Cluster vision method based on SPV and redundancy transmission

Abstract

The invention discloses a cluster vision method based on SPV and redundancy transmission. According to the method, a small number of devices are dynamically selected to construct the visual positioning matrix, only four devices are needed to cover a local area, and other devices do not need visual modules. According to the design, the video transmission requirement is reduced by more than 70%, and the cluster scale can be expanded to more than one hundred by the released bandwidth resources. Meanwhile, the dynamic equipment reconfiguration mechanism adjusts the number of visual nodes in real time according to the target activeness, resource waste is further avoided on the premise of guaranteeing precision, decoupling of bandwidth occupation and cluster scale is achieved, and the redundancy transmission mechanism checks data in parallel through double links; and the standby link ensures low-delay return through transfer of the adjacent nodes. Even if a single node fails or a network fluctuates, the system can seamlessly switch links within 50 milliseconds, so that the system can still stably output a high-precision positioning result in a complex environment, and the industrial-grade reliability requirement is met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cluster vision, and particularly relates to a cluster vision method based on SPV and redundant transmission. BACKGROUND

[0002] Cluster vision positioning is a spatial positioning technology based on the cooperative work of multiple visual sensors. It deploys cameras on multiple nodes (such as unmanned aerial vehicles, robots or mobile devices) in a cluster, uses these sensors to collect image information of targets or environments from different angles. The system uses computer vision algorithms to extract, match and reconstruct features from these multi-view images, and combines the communication and positioning mechanisms (such as GPS, UWB or relative ranging) within the cluster to accurately calculate the position of the target in the global coordinate system, or the spatial attitude and relative position relationship of each node itself. This technology can significantly improve the positioning accuracy and robustness in complex and dynamic environments, and is widely used in cooperative work, environmental mapping, navigation and obstacle avoidance.

[0003] However, the existing technology excessively relies on single-device vision modules and full-volume video transmission: the traditional first-person perspective method requires each device in the cluster to be equipped with a vision module and to independently transmit high-definition video streams, resulting in a linear increase in communication bandwidth, which severely restricts the expansion of the cluster size; at the same time, a single device failure or network fluctuation will cause the positioning to be interrupted, and there is a lack of redundant mechanism to ensure the continuous operation of the system, making it difficult to meet the demand for high reliability in positioning stability. SUMMARY

[0004] The purpose of the present application is to solve the problems mentioned above, and to provide a cluster vision method based on SPV and redundant transmission.

[0005] The technical solution adopted by the present application is as follows: a cluster vision method based on SPV and redundant transmission, the method comprising the following steps: S1: defining the cluster size and the local operation range, determining the candidate nodes that need to be installed with vision devices, and ensuring that the key areas are covered. This step provides a spatial constraint basis for the subsequent construction of the vision matrix.

[0006] S2: dynamically selecting the optimal m devices (m>4) to construct a vision positioning matrix according to the overlapping nature of the local range. The positions and coordinate systems of the devices in the matrix are synchronized in real time, providing a basic framework for multi-view cooperative positioning.

[0007] S3: all devices in the matrix synchronously detect the target and calculate the local coordinates (x_i, y_i) and the deflection angle of the target based on their respective coordinate systems. The detection results of each device are shared in real time through cluster communication, providing data input for global positioning.

[0008] S4: Use shared data to generate multiple sets of distance estimates for the same target. Integrate the estimates through a weighted fusion algorithm (such as Kalman filtering) to preliminarily eliminate single-device errors and improve positioning robustness.

[0009] S5: Input the fused estimates into the redundancy correction model: generate a correction value for every two sets of estimates, and optimize the three-dimensional coordinates (X_0, Y_0, D_0) of the target through mean square estimation. This step relies on the fusion results of step S4 to ensure that the positioning accuracy reaches the centimeter level.

[0010] S7: Based on the target motion trajectory and cluster load state, real-time adjust the role of the visual device, including: Dynamically assign visual nodes to high-activity areas; Reduce visual nodes in low-activity areas and switch data sources to adjacent devices; Predict adjustment strategies through reinforcement learning models to minimize bandwidth occupation and maintain positioning accuracy.

[0011] S8: Design dual-redundancy communication links, including: Main link: directly return control results from visual nodes; Backup link: non-visual nodes transfer data through adjacent visual nodes; Two links verify data consistency in parallel to ensure real-time feedback and fault tolerance.

[0012] S9: Input the positioning results into the guidance control system to generate operation instructions. At the same time, based on historical positioning errors and bandwidth occupation data, iteratively optimize the configuration strategy of the visual matrix and feedback to step S7 to form an adaptive closed loop.

[0013] In a preferred embodiment, in step S1, cluster initialization needs to accurately define the local operation range, set the rectangular area side length parameters Ld and Lh between 10 meters and 100 meters, and the specific value is dynamically adjusted according to the actual scene. Key area coverage requires no blind area, candidate visual node positions are pre-calibrated through RTK-GPS or laser ranging technology, and positioning accuracy is strictly controlled within centimeters. The communication network uses 5G or special wireless protocol, with a transmission delay upper limit of 20 milliseconds to ensure real-time collaboration. The node deployment density needs to meet the minimum spacing of 2 meters to avoid signal interference, while leaving redundant nodes to deal with sudden failures.

[0014] In a preferred embodiment, in step S2, the visual positioning matrix is dynamically composed of at least 4 nodes, the system analyzes the node position overlap in real time, and automatically selects the optimal nodes to construct the matrix. The position synchronization error is ≤1 cm, the matrix update frequency is 1 time per second, and it adapts to the rapid movement of the target. The main node is responsible for data coordination, and the standby node seamlessly switches within 50 milliseconds in case of failure. The lower limit of the node spacing is 2 meters, and the matrix construction algorithm preferentially selects nodes with high field of view overlap rate and load balancing to ensure global coverage stability.

[0015] In a preferred embodiment, in step S3, all nodes in the matrix synchronously capture the target at a frequency of 30 frames per second, and a lightweight YOLOv5 model is used for efficient detection. Each node calculates the target coordinates (xi, yi) and three-dimensional deviation angle based on the local coordinate system, and the detection confidence threshold is set to 0.8, and data below the threshold is automatically discarded. The target data is shared in real time through the time-stamped UDP protocol, with a synchronization error of ≤10 milliseconds, ensuring the spatiotemporal consistency of multi-source data.

[0016] In a preferred embodiment, in step S4, the fusion stage uses an adaptive Kalman filter algorithm, with a process noise covariance QQ initial value of 0.1 and a measurement noise covariance RR of 0.5, and dynamically adjusts the parameters according to the target motion state. The weight of the near target node is set to 0.6, and the weight of the far target node decreases to 0.4, and the position error after fusion is ≤5 cm. The algorithm real-time tracks the target motion pattern, uses linear prediction for uniform motion, and introduces second-order compensation for accelerated motion to improve robustness.

[0017] In a preferred embodiment, in step S5, a correction value is generated every two sets of estimated values, and a mean square estimation algorithm is used to optimize the target three-dimensional coordinates (X_0, Y_0, D_0). The height error is ≤3 cm, the weight of the node within 5 meters of the target is 0.6, and the weight of the node >5 meters is 0.4. The system combines historical trajectories to establish a motion model, predicts the next time position, and eliminates random fluctuations through a sliding window filter to ensure continuity of the output.

[0018] In a preferred embodiment, in step S6, the positions of non-visual nodes are inversely calculated using the triangulation method through at least 3 known nodes in the visual matrix. The node spacing measurement error is ≤1 cm, and the mapping result is updated 10 times per second. The non-visual node position data is broadcast through the cluster network with a broadcast interval of ≤100 milliseconds. If there are less than 3 participating nodes, the system immediately triggers an alarm and starts the redundant node supplement mechanism.

[0019] In a preferred embodiment, in step S7, the activity (such as movement speed, frequency) of the target in the local area and the computing load of each device are first continuously monitored by the cluster sensor. When the target enters a high-activity area (for example, a frequently moving or densely interacting area), the system automatically dispatches adjacent visual devices to join the positioning matrix, improving the detection accuracy in this area. Conversely, if the target is in a low-activity area, the number of visual devices is reduced, and the data collection task is switched to adjacent nodes to avoid resource redundancy. This adjustment process dynamically predicts the optimal strategy through a reinforcement learning model: the model takes historical positioning errors, bandwidth occupancy, and device load as input, trains to generate a decision to "dispatch / reduce visual nodes", and ultimately balances bandwidth occupancy minimization and positioning accuracy.

[0020] The reinforcement learning-driven adaptive mechanism is the innovation support of step S7. The model operates through the following closed loop: State space definition: including target position distribution, device computing load, network bandwidth occupancy; Action space definition: dynamically dispatching or reducing visual devices, and specifying data source switching path; Reward function design: optimizing positioning error reduction and bandwidth occupancy reduction.

[0021] The model updates the strategy in real time through online learning, for example, when a sudden increase in target activity in a certain area is detected, the visual device dispatch action is triggered immediately; if the positioning error is still excessive after load balancing, the strategy is further iterated. This mechanism ensures that the system always responds efficiently in a dynamic environment.

[0022] The regional activity quantification model formula that converts the dynamic adjustment problem into a quantifiable optimization target and deeply combines with reinforcement learning is: ; Where: Vt represents the target movement speed at time t; ΔP / Δt is the target position change rate per unit time; α,β are weight coefficients calibrated by historical data.

[0023] Association with the process: the formula output Ar directly determines the dispatch or reduction of visual devices (Ar exceeding the threshold triggers the action), and is the core input of the reinforcement learning state space.

[0024] The reinforcement learning reward function formula is: ; E is the current positioning error, and Emax is the maximum allowed error; B is the bandwidth occupancy, and Bmax is the upper limit of bandwidth; γ1, γ2 are balance coefficients, solved by constraint optimization.

[0025] Associated with the flow: the reward R real-time evaluates the effectiveness of the action, and feeds back to the policy network, forming a closed-loop optimization.

[0026] In a preferred embodiment, in step S8, the visual node of the main link directly returns data through the TCP protocol, and the check mechanism adopts CRC-32 and sequence number confirmation to ensure integrity; The non-visual node of the backup link is transferred through the adjacent visual node and uses low-delay UDP protocol for transmission.

[0027] The dual-link data consistency is verified by SHA-256 hash value comparison, and the backup link is switched within 50 milliseconds when the main link fails, and there is no data loss in the switching process.

[0028] In a preferred embodiment, in step S9, the positioning result is input into the PID guidance controller, the proportional coefficient Kp=0.8, the integral coefficient Ki=0.2, and the differential coefficient Kd=0.1, and the parameters are dynamically adjusted according to the target maneuverability. The system optimizes the visual matrix configuration every minute: based on the recent 100 times of historical positioning error and real-time bandwidth occupancy, the node selection strategy is updated by gradient descent method, and the optimization result is synchronized in real time to the dynamic device reconfiguration process, forming a closed-loop self-evolution.

[0029] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present application are: 1. In the present application, a small number of devices are dynamically selected to construct a visual positioning matrix, and only 4 devices are needed to cover a local area, and the remaining devices do not need visual modules. This design reduces video transmission demand by more than 70%, and the released bandwidth resources enable the cluster size to be expanded to more than 100 devices. At the same time, the dynamic device reconfiguration mechanism adjusts the number of visual nodes in real time according to the target activity, further avoids resource waste under the premise of ensuring accuracy, and realizes the decoupling of bandwidth occupancy and cluster size.

[0030] 2. In the present application, the redundant transmission mechanism verifies the data in parallel through dual-link: the main link ensures that the control command is completely delivered, and the backup link transmits data through adjacent nodes to ensure low-delay return. Even if a single node fails or the network fluctuates, the system can seamlessly switch links within 50 milliseconds to maintain the command reachability. In addition, multi-device cooperative positioning combines centimeter-level precision measurement and motion trajectory prediction to stabilize the positioning error within 5 centimeters. This dual-insurance design of "hardware redundancy + algorithm fault tolerance" enables the system to still output high-precision positioning results in complex environments, meeting the requirements of industrial-level reliability. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a schematic diagram of the flow principle of the present application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Example: Reference Figure 1 A cluster vision method based on SPV and redundant transmission, comprising the following steps: S1: Define the cluster size and local operational scope, identify candidate nodes where vision devices need to be installed, and ensure coverage of key areas. This step provides a spatial constraint basis for subsequent vision matrix construction.

[0034] S2: Based on the overlap of local areas, dynamically select the optimal m devices (m≥4) to construct a visual positioning matrix. The positions and coordinate systems of the devices within the matrix are synchronized in real time, providing a basic framework for multi-view collaborative positioning.

[0035] S3: All devices within the matrix synchronously detect the target and calculate the target's local coordinates (x_i, y_i) and yaw angle based on their respective coordinate systems. The detection results from each device are shared in real time through cluster communication, providing data input for global positioning.

[0036] S4: Utilize shared data to generate multiple distance estimates for the same target. Integrate the estimates using a weighted fusion algorithm (such as Kalman filtering) to initially eliminate single-device errors and improve positioning robustness.

[0037] S5: Input the fused predicted values ​​into the redundancy correction model: Generate a correction value for every two sets of predicted values, and optimize the target's 3D coordinates (X_0, Y_0, D_0) through mean square estimation. This step relies on the fusion result of step S4 to ensure positioning accuracy at the centimeter level. S6: Map the precise coordinates generated in step 5 to nodes without installed vision devices (e.g., F0). Using the known positional relationships of devices within the matrix, the relative positions of non-vision nodes are calculated backwards to achieve global coverage under SPV.

[0038] S7: Based on the target motion trajectory and cluster load status, adjust the role of the vision device in real time, specifically including: Dynamically add visual nodes to highly active areas; Reduce visual nodes in low-activity areas and switch data sources to nearby devices; By using reinforcement learning models to predict and adjust strategies, bandwidth consumption can be minimized while maintaining positioning accuracy.

[0039] S8: Design a dual-redundant communication link, wherein: Main link: control results are directly returned by vision nodes; Backup link: non-vision nodes transfer data through adjacent vision nodes; The two links verify data consistency in parallel, ensuring real-time feedback and fault tolerance.

[0040] S9: input the positioning result into the guidance control system to generate operation instructions. Meanwhile, based on historical positioning errors and bandwidth occupation data, the configuration strategy of the vision matrix is iteratively optimized and fed back to step S7), forming an adaptive closed loop.

[0041] In step S1, cluster initialization requires accurate demarcation of the local operation range, with the length parameters Ld and Lh of the rectangular region set between 10 meters and 100 meters, and the specific values adjusted dynamically according to the actual scene. The key area coverage requires no blind spots, and the candidate vision node positions are pre-calibrated by RTK-GPS or laser ranging technology, with positioning accuracy strictly controlled at the centimeter level. The communication network uses 5G or a dedicated wireless protocol, with a transmission delay upper limit of 20 milliseconds to ensure real-time collaboration. The node deployment density needs to meet the minimum spacing of 2 meters to avoid signal interference, while redundant nodes are reserved to handle sudden failures.

[0042] In step S2, the vision positioning matrix is dynamically composed of at least 4 nodes, and the system analyzes the node position overlap in real time to automatically select the optimal nodes to build the matrix. The position synchronization error needs to be ≤1 centimeter, and the matrix update frequency is 1 time per second to adapt to fast target movement. The main node is responsible for data coordination, and the backup node seamlessly switches within 50 milliseconds in case of failure. The lower limit of the node spacing is 2 meters, and the matrix construction algorithm preferentially selects nodes with high field of view overlap rate and load balancing to ensure global coverage stability.

[0043] In step S3, all nodes in the matrix capture the target at a frequency of 30 frames per second, and a lightweight YOLOv5 model is used for efficient detection. Each node calculates the target coordinates (xi, yi) and three-dimensional deviation angle based on the local coordinate system, with a detection confidence threshold of 0.8, and data below the threshold is automatically discarded. Target data is shared in real time through time-stamped UDP protocol, with a synchronization error ≤10 milliseconds to ensure the spatiotemporal consistency of multi-source data.

[0044] In step S4, the adaptive Kalman filter algorithm is used in the fusion stage, with the initial value of process noise covariance QQ set to 0.1 and the measurement noise covariance RR set to 0.5. The parameters are dynamically adjusted according to the target motion state. The weight of the near target node is set to 0.6, and the weight of the far target node decreases to 0.4, with a position error ≤5 centimeters after fusion. The algorithm real-time tracks the target motion pattern, uses linear prediction for uniform motion, and introduces second-order compensation for accelerated motion to improve robustness.

[0045] 6. The SPV and redundant transmission-based cluster vision method of claim 1, wherein in step S5, a correction value is generated once every two sets of estimated values, and a mean square estimation algorithm is used to optimize the target three-dimensional coordinates (X_0, Y_0, D_0). The height error is less than or equal to 3 cm, the weight of a node within 5 m of the target is 0.6, and the weight of a node more than 5 m away from the target is 0.4. The system combines historical trajectories to establish a motion model, predicts the position at the next moment, and eliminates random fluctuations through a sliding window filter to ensure continuity of the output.

[0046] 7. The SPV and redundant transmission-based cluster vision method of claim 1, wherein in step S6, the positions of non-vision nodes are inversely calculated by at least three known nodes in the vision matrix using a triangulation method. The measurement error of the distance between nodes is less than or equal to 1 cm, and the mapping result is updated 10 times per second. The position data of the non-vision nodes is broadcasted through the cluster network with a broadcast interval of less than or equal to 100 ms. If there are less than three participating nodes, the system immediately triggers an alarm and starts a redundant node supplement mechanism.

[0047] In step S7, the activity of the target in the local area (such as movement speed and frequency) and the computing load of each device are first continuously monitored through the cluster sensor. When the target enters a high-activity area (such as a frequently moving or densely interacting area), the system automatically dispatches nearby vision devices to join the positioning matrix, improving the detection accuracy in this area. Conversely, if the target is in a low-activity area, the number of vision devices is reduced, and the data collection task is switched to adjacent nodes to avoid resource redundancy. This adjustment process dynamically predicts the optimal strategy through a reinforcement learning model: the model takes historical positioning errors, bandwidth occupancy, and device load as inputs, trains to generate a decision to "dispatch / reduce vision nodes", and finally balances the minimum bandwidth occupancy and positioning accuracy.

[0048] The reinforcement learning-driven adaptive mechanism is the innovation support of step S7. The model operates through the following closed loop: State space definition: including target position distribution, device computing load, and network bandwidth occupancy; Action space definition: dynamically dispatching or reducing vision devices, and specifying data source switching paths; Reward function design: optimizing the reduction of positioning error and bandwidth occupancy.

[0049] The model updates the strategy in real time through online learning, for example, when a sudden increase in target activity is detected in a certain area, the vision device dispatch action is triggered immediately; if the positioning error is still excessive after load balancing, the strategy is further iterated. This mechanism ensures that the system always responds efficiently in a dynamic environment.

[0050] The regional activity quantification model formula that converts the dynamic adjustment problem into a quantifiable optimization target and deeply combines reinforcement learning is: ; Wherein: Vt represents the target motion speed at time t; ΔP / Δt is the target position change rate per unit time; α,β are weight coefficients calibrated by historical data.

[0051] Process association: The formula output Ar directly determines the increase or decrease of visual devices (Ar exceeds the threshold value to trigger action), which is the core input of the reinforcement learning state space.

[0052] The reinforcement learning reward function formula is: ; E is the current positioning error, and Emax is the maximum allowable error; B is the bandwidth occupancy, and Bmax is the upper limit of the bandwidth; γ1,γ2 are balance coefficients solved by constraint optimization.

[0053] Process association: The reward R evaluates the effectiveness of the action in real time and feeds back to the policy network to form a closed-loop optimization.

[0054] In step S8, the visual node of the main link directly returns data through the TCP protocol, and the CRC-32 and sequence number confirmation are used for verification mechanism to ensure integrity; The non-visual node of the backup link transmits through the adjacent visual node and uses low-delay UDP protocol for transmission.

[0055] The dual-link data consistency is verified by SHA-256 hash value comparison, and the backup link is switched within 50 milliseconds when the main link fails, and there is no data loss in the switching process.

[0056] In step S9, the positioning result is input into the PID guidance controller, the proportional coefficient Kp=0.8, the integral coefficient Ki=0.2, and the differential coefficient Kd=0.1, and the parameters are dynamically adjusted according to the target maneuverability. The system optimizes the visual matrix configuration every minute: based on the recent 100 times of historical positioning error and real-time bandwidth occupancy, the node selection policy is updated by gradient descent method, and the optimization result is synchronized to the dynamic device reconfiguration process in real time, forming a closed-loop self-evolution.

[0057] From the above, it can be seen that: In the application, a visual positioning matrix is constructed by dynamically selecting a small number of devices, and only 4 devices are needed to cover a local area, and the remaining devices do not need a visual module. This design reduces video transmission demand by more than 70%, and the released bandwidth resources enable the cluster size to be expanded to more than 100. At the same time, the dynamic device reconfiguration mechanism adjusts the number of visual nodes in real time according to the target activity, further avoids resource waste on the premise of ensuring accuracy, and realizes the decoupling of bandwidth occupation and cluster size.

[0058] In the application, the redundant transmission mechanism checks data in parallel through double links: the main link ensures that the control instruction is delivered completely, and the standby link ensures low-delay return through adjacent node transfer. Even if a single node fails or the network fluctuates, the system can seamlessly switch links within 50 milliseconds, maintaining the instruction reachability. In addition, multi-device cooperative positioning combines centimeter-level precision measurement and motion trajectory prediction to stabilize the positioning error within 5 centimeters. This double insurance design of 'hardware redundancy + algorithm fault tolerance' enables the system to still output high-precision positioning results in complex environments, meeting the requirements of industrial-level reliability.

[0059] It should be noted that, in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0060] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A cluster vision method based on SPV and redundant transmission, characterized in that: The method includes the following steps: S1: Define the cluster size and local operation range, and determine the candidate nodes where vision devices need to be installed to ensure coverage of key areas; this step provides a spatial constraint basis for the subsequent construction of the vision matrix. S2: Based on the overlap of local areas, the optimal m devices are dynamically selected to construct a visual positioning matrix; the positions and coordinate systems of the devices within the matrix are synchronized in real time, providing a basic framework for multi-view collaborative positioning; S3: All devices within the matrix synchronously detect the target and calculate the target's local coordinates (x_i, y_i) and deviation angle based on their respective coordinate systems; the detection results of each device are shared in real time through cluster communication, providing data input for global positioning; S4: Utilize shared data to generate multiple distance estimates for the same target; integrate the estimates through a weighted fusion algorithm to initially eliminate single-device errors and improve positioning robustness; S5: Input the fused prediction values ​​into the redundancy correction model: Generate a correction value for every two sets of prediction values, and optimize the three-dimensional coordinates (X_0,Y_0,D_0) of the target through mean square estimation; This step depends on the fusion result of step S4 to ensure that the positioning accuracy reaches the centimeter level; S6: Map the precise coordinates generated in step S5 to nodes where no vision devices are installed; calculate the relative positions of non-vision nodes in reverse by using the known positional relationships of devices within the matrix, and achieve global coverage under SPV; S7: Based on the target motion trajectory and cluster load status, adjust the role of the vision device in real time, specifically including: Dynamically add visual nodes to highly active areas; Reduce visual nodes in low-activity areas and switch data sources to nearby devices; By using reinforcement learning models to predict and adjust strategies, bandwidth consumption can be minimized while maintaining positioning accuracy. S8: Design a dual-redundant communication link, wherein: Main link: Control results are directly transmitted back from the visual nodes; Backup link: Non-visual nodes relay data through nearby visual nodes; Two links are used to verify data consistency in parallel to ensure the real-time nature and fault tolerance of the feedback. S9: Input the positioning results into the guidance and control system to generate operation commands; at the same time, based on historical positioning errors and bandwidth usage data, iteratively optimize the configuration strategy of the vision matrix and feed it back to step S7 to form an adaptive closed loop.

2. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S1, cluster initialization requires precise delineation of the local operating range. The side length parameters Ld and Lh of the rectangular area are set between 10 meters and 100 meters, and the specific values ​​are dynamically adjusted according to the actual scenario. The key area coverage requires no blind spots. The positions of candidate visual nodes are pre-calibrated using RTK-GPS or laser ranging technology, and the positioning accuracy is strictly controlled at the centimeter level.

3. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S2, the visual positioning matrix is ​​dynamically composed of at least 4 nodes. The system analyzes the overlap of node positions in real time and automatically selects the optimal nodes to construct the matrix. The location synchronization error must be ≤1 cm, the matrix update frequency is once per second to adapt to the rapid movement of the target; the master node is responsible for data coordination, and the backup node can seamlessly switch within 50 milliseconds in case of failure; the minimum node spacing is 2 meters, and the matrix construction algorithm prioritizes nodes with high field-of-view overlap and balanced load to ensure global coverage stability.

4. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S3, all nodes in the matrix synchronously capture the target at a frequency of 30 frames per second, and a lightweight YOLOv5 model is used to achieve efficient detection. Each node calculates the target coordinates (xi,yi) and three-dimensional bias angle based on the local coordinate system. The detection confidence threshold is set to 0.8, and data below the threshold is automatically discarded.

5. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S4, the fusion stage uses an adaptive Kalman filter algorithm, with the initial value of the process noise covariance QQ being 0.1 and the measurement noise covariance RR being 0.

5. The parameters are dynamically adjusted according to the target motion state. The weight of the nearest target node is set to 0.6, and the weight of the far target node is reduced to 0.

4.

6. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S5, a correction value is generated for every two sets of estimated values, and the mean square estimation algorithm is used to optimize the target's three-dimensional coordinates (X_0, Y_0, D_0); the weight of nodes with a height error ≤3 cm and a distance ≤5 m from the target is 0.6, and the weight of nodes >5 m is 0.

4.

7. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S6, the positions of non-visual nodes are inferred by using triangulation based on at least three known nodes in the visual matrix; the node spacing measurement error is ≤1 cm, and the mapping result is updated 10 times per second.

8. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S7, the activity level of the target in the local area and the computing load of each device are continuously monitored by the cluster sensors. When the target enters a highly active area, the system automatically adds nearby vision devices to the positioning matrix to improve the detection accuracy of the area. Conversely, if the target is in a low-activity area, reduce the number of vision devices and switch the data acquisition task to a nearby node to avoid resource redundancy. This adjustment process dynamically predicts the optimal strategy through a reinforcement learning model: the model is trained to generate decisions on "adding / reducing visual nodes" by taking historical positioning errors, bandwidth utilization, and device load as inputs, ultimately achieving a balance between minimizing bandwidth utilization and positioning accuracy. The dynamic adjustment problem is transformed into a quantifiable optimization objective, and the formula for the region activity quantification model, which is deeply integrated with reinforcement learning, is as follows: ； in: Vt represents the target velocity at time t; ΔP / Δt is the rate of change of the target position per unit time; α and β are weighting coefficients, calibrated using historical data; Related to the process: The formula output Ar directly determines the addition or reduction of vision devices and is the core input to the reinforcement learning state space; The formula for the reinforcement learning reward function is: ； E represents the current positioning error. max The maximum allowable error; B represents bandwidth utilization. max This is the maximum bandwidth limit; γ1 and γ2 are equilibrium coefficients, which are solved through constraint optimization. Related to the process: Reward R evaluates the effectiveness of actions in real time and feeds it back to the policy network to form a closed-loop optimization.

9. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S8, the visual node of the main link directly transmits data back via TCP protocol, and the verification mechanism uses CRC-32 and sequence number confirmation to ensure integrity. Non-visual nodes on the backup link relay through a nearby visual node and use low-latency UDP protocol for transmission; Dual-link data consistency is verified by SHA-256 hash value comparison. When the primary link fails, it switches to the backup link within 50 milliseconds, and there is no data loss during the switching process.

10. The cluster vision method based on SPV and redundant transmission as described in claim 1, characterized in that: In step S9, the positioning result is input into the PID guidance controller, with proportional coefficient Kp=0.8, integral coefficient Ki=0.2, and derivative coefficient Kd=0.

1. The parameters are dynamically adjusted according to the target's mobility. The system optimizes the vision matrix configuration once per minute: based on the most recent 100 historical positioning errors and real-time bandwidth occupancy, the node selection strategy is updated using the gradient descent method. The optimization result is synchronized to the dynamic device reconfiguration process in real time, forming a closed-loop self-evolution.

Images