstacle one by one to obtain the scene position coordinates of each static obstacle; Step S24: Map each static obstacle model to the three-dimensional scene building map according to the scene position coordinates of each static obstacle, so as to obtain a global perception scene model.

[0027] In this embodiment, suitable image processing tools and algorithms, such as OpenCV and deep learning models (such as MaskR-CNN) are selected to analyze building structural elements. Mask R-CNN can achieve instance segmentation, which helps to extract different parts of the building structure. Collect and prepare training data to ensure that the data set contains image samples of various building structural elements, such as walls, windows, doors, etc. Train the deep learning model and use a labeled building structure data set to improve the recognition accuracy of the model. Set training parameters, such as learning rate (0.001) and batch size (16), to ensure the convergence and generalization ability of the model. Test the model on the validation set to ensure that the accuracy of the model reaches more than 85% so that it can be used for actual building element analysis. Input the brightness-enhanced panoramic image into the trained deep learning model to perform instance segmentation of building structural elements. The model will output the category and mask information of each building element. Set a confidence threshold (such as 0.6) to retain only structural elements above the threshold to reduce false detection. The information of the extracted building structure elements (including category, location, size, etc.) is stored in a data structure to form a list of multiple building structure elements for subsequent 3D structure perception. Select appropriate 3D modeling tools and algorithms, such as point cloud generation and 3D reconstruction algorithms (such as Structure from Motion, SfM). SfM can generate 3D point clouds from images from multiple perspectives. Determine the parameters to be used, such as the density of the point cloud (10,000 points / square meter) and the reconstruction accuracy (error less than 5 cm). Input the extracted building structure elements and their location information into the 3D reconstruction algorithm to generate the corresponding 3D point cloud. Use camera intrinsic and extrinsic parameters to ensure the accuracy of the point cloud. Construct a 3D building map based on the generated point cloud. The point cloud can be converted into a visual 3D model using a triangular mesh generation algorithm (such as Poisson Surface Reconstruction). Perform post-processing to optimize the surface smoothness and details of the model to ensure that the generated 3D building map conforms to the actual scene. Use methods based on feature point matching and deep learning to locate the 3D position of static obstacles. Feature point matching algorithms (such as ORB or SIFT) can effectively identify the position of obstacles in 3D space. Extract feature points from each static obstacle image frame and match them with the feature points in the 3D building map. Set a matching threshold (such as 0.75) to ensure the accuracy of the match. Calculate the position coordinates of each static obstacle in 3D space using the matched feature points using triangulation. Store the calculated coordinates in a data structure for subsequent processing. Select a suitable mapping algorithm to embed the static obstacle model into the 3D scene building map. The coordinate transformation method can be used to locate and scale the model. According to the 3D position coordinates of each static obstacle, scale its model proportionally and locate it to the corresponding position of the 3D building map.Ensure that the orientation and posture of the model are consistent with the actual scene. Embed all static obstacle models into the 3D scene building map to generate a global perception scene model. Use 3D visualization tools (such as Blender or Unity) for final model display and verification.

[0028] In this embodiment, refer to Figure 4 , is a flowchart of detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Step S31: obtaining a cargo transportation planning log; identifying a plurality of cargoes to be transported within a period according to the cargo transportation planning log; Step S32: Calculating the current positions of multiple goods to be transported within the period, and extracting the current position information of each of the goods to be transported; Step S33: Perform statistics on the final transportation location of the goods in the goods transportation planning log, and mark the transportation destination of each goods to be transported; Step S34: Performing dynamic cargo transportation sequence planning according to the transportation destination of each cargo to be transported, thereby generating an optimal cargo transportation sequence; Step S35: Based on the optimal transportation sequence of the goods and the current location information of each cargo to be transported, a globally available path is planned for the global perception scene model to construct a globally available planned path.

[0029] In this embodiment, the data source of the cargo transportation planning log is determined, for example, data is extracted through the enterprise's logistics management system (WMS) or transportation management system (TMS). These systems usually record detailed information of all transportation activities, including cargo type, quantity, transportation time and transportation route. Use API or database query language (such as SQL) to extract cargo transportation logs from the system. Set extraction conditions, such as time range (past week) and transportation status (such as waiting for transportation, transportation, etc.). Record the extracted fields, such as cargo ID, name, estimated transportation time, starting location and target location, to form a structured data set. Analyze the extracted transportation logs to identify multiple cargoes to be transported within a set period. Use data analysis tools (such as Python's Pandas library) for data processing. Set filtering conditions, such as records with cargo status as "waiting for transportation" to create a list of cargoes to be transported. The identified cargoes to be transported are recorded in a new data structure, including cargo ID, name, current status and transportation information, etc., for use in subsequent steps. This data structure is a table or dictionary for easy subsequent search and processing. Determine the source of the cargo location data, which is usually obtained through RFID tags, GPS positioning, or barcode scanning. Ensure that each cargo to be transported has corresponding location information. Use appropriate hardware (such as RFID readers or GPS receivers) and software systems (such as location tracking management systems) to collect real-time data and record the current location of each cargo. Extract the current location data of each cargo to be transported from the positioning system. Obtain real-time location information through API or database query, including coordinates (longitude, latitude) or coordinates within the warehouse. Use data processing tools (such as Pandas) to merge the location data with the list of cargo to be transported to form a new dataset containing cargo ID and current location information. Store the current location information (such as coordinates, shelf location, etc.) of each cargo to be transported in a data structure to provide basic data for subsequent analysis and planning. Extract the target transportation location of each cargo to be transported from the transportation planning log, usually the destination warehouse or customer address. Ensure that the data contains a clear final transportation location. Use data analysis tools (such as Python's Pandas) to perform statistical processing on the transportation log to extract and analyze the transportation destination of each cargo. Group the transport logs and extract the corresponding transport destination information for each cargo ID to be transported. Use the groupby() function or a similar method to generate a statistical table of the final transport location for each cargo. Store the ID of each cargo to be transported and the corresponding transport destination information in a data structure to form a list of final transport locations for use in subsequent steps. Ensure that the data format is clear, including the cargo ID, transport destination, and related information. Select an appropriate transport sequence planning algorithm, such as a greedy algorithm, a genetic algorithm, or a heuristic search method. Determine the optimal algorithm based on business needs and cargo characteristics.Determine the parameters that affect the transportation sequence, such as transportation time, distance, priority, and cargo type. Set weights so that multiple factors are considered during the planning process. Dynamically sort the transportation destinations of the cargo to be transported using the selected planning algorithm. Use the Dijkstra algorithm to calculate the shortest path from the current location to each target location and generate a transportation sequence based on the path length. Store the generated optimal transportation sequence of the cargo in a data structure to form an ordered list of transportation plans, including the cargo ID and its corresponding transportation sequence and path information. Select an appropriate path planning algorithm, such as the A* algorithm or the RRT (rapid random tree) algorithm. These algorithms are effective in finding feasible paths in complex environments. Ensure that the global perception scene model (such as map data) is ready, including all static obstacle information and pass paths. Generate the scene model using a 3D modeling tool (such as Blender or Unity). Based on the optimal transportation sequence of the cargo and the current location of each cargo to be transported, use the selected path planning algorithm to calculate a feasible path from the current location to the target location.

[0030] In this embodiment, step S4 includes the following steps: Step S41: obtaining real-time transport state parameters of the AGV; evaluating the load capacity of the AGV based on the real-time transport state parameters of the AGV to obtain a load capacity value; Step S42: Calculate the transport speed of the AGV's operating status parameters to generate the AGV transport speed; Step S43: quantifying the cargo transportation efficiency based on the load capacity value and the AGV transportation speed, and generating the transportation efficiency of each cargo; Step S44: Calculate the cargo transportation distance for the globally available planned paths to generate the transportation distance for each cargo; Step S45: Dynamically optimize the globally available planned paths based on the transportation efficiency of each cargo and the transportation distance of each cargo, thereby obtaining a dynamically optimized transportation path.

[0031] In this embodiment, the transport status parameters are collected in real time by sensors on the AGV (such as load sensors, accelerometers, and GPS modules). The sensors should have high precision and reliability to ensure the accuracy of data collection. Set the data collection frequency, such as collecting status data once per second to ensure real-time monitoring. The collected parameters include current load, running speed, driving direction, and battery power. The collected data is transmitted to the central control system or cloud server through a wireless communication module (such as Wi-Fi or Bluetooth). Ensure the stability and real-time nature of data transmission for subsequent analysis. The data is recorded in a database to form a structured real-time transport status data set, including timestamp, load, speed and other information. According to the design specifications and sensor data of the AGV, a load capacity evaluation model is established. The model is a simple linear model, setting the maximum load (such as 500kg) and the current load (such as the value measured by the load sensor). Implement the evaluation formula: Load capacity value = current load / maximum load × 100%, which expresses the load capacity of the AGV in percentage. In the real-time transport status parameters, the location information and timestamp of the AGV are obtained. These data are obtained through GPS modules or odometers to ensure high-precision position tracking. Record the displacement of the AGV within a certain time interval for subsequent speed calculation. Transportation speed = displacement / time interval. Set the time interval to 1 second to calculate the speed of the AGV in real time. If the AGV moves 2 meters in 1 second, the speed is 2 meters per second. Store the calculated AGV transportation speed in the database and associate it with the real-time transportation status parameters and load capacity value to form a complete transportation efficiency evaluation system. Set the calculation formula for transportation efficiency: Transportation efficiency = load capacity value × transportation speed. The load capacity value is expressed as a percentage and the transportation speed is expressed in meters per second. If the load capacity value is 80% (i.e. 400 kg) and the transportation speed is 2 m / s, the transportation efficiency is 320 kg·m / s. For each cargo to be transported, calculate the transportation efficiency using the above formula. Make sure to consider the characteristics and current status of each cargo in the calculation. Store the calculation results in a data structure to form a list containing the transportation efficiency of each cargo. Extract the starting position and target position of each cargo from the globally available planning path. Ensure that the path information is accurate for subsequent distance calculations. The path is displayed and analyzed using visualization tools (such as GIS software). Use the formula: transportation distance = straight-line distance from the starting location to the target location. Use the Haversine formula or Euclidean distance calculation to ensure the calculation result is accurate. If the starting location is (x1, y1) and the target location is (x2, y2), the transportation distance is √((x2 - x1)² + (y2 - y1)²). Store the calculated transportation distance of each cargo in a data structure to form a list containing the transportation distance of each cargo for subsequent optimization.Select a suitable dynamic path optimization algorithm, such as genetic algorithm or particle swarm optimization (PSO). These algorithms can dynamically adjust the path according to the transportation efficiency and distance. Determine the optimization parameters, such as the weight of transportation efficiency (e.g. 0.7) and the weight of transportation distance (e.g. 0.3), so as to balance the influence of the two during the optimization process. Use the selected optimization algorithm to adjust the globally available planning path. Dynamically calculate the optimal path according to the transportation efficiency and transportation distance of each cargo. Optimize the path by calculating the total transportation cost (such as time, energy consumption, etc.).

[0032] In this embodiment, the specific steps of step S5 are: Step S51: Controlling AGV transportation according to the dynamically optimized transportation path, and collecting monitoring videos of the entire transportation process; Step S52: Visually mark the transportation path of the whole transportation process monitoring video, so as to generate a visual transportation path; Step S53: Detect dynamic targets on the visualized transport path and identify dynamic obstacles on the path; Step S54: performing dynamic optical flow tracking on the path dynamic obstacles to generate the moving trajectory of the dynamic obstacles; Step S55: predicting the movement collision probability of the dynamic obstacle movement trajectory based on the dynamic optimized transportation path, thereby generating a path collision prediction probability; Step S56: mining potential collision points based on the path collision prediction probability to generate collision time points and location points.

[0033] In this embodiment, the control system of the AGV is configured according to the dynamically optimized transportation path. Ensure that the navigation system of the AGV (such as laser navigation or visual navigation) can receive and understand the optimized path information. Set the path tracking strategy to ensure that the AGV can accurately travel along the optimized path. Use a PID controller to adjust the speed and direction of the AGV. Install a high-definition camera on the AGV to collect monitoring videos of the entire transportation process in real time. Select a suitable camera model and ensure that its resolution is at least 1080p to obtain a clear video picture. Configure the video acquisition software so that it can record the transportation status of the AGV in real time, including the starting point, end point and path. Start the AGV and start the transportation task according to the set optimized path. During the transportation process, monitor the operating status of the AGV in real time, including speed, load and path deviation. The monitoring video should cover the front, side and rear of the AGV to ensure a comprehensive record of environmental changes during transportation. The collected monitoring video is transmitted to the central control system in real time and stored in a compressed format (such as H.264) to reduce storage space. Record the timestamp of the video and the status information of the AGV for subsequent analysis. Select video processing software (such as OpenCV) to analyze and process the video of the entire transportation process. Make sure the software can process high-definition video. Set the video analysis parameters, including frame rate (such as 30 frames per second) and processing resolution to ensure clear visualization. Through video analysis, extract the actual path of the AGV during transportation. Use image processing algorithms (such as edge detection and contour extraction) to identify the movement trajectory of the AGV. Mark the movement path of the AGV in the video and use different colored lines or arrows to represent the path to generate a visualization effect. Overlay the marked path on the original video to generate a visual transportation path video. Make sure that the video contains time information and transportation status for subsequent analysis and playback. Select a suitable dynamic object detection algorithm, such as YOLO (You Only Look Once) or SSD (SingleShot MultiBox Detector). These deep learning models can detect dynamic objects in the video in real time. Make sure the model has been trained to identify specific types of obstacles (such as pedestrians, vehicles, etc.). Input the visual transportation path video into the selected object detection model to detect dynamic obstacles. Set a detection threshold (such as 0.5) to reduce false detections. For each dynamic obstacle detected, record its category, location (bounding box coordinates), and confidence, and store it in a data structure for subsequent use. Select a suitable optical flow calculation method, such as the Lucas-Kanade optical flow method. This method can effectively estimate the motion of objects in image sequences. Determine the parameters of the optical flow calculation, such as the window size and the number of pyramid layers, to improve the calculation accuracy. For the detected dynamic obstacles, use the optical flow method to calculate their movement trajectory in the video frame. Track the position changes of the obstacles between each frame.Record the moving trajectory of each dynamic obstacle, including timestamp and location coordinates, for subsequent analysis. Design a collision probability prediction model based on the moving trajectory of dynamic obstacles and the dynamic path of AGV. Use a probability model based on time and distance. Set parameters that affect the collision probability, such as obstacle speed, AGV speed, distance threshold, etc. For each dynamic obstacle, calculate the collision probability based on its moving trajectory and the dynamic path of AGV. Use the formula: collision probability = 1 - (safe distance / actual distance). Record the collision prediction probability of each obstacle for subsequent processing and decision-making. Design a potential collision point mining algorithm to identify the time and location points where the AGV collides with the obstacle based on the collision probability and the moving trajectory of the obstacle. Determine the mining parameters, such as time window (e.g. 1 second) and distance threshold. In each time window, analyze the path of the AGV and the trajectory of the dynamic obstacle to identify potential collision points. If the collision probability exceeds the set threshold (e.g. 0.7), record the point. Record the time and location points of potential collisions to form a list of potential collision points for subsequent decision-making and adjustment.

[0034] In this embodiment, the specific steps of step S6 are: Step S61: making an adaptive obstacle avoidance decision according to the collision time point and position point to obtain an adaptive obstacle avoidance strategy; Step S62: fine-tuning the local path of the collision point of the dynamically optimized transportation path based on the adaptive obstacle avoidance strategy to obtain a local fine-tuning path; Step S63: performing path smoothing optimization on the local fine-tuning path to generate a local fine-tuning smoothing path; Step S64: performing iterative reinforcement learning on the local fine-tuning smooth path, thereby constructing a dynamic obstacle avoidance path optimization model.

[0035] In this embodiment, the collision time point and location point data generated in the previous steps are collected and integrated. These data should include detailed information of each potential collision point, such as timestamp, location coordinates, collision probability, etc. A data structure (such as a dictionary or database table) is established to associate this information for use in the subsequent decision-making process. An adaptive obstacle avoidance decision model is designed, and a rule-based decision system or a machine learning model is adopted. The rule-based method is simple and effective, and is suitable for real-time decision-making; while the machine learning method performs better in more complex environments. Set decision criteria, for example, when the collision probability exceeds a set threshold (such as 0.7), the AGV needs to adjust its path immediately. A fuzzy logic system can be used to handle uncertainty. Based on the collision point information, an adaptive obstacle avoidance strategy is generated. If the AGV is about to collide with a dynamic obstacle, a new temporary target point is set according to its current position and target position, and the driving direction and speed of the AGV are adjusted. The adaptive obstacle avoidance strategy is recorded in the system for subsequent path adjustment and decision-making. According to the adaptive obstacle avoidance strategy, the collision points in the dynamically optimized transportation path are analyzed to determine the area that needs local fine-tuning. This usually involves the relative position between the current position of the AGV, the target position, and the collision point. Define the scope of the local area, such as a radius (such as 2 meters) centered on the collision point, within which the path is fine-tuned. Ensure that the local path adjustment does not affect the overall planning of the global path. Select an appropriate path fine-tuning algorithm, such as the A* algorithm or the Dijkstra algorithm, to calculate the new local fine-tuning path. These algorithms can find new feasible paths in the local area. Generate a new path in the local area to avoid the collision point and ensure that the AGV can pass smoothly. Set a new target point, adjust the AGV's driving route and update the path information. Combine the fine-tuned local path with the global path to form a new dynamically optimized transportation path. Select an appropriate path smoothing algorithm, such as Bezier curves, B-spline curves, or Ramer-Douglas-Peucker algorithm. Bezier curves and B-spline curves can generate smooth curved paths, while the Ramer-Douglas-Peucker algorithm can effectively reduce the number of path points. Determine the parameters of the smoothing algorithm, such as the number of control points and tolerance, to maintain the flexibility and feasibility of the path. Smooth the local fine-tuning path to generate a local fine-tuning smoothed path. Use the selected smoothing algorithm to adjust the position of the waypoints to eliminate sharp corners and unnecessary turns in the path. Verify that the generated smooth path meets the kinematic constraints of the AGV, including the maximum and minimum turning radius. Ensure that the smooth path is feasible in actual operation. Select a suitable reinforcement learning algorithm, such as Q-learning or deep Q network (DQN). These algorithms can learn the optimal obstacle avoidance strategy through interaction with the environment. Define the state space, including the current position of the AGV, the target position, the position and speed of dynamic obstacles, etc.Define the action space, including AGV speed adjustment, steering and other operations. Perform AGV path optimization training in a simulated environment. AGV selects an action based on the current state, updates the state after execution, and obtains rewards (such as positive rewards for successfully avoiding obstacles and negative rewards for collisions). Perform multiple rounds of training, gradually adjust the Q value or policy network, and learn the optimal obstacle avoidance path.

[0036] In this embodiment, an intelligent obstacle avoidance and path optimization system for AGV is provided, which is used to execute the intelligent obstacle avoidance and path optimization method for AGV as described above, including: The obstacle recognition module is used to obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct each static obstacle model; A global perception module, used to perceive the three-dimensional structure of the scene in the omnidirectional image, and perform three-dimensional mapping based on each static obstacle model, so as to obtain a global perception scene model; The available path planning module is used to obtain the cargo transportation planning log; perform global available path planning for cargo on the global perception scene model according to the cargo transportation planning log, and construct a global available planning path; The path optimization module is used to perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; The collision prediction module is used to mine potential collision points of dynamic obstacles according to the dynamic optimized transportation path to generate collision time and location points; The local path fine-tuning module is used to make adaptive obstacle avoidance decisions based on the collision time and location points, and to fine-tune the local path at the collision point, thereby building a dynamic obstacle avoidance path optimization model.

[0037] Through omnidirectional image acquisition, the present invention enables AGV to obtain environmental data from different angles in real time, avoiding the blind spots of traditional perception systems, thus ensuring comprehensive perception of the entire activity area. The recognition of static obstacles is based on computer vision and deep learning algorithms (such as convolutional neural networks (CNNs)), which can efficiently and accurately distinguish various static obstacles from images, such as walls, equipment, shelves, etc. The recognition accuracy improves the quality of path planning and obstacle avoidance decisions, and generates a detailed geometric model for each static obstacle, including information such as the shape, size, and material of the obstacle, providing important data support for subsequent three-dimensional mapping and path planning. Through the three-dimensional structure perception of the scene, AGV can more accurately understand the spatial layout of the environment, including the height of the obstacle. , distribution and distance information. Through stereoscopic perception, AGV can avoid not only planar obstacles, but also uneven objects. Combined with the static obstacle model, AGV can build a detailed environmental map in three-dimensional space. This process not only includes the location of obstacles, but also can represent the spatial structure of the site, so that path planning can make efficient decisions in complex environments. The global perception scene model provides richer and more accurate environmental information for path planning, helping AGV to better identify road capacity, optimal paths and potential obstacles, and improve the accuracy and reliability of path planning. By obtaining the cargo transportation planning log, the system can customize path planning for AGV according to specific transportation tasks and goals, the type, weight and destination of the cargo. Factors such as the environment and the speed of the vehicle will affect the selection and optimization of the path. Combined with the global perception scenario model, the path planning module can take into account all obstacles in the environment, road conditions, traffic flow and other factors to develop the most suitable global path. This not only reduces the redundant part of the path, but also ensures the efficiency and safety of the transportation process. During the operation of the AGV, the environment changes (the emergence of new obstacles or temporary blockages on the path). The path optimization module can dynamically adjust the global planning path to cope with these changes and ensure the efficient completion of the transportation task. Through dynamic optimization, the system can select the optimal path based on real-time traffic conditions, cargo demand, energy consumption and other factors, reduce unnecessary detours and time waste, and reduce transportation costs. By analyzing dynamic The position, speed and movement trajectory of obstacles, the collision prediction module can evaluate the potential collision points with the AGV path in real time, which provides timely warning for the AGV and helps it make avoidance decisions in advance. Based on collision prediction, the AGV avoids high-risk areas in advance to avoid accidental collisions. Avoiding collisions not only ensures the safety of equipment and goods, but also reduces the maintenance cost of the system. Through accurate collision point prediction, the AGV can maintain a high stability during real-time transportation and avoid path interruptions or unexpected downtime. Based on the real-time predicted collision point information, the local path fine-tuning module quickly makes adaptive obstacle avoidance decisions according to the current position and motion state of the AGV. The AGV can flexibly choose strategies such as stopping, detouring or adjusting speed to avoid collisions.Local path fine-tuning can minimize unnecessary path adjustments and only make fine-tuning when necessary, thereby avoiding large-scale path planning modifications and ensuring the continuity and stability of transportation tasks.

[0038] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0039] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. An intelligent obstacle avoidance and path optimization method for AGV, characterized in that: The following steps are involved: Step S1: Obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct a model of each static obstacle; Step S2: performing three-dimensional structure perception of the scene on the omnidirectional image, and performing three-dimensional mapping based on each static obstacle model, thereby obtaining a global perception scene model; Step S3: Obtain cargo transportation planning logs; perform cargo global available path planning on the global perception scene model according to the cargo transportation planning logs, and construct a globally available planning path; Step S4: Perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; Step S5: mining potential collision points of dynamic obstacles according to the dynamically optimized transport path to generate collision time points and location points; Step S6: Adaptive obstacle avoidance decision is made according to the collision time point and location point, and local path fine-tuning of the collision point is performed to construct a dynamic obstacle avoidance path optimization model.

2. The intelligent obstacle avoidance and path optimization method for AGV according to claim 1 is characterized in that: The specific steps of step S1 are: Step S11: Acquire a full-range image of the AGV activity area based on multiple high-definition cameras; Step S12: performing panoramic image stitching processing on the omnidirectional images of the AGV activity area to construct a panoramic image of the activity area; Step S13: performing image brightness enhancement processing on the panoramic image of the active area to generate a brightness enhanced panoramic image; Step S14: static obstacle recognition is performed on the brightness enhanced panoramic image to construct each static obstacle model.

3. The intelligent obstacle avoidance and path optimization method for AGV according to claim 2 is characterized in that: The specific steps of step S14 are: Perform image visual recognition on the brightness-enhanced panoramic image and mark all object nodes in the area; Perform deep semantic segmentation on all object nodes in the area and extract static obstacles; Detect edge contours of static obstacles; Perform boundary frame division processing according to the edge contour line to obtain multiple static obstacle image frames; Calculate the geometric shape of each obstacle for multiple static obstacle image frames one by one to generate the geometric shape features of each obstacle; Performing ambient light reflection feature recognition on multiple static obstacle image frames to generate obstacle material characteristics; Obstacle morphology fitting is performed based on the geometric features and material properties of each obstacle to construct each static obstacle model.

4. The intelligent obstacle avoidance and path optimization method for AGV according to claim 1, characterized in that: The specific steps of step S2 are: Step S21: performing architectural structural element analysis on the brightness enhanced panoramic image to obtain a plurality of architectural structural elements; Step S22: performing scene three-dimensional structure perception according to multiple building structural elements to generate a three-dimensional scene building map; Step S23: locating the three-dimensional position of each static obstacle one by one to obtain the scene position coordinates of each static obstacle; Step S24: Map each static obstacle model to the three-dimensional scene building map according to the scene position coordinates of each static obstacle, so as to obtain a global perception scene model.

5. The intelligent obstacle avoidance and path optimization method for AGV according to claim 1, characterized in that: The specific steps of step S3 are: Step S31: obtaining a cargo transportation planning log; identifying a plurality of cargoes to be transported within a period according to the cargo transportation planning log; Step S32: Calculating the current positions of multiple goods to be transported within the period, and extracting the current position information of each of the goods to be transported; Step S33: Perform statistics on the final transportation location of the goods in the goods transportation planning log, and mark the transportation destination of each goods to be transported; Step S34: Performing dynamic cargo transportation sequence planning according to the transportation destination of each cargo to be transported, thereby generating an optimal cargo transportation sequence; Step S35: Based on the optimal transportation sequence of the goods and the current location information of each cargo to be transported, a globally available path is planned for the global perception scene model to construct a globally available planned path.

6. The intelligent obstacle avoidance and path optimization method for AGV according to claim 1, characterized in that: The specific steps of step S4 are: Step S41: obtaining real-time transport state parameters of the AGV; evaluating the load capacity of the AGV based on the real-time transport state parameters of the AGV to obtain a load capacity value; Step S42: Calculate the transport speed of the AGV's operating status parameters to generate the AGV transport speed; Step S43: quantifying the cargo transportation efficiency based on the load capacity value and the AGV transportation speed, and generating the transportation efficiency of each cargo; Step S44: Calculate the cargo transportation distance for the globally available planned paths to generate the transportation distance for each cargo; Step S45: Dynamically optimize the globally available planned paths based on the transportation efficiency of each cargo and the transportation distance of each cargo, thereby obtaining a dynamically optimized transportation path.

7. The intelligent obstacle avoidance and path optimization method for AGV according to claim 1, characterized in that: The specific steps of step S5 are: Step S51: Controlling AGV transportation according to the dynamically optimized transportation path, and collecting monitoring videos of the entire transportation process; Step S52: Visually mark the transportation path of the whole transportation process monitoring video, so as to generate a visual transportation path; Step S53: Detect dynamic targets on the visualized transport path and identify dynamic obstacles on the path; Step S54: performing dynamic optical flow tracking on the path dynamic obstacles to generate the moving trajectory of the dynamic obstacles; Step S55: predicting the movement collision probability of the dynamic obstacle movement trajectory based on the dynamic optimized transportation path, thereby generating a path collision prediction probability; Step S56: mining potential collision points based on the path collision prediction probability to generate collision time points and location points.

8. The intelligent obstacle avoidance and path optimization method for AGV according to claim 1, characterized in that: The specific steps of step S6 are: Step S61: making an adaptive obstacle avoidance decision according to the collision time point and position point to obtain an adaptive obstacle avoidance strategy; Step S62: fine-tuning the local path of the collision point of the dynamically optimized transportation path based on the adaptive obstacle avoidance strategy to obtain a local fine-tuning path; Step S63: performing path smoothing optimization on the local fine-tuning path to generate a local fine-tuning smoothing path; Step S64: performing iterative reinforcement learning on the local fine-tuning smooth path, thereby constructing a dynamic obstacle avoidance path optimization model.

9. An intelligent obstacle avoidance and path optimization system for AGV, characterized in that: The method for executing the intelligent obstacle avoidance and path optimization of the AGV according to claim 1 comprises: The obstacle recognition module is used to obtain an omnidirectional image of the AGV activity area; perform static obstacle recognition on the omnidirectional image to construct each static obstacle model; A global perception module, used to perceive the three-dimensional structure of the scene in the omnidirectional image, and perform three-dimensional mapping based on each static obstacle model, so as to obtain a global perception scene model; The available path planning module is used to obtain the cargo transportation planning log; perform global available path planning for cargo on the global perception scene model according to the cargo transportation planning log, and construct a global available planning path; A path optimization module is used to perform dynamic path optimization based on the globally available planned paths, thereby obtaining a dynamically optimized transportation path; The collision prediction module is used to mine potential collision points of dynamic obstacles according to the dynamic optimized transportation path to generate collision time and location points; The local path fine-tuning module is used to make adaptive obstacle avoidance decisions based on the collision time and location points, and to fine-tune the local path at the collision point, thereby building a dynamic obstacle avoidance path optimization model.

Images