Complex dynamic industrial visual positioning method and system based on multi-agent collaboration

By employing a multi-agent collaborative visual positioning method, the problems of insufficient field of view coverage and decreased positioning accuracy in complex and dynamic industrial scenarios are solved, achieving high-precision and stable visual positioning and adapting to the challenges of various environmental interferences and insufficient node collaborative scheduling.

CN122090040APending Publication Date: 2026-05-26NANJING ALLCAM INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610525206.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-21
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing industrial visual positioning technologies suffer from problems such as insufficient field of view coverage, decreased positioning accuracy, high system latency, poor environmental adaptability, unstable feature matching, and insufficient node collaborative scheduling in complex and dynamic industrial scenarios, and cannot meet the high-precision positioning requirements of high-speed dynamic targets.

Method used

A complex dynamic industrial visual positioning method based on multi-agent collaboration is adopted. By distributing multiple visual agent nodes with independent edge acquisition and computing capabilities, and combining deep convolutional neural networks and multi-view geometric constraint models, image preprocessing, feature extraction and matching are performed, nonlinear compensation optimization is executed, and task allocation and collaborative scheduling are carried out through deep reinforcement learning to achieve high-precision and stable positioning.

Benefits of technology

It improves positioning accuracy and stability in complex and dynamic industrial scenarios, reduces system latency and data transmission pressure, achieves adaptability to various environmental interferences, and ensures stable operation of the system when a single node is abnormal.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122090040A_ABST
    Figure CN122090040A_ABST
Patent Text Reader

Abstract

The invention discloses a complex dynamic industrial visual positioning method and system based on multi-agent collaboration, and relates to the technical field of machine vision, distributed visual agent nodes are deployed in an industrial dynamic operation scene, and node microsecond-level time synchronization, space joint calibration and communication networking are completed; the nodes synchronously collect industrial images according to a preset time sequence, and a standardized data set is generated through preprocessing; dynamic target feature extraction is completed based on an improved deep convolutional neural network, and cross-node feature matching is completed in combination with epipolar constraints; a multi-view geometric constraint model is constructed to solve the three-dimensional pose of the target, and a positioning result is generated through dynamic interference compensation optimization; and outputting a positioning result, completing closed-loop feedback, and dynamically adjusting node parameters to realize multi-agent collaborative global optimization. According to the invention, the precision and the anti-interference capability of visual positioning in a complex dynamic industrial scene are improved, the multi-node load balancing is optimized, and the visual positioning operation requirements of various industrial dynamic scenes can be met.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and in particular to a complex dynamic industrial vision positioning method and system based on multi-agent collaboration. Background Technology

[0002] With the continuous development of industrial automation technology, machine vision positioning has become a core supporting technology in intelligent manufacturing, industrial robot operations, and automated production line processing. In dynamic industrial operation scenarios, target workpieces are often in continuous motion, and the scene contains various complex interference factors such as light fluctuations, dust interference, welding slag splashes, and partial occlusion. This places higher demands on the positioning accuracy, real-time response, environmental anti-interference ability, and multi-node collaborative operation capability of vision positioning systems. Traditional single-camera vision positioning solutions are limited by the imaging field of view and can only complete target positioning within a small scene. In large-scale industrial operations, they suffer from insufficient field of view coverage, are prone to target loss during the tracking of dynamically moving targets, and experience a significant decrease in positioning accuracy when the target is partially occluded. They cannot adapt to the operational needs of complex and dynamic industrial scenarios. While centralized visual positioning solutions based on multiple cameras can achieve large-scale field-of-view coverage through multi-camera deployment, the front-end cameras in the system architecture are only responsible for image acquisition. All image data must be transmitted to the central end for unified processing, resulting in problems such as high data transmission bandwidth consumption, concentrated computing pressure at the central end, and high system processing latency. Each acquisition node does not have independent edge computing capabilities and cannot adaptively adjust acquisition parameters according to real-time changes in the scene, resulting in insufficient environmental adaptability in dynamically changing industrial scenarios.

[0003] Current industrial vision positioning technologies often employ a single type of feature descriptor for feature extraction and matching, failing to adequately consider the impact of various dynamic disturbances in industrial scenarios on the stability of feature descriptors. Under conditions such as sudden changes in illumination, motion blur, and partial occlusion, the discriminative power of feature descriptors significantly decreases, leading to an increased mismatch rate. This fails to provide highly reliable feature matching pairs for subsequent pose calculation, directly affecting the final positioning accuracy. In the pose calculation stage, existing technologies primarily optimize 3D pose calculation for static targets. For continuously moving dynamic targets in industrial scenarios, they do not fully incorporate the continuous motion characteristics of the target into the constraint conditions, easily leading to accumulated errors during pose calculation and insufficient positioning stability in high-speed dynamic scenes. Meanwhile, existing multi-node visual positioning systems lack intelligent collaborative scheduling mechanisms. The task allocation and operating parameters of each node are mostly fixed settings, and cannot be dynamically adjusted according to the real-time computing load, image acquisition quality, and positioning accuracy data of each node. The overall load balance of the system is poor. When some nodes experience hardware failure, communication interruption, or abnormal acquisition, the positioning performance of the entire system will be directly affected. Moreover, most existing systems are in open-loop control mode and lack effective closed-loop feedback optimization mechanisms. They cannot optimize the relevant parameters of the front-end positioning link based on the pose deviation data after the industrial operation is executed. The positioning accuracy of the system is difficult to continuously optimize and cannot fully adapt to the normalized operation requirements in complex and dynamic industrial scenarios. Summary of the Invention

[0004] The present invention proposes a complex dynamic industrial visual positioning method and system based on multi-agent collaboration to solve the problems mentioned in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a complex dynamic industrial visual positioning method based on multi-agent collaboration, comprising the following steps: The multi-agent node deployment and initialization steps involve the distributed deployment of multiple visual agent nodes with independent edge acquisition and computing capabilities within a dynamic industrial operation scenario. The multi-source industrial image synchronous acquisition and preprocessing steps involve each visual intelligent agent node synchronously acquiring industrial images in the dynamic operation scene according to the preset hardware triggering sequence, and performing denoising, enhancement, and distortion correction preprocessing on the original images in sequence. The dynamic target feature extraction and matching steps are based on an improved deep convolutional neural network to perform target detection and feature point extraction on standardized industrial images. The feature matching and target association across intelligent agent nodes are completed by combining prior information on the dynamic target motion trajectory and epipolar constraints. The multi-agent collaborative pose calculation step constructs a multi-view geometric constraint model based on the matched feature point pairs and the spatial calibration parameters of each agent node. The dynamic interference compensation and positioning result optimization steps involve performing nonlinear compensation optimization on the calculated 3D pose data based on dynamic interference parameters such as lighting fluctuations, workpiece motion blur, and local occlusion within the scene. The positioning result output and closed-loop feedback steps output the initial positioning result to the industrial operation execution end, and at the same time collect the pose deviation data after the operation is executed; The multi-agent collaborative scheduling and global optimization steps dynamically adjust the triggering sequence, computational weight, and task allocation of each node based on the computational load, image acquisition quality, and positioning accuracy data of each agent node. The scheduling parameters are then fed back to the multi-source industrial image synchronous acquisition and preprocessing steps, as well as the dynamic target feature extraction and matching steps.

[0006] Furthermore, it also includes a dynamic target feature robustness enhancement and accurate matching unit. Addressing issues such as motion blur, abrupt changes in illumination, and partial occlusion in complex dynamic industrial scenarios, it performs multi-dimensional feature fusion and similarity calculation on the extracted feature points. The calculation process is completed using the following formula: ; For image With images The comprehensive matching similarity of corresponding feature point pairs. These are the gradient feature weight coefficients. , Images , The Middle The gradient direction histogram eigenvectors of each feature point These are the texture feature weight coefficients. , Images , The Middle The gray-level co-occurrence matrix and texture feature vector of each feature point These are the weighting coefficients for local structural features. , Images , The Middle Local binary pattern feature vectors of feature points The feature dimension is the result of uniform normalization for all feature vectors. The spatial context constraint attenuation coefficient, , These are the spatial coordinate vectors of the corresponding feature points in their respective images after epipolar constraint verification. Let Euclidean distance be the spatial coordinates of two feature points.

[0007] Furthermore, it also includes a high-precision time synchronization and spatial joint calibration unit for multiple agent nodes. Through a dual synchronization mechanism combining hardware trigger signals and high-precision network time protocols, and based on a 3D calibration board and hand-eye calibration algorithms, it completes the individual calibration of camera intrinsic parameters and distortion parameters of each agent node, as well as the joint calibration of the camera coordinate system of each node and the global coordinate system of the industrial scene, to build a unified global spatial coordinate system and eliminate spatial coordinate deviations between multiple agent nodes.

[0008] Furthermore, in the multi-source industrial image synchronous acquisition and preprocessing steps, an adaptive bilateral filtering algorithm is used to denoise the original image to address image noise caused by welding slag splashes, dust interference, and lens vibration in industrial scenes. While filtering out impulse noise and Gaussian noise, the edge features of the target workpiece are preserved. To address the problem of insufficient image contrast caused by illumination fluctuations, a contrast-limited adaptive histogram equalization algorithm is used to enhance the image. To address image distortion caused by camera lens distortion and installation angle, pixel-level distortion correction is performed based on the distortion coefficients involved in the camera. At the same time, a quality score is performed on the preprocessed image.

[0009] Furthermore, in the dynamic target feature extraction and matching step, an improved YOLO deep convolutional neural network is used to complete the real-time detection of dynamic targets, identify and locate the target workpiece region in the image, remove invalid feature points in the background region, and then perform FAST feature point detection and BRIEF descriptor generation on the target workpiece region through an improved ORB algorithm. At the same time, the Kalman filter algorithm is combined to predict the motion trajectory of the dynamic target, generate the motion prior region of the target workpiece between adjacent frames and cross-agent nodes, and remove mismatched pairs in the matching results based on the random sampling consensus algorithm.

[0010] Furthermore, in the multi-agent collaborative pose calculation and global optimization process, a multi-view nonlinear joint optimization objective function is constructed. Global bundle adjustment optimization is performed on the pose observations of multiple agent nodes and the 3D pose of the target workpiece. The optimization process is completed using the following formula: ; This represents the total number of nodes in the visual intelligent agent. The total number of three-dimensional feature points on the target workpiece. , The first The rotation matrix and translation vector of the agent's camera coordinate system relative to the global coordinate system. For the first The three-dimensional coordinates of each feature point in the global coordinate system For the first The feature point at the th ... Pixel observation coordinates on an agent's image For camera pinhole imaging projection function, For Huber robust kernel function, These are the dynamic motion constraint weighting coefficients. In camera pixel equivalents, This represents the change in the translation vector between adjacent acquisition times. Let be the velocity vector of the target workpiece. This represents the change in the rotation matrix between adjacent acquisition times.

[0011] Furthermore, in the multi-agent collaborative scheduling and global optimization steps, a multi-agent task allocation and collaborative scheduling model based on deep reinforcement learning is constructed. The CPU utilization rate, memory utilization rate, image acquisition frame rate, feature matching accuracy, and localization solution accuracy of each agent node are used as state inputs. The task allocation weights, image acquisition triggering sequence, and computing resource allocation of each node are used as action outputs. The global localization accuracy, solution time, and system load balance are used as reward functions. Through offline pre-training and online self-learning, the task allocation and running parameters of each agent node are dynamically adjusted. When an agent node experiences problems such as image quality degradation, excessive computing load, or insufficient localization accuracy, its localization tasks are automatically allocated to adjacent low-load agent nodes, while the acquisition triggering sequence and fusion weights of each node are adjusted.

[0012] Furthermore, the complex dynamic industrial vision positioning system based on multi-agent collaboration includes the following modules: The multi-agent node deployment module is used to complete the network configuration, time synchronization, camera calibration and spatial coordinate system of distributed visual agent nodes; The multi-source industrial image acquisition and preprocessing module is used to drive each visual intelligent agent node to synchronously acquire industrial scene images and complete the preprocessing of the original images, including denoising, enhancement, and distortion correction. The dynamic target feature extraction and matching module is used to complete the detection and identification of dynamic target workpieces, feature point extraction, cross-node feature matching and mismatch removal; The multi-agent collaborative pose calculation module is used to calculate the 3D pose data of dynamic targets in the global coordinate system based on a multi-view geometric constraint model, combined with feature matching pairs and node calibration parameters. The dynamic interference compensation and optimization module is used to identify the type and intensity of dynamic interference in the scene and perform nonlinear compensation optimization on the solved pose data. The positioning result output and feedback module is used to output the positioning result to the industrial operation execution terminal, and at the same time collect the operation posture deviation data; The multi-agent global collaborative scheduling module is used to dynamically adjust task allocation, triggering timing, and computational weights based on the running status and positioning performance of each agent node.

[0013] Furthermore, the multi-source industrial image acquisition and preprocessing module is equipped with a multi-channel hardware trigger unit, a distributed edge computing unit, and an image quality assessment unit. The multi-channel hardware trigger unit can output multiple synchronous hardware trigger signals to control the industrial cameras of each visual agent node to complete synchronous image acquisition according to a preset time sequence. The distributed edge computing unit is deployed locally on each visual agent node and can complete image preprocessing operations at the node end. The image quality assessment unit can perform multi-dimensional quality scoring on the preprocessed image and automatically trigger the secondary acquisition and parameter adjustment process for low-quality images.

[0014] Furthermore, the multi-agent global collaborative scheduling module incorporates a deep reinforcement learning scheduling model, a load balancing management unit, and a fault redundancy switching unit. The deep reinforcement learning scheduling model can dynamically optimize the task allocation and operating parameters of each agent node based on the real-time operating status and positioning performance of each agent node. The load balancing management unit can monitor the hardware resource usage of each node in real time and dynamically migrate the computing tasks of high-load nodes to low-load nodes. The fault redundancy switching unit can monitor the operating status of each agent node in real time. When a node experiences a hardware failure, communication interruption, or abnormal data acquisition, it automatically switches the positioning task it is undertaking to an adjacent backup agent node and triggers a fault warning.

[0015] Compared with existing technologies, the beneficial effects of this invention are: This invention employs a distributed multi-agent node deployment architecture, configuring each visual agent node with independent edge acquisition and computing capabilities. Basic image preprocessing and feature extraction can be performed locally on each node, reducing the computational burden on the central endpoint and the overall system data transmission bandwidth usage, shortening system processing response latency, and improving overall system stability. Through a similarity calculation method based on multi-dimensional low-level visual feature fusion, combined with epipolar constraints and prior information on dynamic target motion trajectories, the accuracy of feature matching in complex dynamic industrial scenarios can be improved, and the feature mismatch rate under various environmental interferences can be reduced, providing a highly reliable feature matching foundation for subsequent multi-view pose calculation.

[0016] This invention utilizes a multi-view nonlinear joint optimization method that combines the continuous motion characteristics of the target to simultaneously optimize the pose parameters of each agent node and the 3D coordinates of the target workpiece. This reduces the cumulative error in pose calculation under dynamic scenes. The application of a robust kernel function can suppress the negative impact of outlier observations caused by dynamic disturbances on the optimization results, improving the accuracy and stability of dynamic target pose calculation. Through a multi-agent collaborative scheduling mechanism based on deep reinforcement learning, the task allocation, acquisition triggering sequence, and computation weight of each agent node can be dynamically adjusted according to their real-time operating status, achieving overall system load balancing. When some nodes experience performance degradation or operational abnormalities, task migration and allocation can be automatically completed, maintaining stable system operation even under single-node abnormal conditions.

[0017] This invention utilizes a closed-loop feedback mechanism to optimize relevant parameters in the front-end feature extraction and pose calculation stages based on pose deviation data after industrial operations are performed. This enables continuous iterative optimization of positioning accuracy. Nonlinear compensation optimization steps, designed to address various dynamic disturbances in industrial scenarios, enhance the system's adaptability to complex and dynamic industrial environments. Simultaneously, a high-precision time synchronization and spatial joint calibration mechanism provides a unified time and space reference for cross-node feature matching and collaborative pose calculation, eliminating time and space coordinate deviations between multiple nodes. This further improves the accuracy and consistency of multi-agent collaborative positioning, better adapting to the visual positioning needs of various complex and dynamic industrial scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic block diagram of the complex dynamic industrial visual positioning method and system based on multi-agent collaboration proposed in this invention. Figure 2 Detailed flowchart for node deployment, high-precision synchronization, and spatial calibration; Figure 3 A flowchart of the closed-loop process for image preprocessing and adaptive quality assessment; Figure 4 Flowchart for enhanced matching of dynamic features and multi-view pose calculation; Figure 5 Optimize the logic graph for multi-agent global scheduling and reinforcement learning. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0022] Reference Figures 1 to 5 A complex dynamic industrial vision localization method based on multi-agent collaboration includes the following steps: The multi-agent node deployment and initialization steps involve distributing multiple visual intelligent agent nodes with independent edge acquisition and computing capabilities in a dynamic industrial operation scenario, and completing microsecond-level time synchronization, spatial joint calibration and communication network configuration for each node. The multi-source industrial image synchronous acquisition and preprocessing steps involve each visual intelligent agent node synchronously acquiring industrial images in a dynamic operation scenario according to a preset hardware triggering sequence, and sequentially performing denoising, enhancement, and distortion correction preprocessing on the original images to generate a standardized industrial image dataset. The dynamic target feature extraction and matching steps are based on an improved deep convolutional neural network to perform target detection and feature point extraction on standardized industrial images. The feature matching and target association across intelligent agent nodes are completed by combining prior information on the dynamic target motion trajectory and epipolar constraints. The multi-agent collaborative pose calculation step is to construct a multi-view geometric constraint model based on the matched feature point pairs and the spatial calibration parameters of each agent node, and calculate the three-dimensional pose data of the dynamic target in the global coordinate system of the industrial scene. The dynamic interference compensation and positioning result optimization steps are based on dynamic interference parameters such as lighting fluctuations, workpiece motion blur, and local occlusion within the scene. Nonlinear compensation optimization is performed on the calculated 3D pose data to generate initial positioning results. The positioning result output and closed-loop feedback steps output the initial positioning result to the industrial operation execution end, and at the same time collect the pose deviation data after the operation is executed, and feed it back to the feature extraction and pose calculation steps. The multi-agent collaborative scheduling and global optimization steps dynamically adjust the triggering sequence, computational weight, and task allocation of each node based on the computational load, image acquisition quality, and positioning accuracy data of each agent node. This achieves global optimization of multi-agent collaborative positioning and feeds back the scheduling parameters to the multi-source industrial image synchronous acquisition and preprocessing steps, as well as the dynamic target feature extraction and matching steps.

[0023] This invention also includes a dynamic target feature robustness enhancement and accurate matching unit. Addressing issues such as motion blur, abrupt changes in illumination, and partial occlusion in complex dynamic industrial scenarios, it performs multi-dimensional feature fusion and similarity calculation on the extracted feature points. The calculation process is implemented using the following formula: ; For image With images The comprehensive matching similarity of corresponding feature point pairs. These are the gradient feature weight coefficients. , Images , The Middle The gradient direction histogram eigenvectors of each feature point These are the texture feature weight coefficients. , Images , The Middle The gray-level co-occurrence matrix and texture feature vector of each feature point These are the weighting coefficients for local structural features. , Images , The Middle Local binary pattern feature vectors of feature points The feature dimension is the result of uniform normalization for all feature vectors. The spatial context constraint attenuation coefficient, , These are the spatial coordinate vectors of the corresponding feature points in their respective images after epipolar constraint verification. Let be the Euclidean distance between the spatial coordinates of two feature points, where e is the natural constant, approximately equal to 2.718, and + + =1, the weighting coefficient can be dynamically adjusted according to the type of interference in the industrial scene, such as sudden changes in illumination. Set to 0.5 Set to 0.3 Set to 0.2 for motion blur scenes Set to 0.5 Set to 0.3 Set to 0.2 for partial occlusion scenes. Set to 0.5 Set to 0.3 Setting it to 0.2, by integrating the similarity calculation of multi-dimensional low-level visual features and spatial context constraints, it can effectively distinguish between real matching feature points and mismatched feature points, reduce the feature mismatch rate in dynamic industrial scenarios, provide highly reliable feature matching pairs for subsequent multi-view pose solving, and adapt to different types of dynamic industrial operation scenarios.

[0024] This invention also includes a high-precision time synchronization and spatial joint calibration unit for multiple agent nodes. Through a dual synchronization mechanism combining hardware trigger signals and high-precision network time protocols, the image acquisition trigger synchronization error of each visual agent node is controlled within 1 microsecond. At the same time, based on a 3D calibration board and hand-eye calibration algorithm, the intrinsic parameters and distortion parameters of the camera of each agent node are individually calibrated, as well as the joint calibration of the camera coordinate system of each node and the global coordinate system of the industrial scene are completed. A unified global spatial coordinate system is constructed to eliminate spatial coordinate deviations between multiple agent nodes and provide a unified spatial reference for cross-node feature matching and collaborative pose calculation.

[0025] In this invention, during the multi-source industrial image synchronous acquisition and preprocessing step, an adaptive bilateral filtering algorithm is used to denoise the original image to address image noise caused by welding slag splashes, dust interference, and lens vibration in industrial scenes. This algorithm filters out impulse noise and Gaussian noise while preserving the edge features of the target workpiece. To address the problem of insufficient image contrast caused by illumination fluctuations, a contrast-limited adaptive histogram equalization algorithm is used to enhance the image. For image distortion caused by camera lens distortion and installation angle, pixel-level distortion correction is performed based on the distortion coefficients involved in the camera. Simultaneously, a quality score is performed on the preprocessed image. The scoring indicators include image sharpness, contrast, and effective target proportion. Only standardized images with a quality score higher than a preset threshold are input into the subsequent feature extraction stage. Images with a quality score below the threshold trigger the corresponding intelligent agent node's acquisition parameter readjustment and secondary acquisition process.

[0026] In this invention, the dynamic target feature extraction and matching step employs an improved YOLO (You Only LookOnce) deep convolutional neural network to perform real-time detection of dynamic targets, identify and locate the target workpiece region in the image, and remove invalid feature points from the background region. Then, an improved ORB (OrientedFAST and RotatedBRIEF) algorithm is used to perform FAST (Features from Accelerated Segment Test) feature point detection and BRIEF (Binary Robust Independent Elementary Features) descriptor generation on the target workpiece region. Simultaneously, a Kalman filter algorithm is used to predict the motion trajectory of the dynamic target, generating the prior motion region of the target workpiece between adjacent frames and across agent nodes. Finally, a random sampling consensus algorithm is used to remove mismatched pairs from the matching results.

[0027] In this invention, during the multi-agent collaborative pose calculation and global optimization process, a multi-view nonlinear joint optimization objective function is constructed. Global bundle adjustment optimization is performed on the pose observations of multiple agent nodes and the 3D pose of the target workpiece to reduce the cumulative positioning error in dynamic scenes. The optimization process is achieved through the following formula: ; This represents the total number of nodes in the visual intelligent agent. The total number of three-dimensional feature points on the target workpiece. , The first The rotation matrix and translation vector of the agent's camera coordinate system relative to the global coordinate system. For the first The three-dimensional coordinates of each feature point in the global coordinate system For the first The feature point at the th ... Pixel observation coordinates on an agent's image For camera pinhole imaging projection function, For Huber robust kernel function, These are the dynamic motion constraint weighting coefficients. This refers to the camera pixel equivalent, which is the physical distance in the world coordinate system corresponding to a unit pixel in the image plane, expressed in pixels per meter. This represents the change in the translation vector between adjacent acquisition times. Let be the velocity vector of the target workpiece. The rotation matrix change at adjacent acquisition times is represented by the objective function of minimizing the reprojection error and the dynamic motion constraint error. This allows for the simultaneous optimization of the pose parameters of each agent node and the 3D coordinates of the target workpiece. The two errors are additive after being standardized by pixel equivalents. The robust kernel function can effectively suppress the negative impact of out-of-point observations caused by dynamic interference on the optimization results. The dynamic motion constraint term can fit the continuous motion characteristics of the target workpiece in industrial scenarios, improving the pose calculation accuracy and stability in high-speed dynamic scenarios.

[0028] In this invention, the multi-agent collaborative scheduling and global optimization steps involve constructing a multi-agent task allocation and collaborative scheduling model based on deep reinforcement learning. The model uses the CPU utilization rate, memory usage rate, image acquisition frame rate, feature matching accuracy, and localization accuracy of each agent node as state inputs, and the task allocation weights, image acquisition triggering sequences, and computational resource allocations of each node as action outputs. The reward functions are global localization accuracy, computation time, and system load balancing. Through offline pre-training and online self-learning, the model dynamically adjusts the task allocation and operating parameters of each agent node. When an agent node experiences problems such as image quality degradation, excessive computational load, or insufficient localization accuracy, its localization tasks are automatically allocated to adjacent low-load agent nodes. Simultaneously, the acquisition triggering sequences and fusion weights of each node are adjusted, enabling the system to sustainably maintain high localization accuracy and low computational latency in complex and dynamic industrial scenarios, thus achieving adaptive global optimization of the multi-agent collaborative localization system.

[0029] This invention includes the following modules: The multi-agent node deployment module is used to complete the network configuration, time synchronization, camera calibration and spatial coordinate system integration of distributed visual agent nodes, and to build the hardware foundation and spatial reference for multi-agent collaborative positioning. The multi-source industrial image acquisition and preprocessing module is used to drive each visual intelligent agent node to synchronously acquire industrial scene images, complete the denoising, enhancement, and distortion correction preprocessing of the original images, and generate a standardized industrial image dataset. The dynamic target feature extraction and matching module is used to complete the detection and identification of dynamic target workpieces, feature point extraction, cross-node feature matching and mismatch elimination, and output highly reliable feature matching pairs. The multi-agent collaborative pose calculation module is used to calculate the three-dimensional pose data of the dynamic target in the global coordinate system based on the multi-view geometric constraint model, combined with feature matching pairs and node calibration parameters, and to perform global nonlinear optimization. The dynamic interference compensation and optimization module is used to identify the type and intensity of dynamic interference in the scene, perform nonlinear compensation optimization on the calculated pose data, and generate the final positioning result of the target workpiece. The positioning result output and feedback module is used to output the positioning result to the industrial operation execution end, and at the same time collect the operation posture deviation data and feed it back to the front-end processing stage to form a closed-loop optimization. The multi-agent global collaborative scheduling module is used to dynamically adjust task allocation, trigger timing and computation weight based on the running status and positioning performance of each agent node, so as to realize global collaborative optimization of the multi-agent system.

[0030] In this invention, the multi-source industrial image acquisition and preprocessing module is equipped with a multi-channel hardware triggering unit, a distributed edge computing unit, and an image quality assessment unit. The multi-channel hardware triggering unit can output multiple synchronous hardware triggering signals to control the industrial cameras of each visual agent node to complete synchronous image acquisition according to a preset timing sequence. The synchronous triggering accuracy can reach 1 microsecond. The distributed edge computing unit is deployed locally on each visual agent node and can complete image preprocessing operations at the node end, reducing the computing pressure and data transmission bandwidth occupation at the central end. The image quality assessment unit can perform multi-dimensional quality scoring on the preprocessed image and automatically trigger the secondary acquisition and parameter adjustment process of low-quality images, so that the image data input to subsequent stages meets the positioning accuracy requirements.

[0031] In this invention, the multi-agent global collaborative scheduling module incorporates a deep reinforcement learning scheduling model, a load balancing management unit, and a fault redundancy switching unit. The deep reinforcement learning scheduling model can dynamically optimize the task allocation and operating parameters of each agent node based on the real-time operating status and positioning performance of each agent node. The load balancing management unit can monitor the hardware resource usage of each node in real time and dynamically migrate the computing tasks of high-load nodes to low-load nodes to maintain the overall load balance of the system. The fault redundancy switching unit can monitor the operating status of each agent node in real time. When a node experiences hardware failure, communication interruption, or abnormal data acquisition, it automatically switches the positioning tasks it undertakes to an adjacent backup agent node and triggers a fault warning, enabling the system to continue to operate stably even in the event of a single node failure, thus maintaining the continuity of industrial positioning operations.

[0032] The following two examples further illustrate the specific implementation of this system: Example 1: Dynamic Industrial Vision Positioning Example for Automobile Body Welding Production Line This embodiment is applied to the body-in-white welding station in automobile manufacturing. Within the welding station, the body-in-white moves continuously and at a constant speed via a ground conveyor line, with a speed of up to 0.2 meters per second. The visual positioning system needs to complete high-precision three-dimensional positioning of the welding points of the body-in-white to provide accurate pose data for the industrial welding robot. The working scene contains complex dynamic interference factors such as welding arc light, welding slag spatter, dust interference, and partial tooling obstruction. At the same time, the lateral span of the welding station reaches 6 meters, requiring continuous dynamic positioning within a large scene range.

[0033] After system startup, the deployment and initialization of multi-agent nodes are completed first. Four groups of visual agent nodes are distributed around the welding station. Each group of nodes is equipped with an industrial area array camera, an edge computing unit, a multi-channel hardware trigger unit, and an industrial communication module. All nodes are configured for communication networking via industrial Ethernet. A dual synchronization mechanism combining hardware trigger signals and a high-precision network time protocol is used to complete the time synchronization configuration of each node, keeping the synchronization error of image acquisition triggering within 1 microsecond. Based on a 3D calibration board and hand-eye calibration algorithms, the intrinsic parameters and distortion parameters of each node's camera are individually calibrated, as well as the joint calibration of the camera coordinate system of each node and the global coordinate system of the welding station are performed sequentially, constructing a unified global spatial coordinate system and eliminating spatial coordinate deviations between multiple nodes.

[0034] Each visual agent node synchronously acquires industrial images of the moving body-in-white within the welding station according to a preset hardware trigger sequence. For the acquired raw images, preprocessing is performed in the local edge computing unit of each node. To address image noise caused by welding slag spatter, dust interference, and lens vibration, an adaptive bilateral filtering algorithm is used for denoising, filtering out impulse noise and Gaussian noise while preserving the edge features of the welding points on the body-in-white. To address illumination fluctuations and insufficient image contrast caused by welding arc light, a contrast-limited adaptive histogram equalization algorithm is used for image enhancement. For image distortion caused by camera lens distortion and installation angle, pixel-level distortion correction is performed based on the distortion coefficients involved in the camera. After preprocessing, a multi-dimensional quality score is applied to the images, including image sharpness, contrast, and effective target proportion. Only standardized images with quality scores above a preset threshold are input into the subsequent feature extraction stage. Images with quality scores below the threshold trigger the corresponding node's acquisition parameters readjustment and a secondary acquisition process.

[0035] In the dynamic target feature extraction and matching stage, an improved YOLO deep convolutional neural network is used to perform real-time detection of the welding area of ​​the body-in-white, identify and locate the target area of ​​the welding point in the image, and remove invalid feature points in the background area. Then, an improved ORB algorithm is used to perform FAST feature point detection and BRIEF descriptor generation on the target area. The Kalman filter algorithm is combined to predict the motion trajectory of the body-in-white, generating the motion prior region of the body-in-white between adjacent frames and cross-agent nodes, constraining the search range of feature matching within the motion prior region, thus narrowing the search space of feature matching. At the same time, epipolar constraints are used to complete feature matching and target association across intelligent agent nodes. Based on the random sampling consensus algorithm, mismatched pairs in the matching results are removed, and the feature matching results are purified.

[0036] In the multi-agent collaborative pose calculation stage, a multi-view geometric constraint model is constructed based on the matched feature point pairs and the spatial calibration parameters of each node. This model calculates the 3D pose data of the welding points on the body-in-white in the global coordinate system. Simultaneously, a multi-view nonlinear joint optimization objective function is constructed, and global bundle adjustment optimization is performed on the pose observations of multiple nodes and the target 3D pose to reduce the cumulative positioning error in dynamic scenes. For dynamic interference parameters such as welding arc light, local occlusion, and motion blur within the scene, nonlinear compensation optimization is performed on the calculated 3D pose data to generate initial positioning results, which are then output to the execution end of the industrial welding robot in real time. The system synchronously collects pose deviation data after the welding robot's operation and feeds it back to the feature extraction and pose calculation stages to complete closed-loop optimization.

[0037] In the multi-agent collaborative scheduling and global optimization phase, a multi-agent task allocation and collaborative scheduling model based on deep reinforcement learning is constructed. The model uses each node's CPU utilization, memory usage, image acquisition frame rate, feature matching accuracy, and localization accuracy as state inputs, and each node's task allocation weights, image acquisition triggering sequence, and computational resource allocation as action outputs. The reward function is global localization accuracy, computation time, and system load balancing. Through offline pre-training and online self-learning, the model dynamically adjusts the task allocation and operating parameters of each node. When a node experiences image quality degradation, excessive computational load, or insufficient localization accuracy, its localization tasks are automatically reassigned to adjacent low-load nodes. Simultaneously, the acquisition triggering sequence and fusion weights of each node are adjusted, enabling the system to continuously maintain high localization accuracy and low computation latency in complex dynamic welding scenarios.

[0038] Table 1: Performance of Positioning in Automotive Welding Production Line Scenarios

[0039] Table 1 shows the data from statistical analysis of 1000 consecutive welding operations conducted by the two systems in this embodiment under the same welding station, production line operating speed, and environmental interference conditions. Traditional centralized multi-camera positioning systems employ a unified central computing architecture, which suffers from concentrated computational pressure and high response latency. Under interference from welding arc light and weld spatter, feature matching accuracy is low, and a single node failure can directly lead to system-wide failure, failing to meet the demands of continuous welding operations. The system of this invention, through distributed edge computing and multi-agent collaborative scheduling, improves system response speed and positioning accuracy, exhibits superior anti-interference capabilities and operational stability, and is fully adaptable to the dynamic positioning requirements of automotive welding production lines.

[0040] Example 2: Dynamic Industrial Vision Positioning Example for High-Speed ​​Sorting Production Line of 3C Products This embodiment is applied to a high-speed sorting station for mobile phone frames in 3C product manufacturing. The mobile phone frames are continuously transported at high speed via a conveyor belt, with a conveying speed of up to 1 meter per second. The vision positioning system needs to complete high-precision three-dimensional positioning of the mobile phone frame gripping points to provide accurate pose data for the high-speed parallel sorting robot. The working scene contains complex dynamic interference factors such as image blurring caused by high-speed movement, fluctuations in production line lighting, partial occlusion caused by product stacking, and conveyor belt vibration. The longitudinal conveying distance of the sorting station reaches 8 meters, requiring high-speed continuous dynamic positioning over a long distance.

[0041] After system startup, the deployment and initialization of multi-agent nodes are completed first. Six groups of visual agent nodes are distributed and deployed above the conveyor belt of the sorting station. Each group of nodes is equipped with an industrial linear scan camera, an edge computing unit, a multi-channel hardware trigger unit, and an industrial communication module. All nodes are configured for communication networking via industrial Ethernet. A dual synchronization mechanism combining hardware trigger signals and a high-precision network time protocol is used to complete the time synchronization configuration of each node, keeping the synchronization error of image acquisition triggering within 1 microsecond. Based on a 3D calibration board and hand-eye calibration algorithms, the intrinsic parameters and distortion parameters of each node's camera are individually calibrated, as well as the joint calibration of the camera coordinate system of each node and the global coordinate system of the sorting station, constructing a unified global spatial coordinate system and eliminating spatial coordinate deviations between multiple nodes.

[0042] Each visual intelligent agent node, following a preset hardware trigger sequence, synchronizes with the conveyor belt encoder signal to acquire industrial images of high-speed moving mobile phone frames within the sorting station. For the acquired raw images, preprocessing is performed in the node's local edge computing unit. To address image noise and motion blur caused by conveyor belt vibration and high-speed movement, an adaptive bilateral filtering algorithm is used for denoising, preserving edge features of the phone frame's capture points while filtering out noise. To address illumination fluctuations and insufficient image contrast caused by production line lighting, a contrast-limited adaptive histogram equalization algorithm is used for image enhancement. For image distortion caused by camera lens distortion and installation angle, pixel-level distortion correction is performed based on the camera's in-camera distortion coefficients. After preprocessing, a multi-dimensional quality score is applied to the images, including image sharpness, contrast, and effective target proportion. Only standardized images with quality scores above a preset threshold are input into the subsequent feature extraction stage. Images with quality scores below the threshold trigger the corresponding node's acquisition parameters readjustment and a secondary acquisition process.

[0043] In the dynamic target feature extraction and matching stage, an improved YOLO deep convolutional neural network is used to perform real-time detection of the phone's mid-frame. This identifies and locates the target region of the mid-frame in the image, removing invalid feature points from background areas and stacked products. Then, an improved ORB algorithm is used to perform FAST feature point detection and BRIEF descriptor generation on the target region. A Kalman filter algorithm is combined to predict the motion trajectory of the phone's mid-frame, generating a priori motion region for the mid-frame between adjacent frames and across agent nodes. This constrains the search range for feature matching within the prior motion region, narrowing the search space. Simultaneously, epipolar constraints are used to complete feature matching and target association across agent nodes. A random sampling consensus algorithm is used to remove mismatches from the matching results, refining the feature matching results.

[0044] In the multi-agent collaborative pose calculation stage, a multi-view geometric constraint model is constructed based on the matched feature point pairs and the spatial calibration parameters of each node. This model calculates the 3D pose data of the phone's frame capture points in the global coordinate system. Simultaneously, a multi-view nonlinear joint optimization objective function is constructed, and global bundle adjustment optimization is performed on the pose observations of multiple nodes and the target 3D pose to reduce the cumulative positioning error in high-speed dynamic scenes. For dynamic interference parameters such as motion blur, illumination fluctuations, and stacked occlusion within the scene, nonlinear compensation optimization is performed on the calculated 3D pose data to generate initial positioning results. These initial positioning results are output to the execution end of the parallel sorting robot in real time. The system synchronously collects pose deviation data after the sorting robot's operation and feeds it back to the feature extraction and pose calculation stages to complete closed-loop optimization.

[0045] In the multi-agent collaborative scheduling and global optimization phase, a multi-agent task allocation and collaborative scheduling model based on deep reinforcement learning is constructed. The model uses the CPU utilization, memory usage, image acquisition frame rate, feature matching accuracy, and localization accuracy of each node as state inputs, and the task allocation weights, image acquisition triggering sequences, and computational resource allocation of each node as action outputs. The reward functions are global localization accuracy, computation time, and system load balancing. Through offline pre-training and online self-learning, the model dynamically adjusts the task allocation and operating parameters of each node. When a node experiences image quality degradation, excessive computational load, or insufficient localization accuracy, its localization tasks are automatically reassigned to adjacent low-load nodes. Simultaneously, the acquisition triggering sequences and fusion weights of each node are adjusted, enabling the system to continuously maintain high localization accuracy and low computation latency in high-speed dynamic sorting scenarios.

[0046] Table 2: Comparison of Positioning Performance in 3C Sorting Production Line Scenarios

[0047] Table 2 shows the data from the statistical analysis of 2000 consecutive sorting operations conducted by the two systems in this embodiment under the same sorting station, the same conveyor belt operating parameters, and the same environmental interference conditions. Traditional single-camera high-speed positioning systems are limited by their imaging field of view and computing power, and can only adapt to low-speed production line operation scenarios. Under high-speed movement, they are prone to motion blur, resulting in a significant decrease in feature matching accuracy and positioning precision, and insufficient positioning stability during long-term continuous operation. The system of this invention, through distributed multi-node deployment and multi-agent collaborative optimization, improves positioning accuracy and response speed in high-speed dynamic scenarios, adapting to the needs of higher-speed production line operations. It exhibits superior positioning stability during long-term continuous operation and can fully meet the dynamic positioning requirements of high-speed sorting production lines for 3C products.

[0048] refer to Figure 1 This diagram illustrates the closed-loop process of the system's entire lifecycle, from hardware deployment to positioning output. First, visual agents are deployed in a distributed manner at the industrial site, achieving microsecond-level synchronization and calibration. The system synchronously acquires images and performs standardized preprocessing. Then, deep learning is used to extract features and, combined with motion priors, cross-node matching is performed. The target's 3D pose is calculated using multi-view geometric constraints, and nonlinear compensation is performed to address dynamic disturbances such as illumination and occlusion. After the positioning results are output to the execution end, the node load and computational weights are dynamically adjusted through bias feedback and a reinforcement learning scheduling model, achieving global adaptive optimization of system performance.

[0049] Reference Figure 2 This figure details the key technical nodes in the system initialization phase. The system utilizes a dual mechanism of hardware trigger signals and network time protocols to reduce the synchronization error of data acquisition at each node to within 1 microsecond. In the spatial dimension, through joint observation of the calibration board, single-camera intrinsic parameter calibration and multi-camera hand-eye calibration are completed respectively. Ultimately, all agent coordinate systems are normalized to a unified global coordinate system for the industrial scene, eliminating spatial bias and providing a unified spatial reference for subsequent cross-node feature association.

[0050] Reference Figure 3 This figure illustrates how the system addresses image quality issues arising from harsh industrial environments (such as welding slag, dust, and vibration). The original image undergoes bilateral filtering for noise reduction to preserve edges and is enhanced using Limiting Contrast Adaptive Histogram Equalization (CLAHE). The preprocessed image then enters a quality assessment unit, scoring it based on three dimensions: sharpness, contrast, and effective target proportion. Only high-scoring images proceed to feature extraction; low-quality images trigger an adaptive adjustment mechanism, controlling the agent to readjust acquisition parameters and perform a second acquisition to ensure the reliability of input for subsequent localization algorithms.

[0051] Reference Figure 4This figure illustrates the robust localization logic in complex dynamic scenes. The system employs an improved convolutional neural network (YOLO) to delineate the target region, utilizes the ORB algorithm to extract features, and combines Kalman filtering to predict motion trajectories to narrow the search space. During the matching process, multi-dimensional features such as gradient, texture, and structure are fused, and epipolar constraints are added to effectively filter out mismatched points. In the pose calculation stage, a joint objective function incorporating reprojection error and dynamic motion constraints is constructed, and a robust kernel function is used to suppress outlier observations, achieving high-precision 3D pose calculation.

[0052] refer to Figure 5 This diagram illustrates the system's intelligent scheduling and fault-tolerance mechanisms. The system monitors the hardware resource usage, data acquisition quality, and positioning accuracy of each node in real time, using these as input states for the reinforcement learning model. The model dynamically adjusts the task weights and trigger sequences of each node through online learning. When a single node failure or excessive load is detected, the load balancing management unit smoothly migrates tasks to low-load nodes and activates a backup agent through a fault redundancy switching mechanism to ensure the continuity of positioning operations and achieve optimal allocation and self-healing of the overall system performance.

[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A complex dynamic industrial visual positioning method based on multi-agent collaboration, characterized in that, Includes the following steps: The multi-agent node deployment and initialization steps involve the distributed deployment of multiple visual agent nodes with independent edge acquisition and computing capabilities within a dynamic industrial operation scenario. The multi-source industrial image synchronous acquisition and preprocessing steps involve each visual intelligent agent node synchronously acquiring industrial images in the dynamic operation scene according to the preset hardware triggering sequence, and performing denoising, enhancement, and distortion correction preprocessing on the original images in sequence. The dynamic target feature extraction and matching steps are based on an improved deep convolutional neural network to perform target detection and feature point extraction on standardized industrial images. The feature matching and target association across intelligent agent nodes are completed by combining prior information on the dynamic target motion trajectory and epipolar constraints. The multi-agent collaborative pose calculation step constructs a multi-view geometric constraint model based on the matched feature point pairs and the spatial calibration parameters of each agent node. The dynamic interference compensation and positioning result optimization steps involve performing nonlinear compensation optimization on the calculated 3D pose data based on dynamic interference parameters such as lighting fluctuations, workpiece motion blur, and local occlusion within the scene. The positioning result output and closed-loop feedback steps output the initial positioning result to the industrial operation execution end, and at the same time collect the pose deviation data after the operation is executed; The multi-agent collaborative scheduling and global optimization steps dynamically adjust the triggering sequence, computational weight, and task allocation of each node based on the computational load, image acquisition quality, and positioning accuracy data of each agent node. The scheduling parameters are then fed back to the multi-source industrial image synchronous acquisition and preprocessing steps, as well as the dynamic target feature extraction and matching steps.

2. The complex dynamic industrial visual positioning method based on multi-agent collaboration according to claim 1, characterized in that, It also includes steps for enhancing the robustness and accurate matching of dynamic target features. Addressing issues such as motion blur, abrupt changes in illumination, and partial occlusion in complex dynamic industrial scenarios, it performs multi-dimensional feature fusion and similarity calculation on the extracted feature points. The calculation process is completed using the following formula: ; For image With images The comprehensive matching similarity of corresponding feature point pairs. These are the gradient feature weight coefficients. , Images , The Middle The gradient direction histogram eigenvectors of each feature point These are the texture feature weight coefficients. , Images , The Middle The gray-level co-occurrence matrix and texture feature vector of each feature point These are the weighting coefficients for local structural features. , Images , The Middle Local binary pattern feature vectors of feature points The feature dimension is the result of uniform normalization for all feature vectors. The spatial context constraint attenuation coefficient, , These are the spatial coordinate vectors of the corresponding feature points in their respective images after epipolar constraint verification. Let Euclidean distance be the spatial coordinates of two feature points.

3. The complex dynamic industrial visual positioning method based on multi-agent collaboration according to claim 1, characterized in that, It also includes high-precision time synchronization and joint spatial calibration steps for multiple agent nodes. Through a dual synchronization mechanism combining hardware trigger signals and high-precision network time protocols, and based on a 3D calibration board and hand-eye calibration algorithm, it completes the individual calibration of camera intrinsic parameters and distortion parameters of each agent node, as well as the joint calibration of the camera coordinate system of each node and the global coordinate system of the industrial scene, to build a unified global spatial coordinate system and eliminate spatial coordinate deviations between multiple agent nodes.

4. The complex dynamic industrial visual positioning method based on multi-agent collaboration according to claim 1, characterized in that, In the multi-source industrial image synchronous acquisition and preprocessing steps, an adaptive bilateral filtering algorithm is used to denoise the original image to address image noise caused by welding slag splashes, dust interference, and lens vibration in industrial scenes. While filtering out impulse noise and Gaussian noise, the edge features of the target workpiece are preserved. To address the problem of insufficient image contrast caused by illumination fluctuations, a contrast-limited adaptive histogram equalization algorithm is used to enhance the image. For image distortion caused by camera lens distortion and installation angle, pixel-level distortion correction is performed based on the distortion coefficients involved in the camera. At the same time, a quality score is performed on the preprocessed image.

5. The complex dynamic industrial visual positioning method based on multi-agent collaboration according to claim 1, characterized in that, In the dynamic target feature extraction and matching step, an improved YOLO deep convolutional neural network is used to complete the real-time detection of dynamic targets, identify and locate the target workpiece region in the image, remove invalid feature points in the background region, and then perform FAST feature point detection and BRIEF descriptor generation on the target workpiece region through an improved ORB algorithm. At the same time, the Kalman filter algorithm is combined to predict the motion trajectory of the dynamic target, generate the motion prior region of the target workpiece between adjacent frames and cross-agent nodes, and remove mismatched pairs in the matching results based on the random sampling consensus algorithm.

6. The complex dynamic industrial visual positioning method based on multi-agent collaboration according to claim 1, characterized in that, In the multi-agent collaborative pose calculation and global optimization step, a multi-view nonlinear joint optimization objective function is constructed. Global bundle adjustment optimization is performed on the pose observations of multiple agent nodes and the 3D pose of the target workpiece. The optimization process is completed by the following formula: ; This represents the total number of nodes in the visual intelligent agent. The total number of three-dimensional feature points on the target workpiece. , The first The rotation matrix and translation vector of the agent's camera coordinate system relative to the global coordinate system. For the first The three-dimensional coordinates of each feature point in the global coordinate system For the first The feature point at the th ... Pixel observation coordinates on an agent's image For camera pinhole imaging projection function, For Huber robust kernel function, These are the dynamic motion constraint weighting coefficients. In camera pixel equivalents, This represents the change in the translation vector between adjacent acquisition times. Let the velocity vector of the target workpiece be denoted as . This represents the change in the rotation matrix between adjacent acquisition times.

7. The complex dynamic industrial visual positioning method based on multi-agent collaboration according to claim 1, characterized in that, In the multi-agent collaborative scheduling and global optimization step, a multi-agent task allocation and collaborative scheduling model based on deep reinforcement learning is constructed. The CPU utilization rate, memory utilization rate, image acquisition frame rate, feature matching accuracy, and localization solution accuracy of the agent nodes are used as state inputs. The task allocation weights of each node, image acquisition triggering sequence, and computing resource allocation are used as action outputs. The global localization accuracy, solution time, and system load balance are used as reward functions. Through offline pre-training and online self-learning, the task allocation and running parameters of each agent node are dynamically adjusted. When an agent node experiences problems such as image quality degradation, excessive computing load, or insufficient localization accuracy, its localization tasks are automatically allocated to adjacent low-load agent nodes, while the acquisition triggering sequence and fusion weights of each node are adjusted.

8. A complex dynamic industrial vision positioning system based on multi-agent collaboration, characterized in that, The method for complex dynamic industrial visual positioning based on multi-agent collaboration, as described in any one of claims 1-7, includes the following modules: The multi-agent node deployment module is used to complete the network configuration, time synchronization, camera calibration and spatial coordinate system of distributed visual agent nodes; The multi-source industrial image acquisition and preprocessing module is used to drive each visual intelligent agent node to synchronously acquire industrial scene images and complete the preprocessing of the original images, including denoising, enhancement, and distortion correction. The dynamic target feature extraction and matching module is used to complete the detection and identification of dynamic target workpieces, feature point extraction, cross-node feature matching and mismatch removal; The multi-agent collaborative pose calculation module is used to calculate the 3D pose data of dynamic targets in the global coordinate system based on a multi-view geometric constraint model, combined with feature matching pairs and node calibration parameters. The dynamic interference compensation and optimization module is used to identify the type and intensity of dynamic interference in the scene and perform nonlinear compensation optimization on the solved pose data. The positioning result output and feedback module is used to output the positioning result to the industrial operation execution terminal, and at the same time collect the operation posture deviation data; The multi-agent global collaborative scheduling module is used to dynamically adjust task allocation, triggering timing, and computational weights based on the running status and positioning performance of each agent node.

9. The complex dynamic industrial vision positioning system based on multi-agent collaboration according to claim 8, characterized in that, The multi-source industrial image acquisition and preprocessing module is equipped with a multi-channel hardware trigger unit, a distributed edge computing unit, and an image quality assessment unit. The multi-channel hardware trigger unit can output multiple synchronous hardware trigger signals to control the industrial cameras of each visual agent node to complete synchronous image acquisition according to a preset timing sequence. The distributed edge computing unit is deployed locally on each visual agent node and can complete image preprocessing operations at the node end. The image quality assessment unit can perform multi-dimensional quality scoring on the preprocessed image and automatically trigger the secondary acquisition and parameter adjustment process for low-quality images.

10. The complex dynamic industrial vision positioning system based on multi-agent collaboration according to claim 8, characterized in that, The multi-agent global collaborative scheduling module incorporates a deep reinforcement learning scheduling model, a load balancing management unit, and a fault redundancy switching unit. The deep reinforcement learning scheduling model can dynamically optimize the task allocation and operating parameters of each agent node based on its real-time operating status and positioning performance. The load balancing management unit can monitor the hardware resource usage of each node in real time and dynamically migrate the computing tasks of high-load nodes to low-load nodes. The fault redundancy switching unit can monitor the operating status of each agent node in real time. When a node experiences a hardware failure, communication interruption, or abnormal data acquisition, it automatically switches the positioning task it is undertaking to an adjacent backup agent node and triggers a fault warning.