Multi-unmanned aerial vehicle cooperative identity recognition system based on feedback driving

Abstract

The invention is suitable for the technical field of unmanned aerial vehicle communication, and particularly relates to a multi-unmanned aerial vehicle cooperative identity recognition system based on feedback driving, and the system comprises an unmanned aerial vehicle cluster which is used for carrying out data collection and carrying out the semantic extraction of collected video data, and obtaining semantic features; the central control terminal is used for acquiring global state information and semantic features from the unmanned aerial vehicle and generating an action decision; and the feedback scheduling control module is used for performing path planning, error control and communication link modeling according to the action decision, and generating a feedback control instruction. According to the invention, through edge intelligent sensing, center feedback scheduling, dynamic path planning and communication resource optimization, a highly-integrated multi-unmanned aerial vehicle cooperative identity recognition system is formed, the recognition accuracy and the response speed are significantly improved, energy and communication resource utilization is optimized, and the system is suitable for large-scale popularization and application. And stable operation and efficient task completion in a complex and changeable environment are ensured.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) communication technology, and particularly relates to a feedback-driven multi-UAV collaborative identity recognition system. Background Technology

[0002] Current multi-UAV collaborative identification systems generally employ static task allocation and fixed communication strategies. Static task allocation enables basic image acquisition and centralized processing, providing some path planning and automated identification capabilities. The system monitors target areas through preset flight paths or coverage searches, and combined with centralized analysis from a central terminal, it can achieve basic functions of environmental perception and identification for multiple UAVs. However, static task allocation struggles to adapt to dynamic and complex environments, exhibiting lag in response to rapidly moving or sudden targets and lacking flexible adjustment mechanisms, thus impacting identification efficiency and accuracy. Furthermore, single-visual-modal recognition is limited by image quality, occlusion, and changes in lighting, potentially leading to insufficient stability and robustness, making it difficult to meet the high-precision requirements of complex scenarios. Summary of the Invention

[0003] The purpose of this invention is to provide a feedback-driven multi-UAV collaborative identity recognition system, which aims to solve the problems of existing multi-UAV identity recognition systems, such as the single task execution mode and lack of dynamic scheduling capability. They usually rely on preset routes or fixed area scanning, making it difficult to adjust the acquisition strategy in a timely manner based on the recognition results, resulting in insufficient coverage of key areas and target omissions.

[0004] This invention is implemented as follows: a feedback-driven multi-UAV cooperative identity recognition system, the system comprising: A drone swarm, which is used to collect data and extract semantic features from the collected video data; A central control terminal is used to acquire global state information and semantic features from the UAV and generate action decisions. The feedback scheduling control module is used to perform path planning, error control, and communication link modeling based on action decisions, and to generate feedback control commands. It also includes an edge recognition terminal and a distributed online learning module for continuous optimization. The edge recognition terminal is used to incrementally train the data collected by the drone swarm and update the local model parameters.

[0005] Preferably, in the process of collecting data and extracting semantic features from the collected video data, the drone swarm... In time The captured video frames are The edge recognition model parameters are The features extracted from the drone are represented as follows: ; in Including facial information, clothing texture, motion trajectory, spatial location, and RF signal features, the feature vector is... Input lightweight classification network Perform identity recognition and calculate the recognition confidence level using a soft maximum function. : .

[0006] Preferably, the edge feature extraction network incorporates a channel attention mechanism. Key features are weighted, and edge deployment efficiency is optimized by combining pruning and quantization-aware training (QAT).

[0007] in It is the feature vector after weighted optimization; The drone swarm includes a language generation module and cross-modal features. Then input into the semantic generation module Generate text description :

[0008] Among them, language model parameters We employ a deep pre-trained BERT architecture and fine-tune it to adapt to the visual semantic translation task. Features generated by edge feature extraction network Confidence level and text description The data is transmitted to the central control terminal via a wireless communication network, and the current global status information of each UAV is uploaded simultaneously.

[0009] Preferably, the global state information is the global state vector of the UAV. :

[0010] in, For spatial location information, For speed, Represents remaining energy. This indicates the current status of the communication link.

[0011] Preferably, the central control terminal employs a deep reinforcement learning policy network, and the parameters of the deep reinforcement learning policy network are set as follows: Input status With semantic information Output action decision :

[0012] The objective function for this deep reinforcement learning policy network is:

[0013] in, Corresponding task completion time, As an energy consumption indicator, , , These are the weighting coefficients for each indicator.

[0014] Preferably, the feedback scheduling control module includes a path control submodule, an error control submodule, and a communication link modeling submodule. The path control submodule integrates an artificial intelligence path search algorithm and a dynamic window method to generate a dynamic obstacle avoidance path based on the current flight status and target position. The feedback scheduling control module will take action Converted into drone control commands :

[0015] in This is the feedback controller function.

[0016] Preferably, the error control submodule is adjusted in real time by a PID control law, and the error calculation is expressed as follows:

[0017] Where t represents the discrete time step number of the current control cycle. The tracking error at the current moment is defined as the target position. Compared with the actual location of the drone The difference, , , These are the proportional, integral, and differential gains, respectively.

[0018] Preferably, the communication link modeling submodule defines the signal-to-noise ratio based on the wireless channel characteristics between UAVs. Characterizing the first frame and the first Link quality between drones; bandwidth According to Shannon's formula, the result is:

[0019] Corresponding transmission delay The calculation is as follows:

[0020] in, The allocation of communication resources is dynamically adjusted according to the size of the data to be transmitted, so as to avoid link congestion.

[0021] Preferably, the edge recognition terminal performs incremental training on the collected high-value samples to update the local model parameters:

[0022] in For learning rate, The loss function is trained at the edge. The policy network π is fine-tuned online based on task performance:

[0023] Where γ is the policy network learning rate. To reinforce the learning objective function.

[0024] Preferably, the training and deployment of the system are divided into three stages, including: In the offline phase, the recognition and language generation models are trained based on publicly available multimodal datasets, and the parameters are compressed and optimized to adapt to the operating requirements of embedded devices. During the simulation phase, a multi-objective dynamic scenario is constructed based on the simulation platform, and a scheduling policy network is trained to optimize the overall system performance. During the online deployment phase, model parameters are distributed to the drone cluster and central terminal, enabling the system to achieve adaptive closed-loop operation with simultaneous data collection and feedback in the actual environment.

[0025] The feedback-driven multi-UAV collaborative identity recognition system provided by this invention forms a highly integrated multi-UAV collaborative identity recognition system through edge intelligent perception, central feedback scheduling, dynamic path planning and communication resource optimization. It significantly improves the recognition accuracy and response speed, optimizes the utilization of energy and communication resources, and ensures stable operation and efficient task completion in complex and ever-changing environments. Attached Figure Description

[0026] Figure 1 This is an architecture diagram of a feedback-driven multi-UAV collaborative identity recognition system provided in an embodiment of the present invention. Detailed Implementation

[0027] 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.

[0028] Multi-UAV collaborative perception refers to the process by which multiple unmanned aerial vehicles (UAVs) with autonomous perception and communication capabilities collaborate dynamically through wireless communication networks, thereby enhancing the overall system's environmental understanding and task completion efficiency. In this invention, multiple UAVs dynamically adjust their identification tasks and flight trajectories through a feedback mechanism from a central control terminal (i.e., a central control node), achieving a highly efficient aerial collaborative perception system that significantly improves identification efficiency and system stability.

[0029] Edge computing technology refers to performing local data processing and preliminary analysis on devices close to the data acquisition end (such as cameras, sensors, and drones), reducing data transmission latency, bandwidth burden, and alleviating the computational pressure on the central terminal. Traditional image recognition often relies on centralized processing at a central terminal, resulting in response latency and bandwidth bottlenecks. In this invention, the edge drone possesses preliminary image recognition and feature extraction capabilities, enabling real-time target detection and confidence assessment, and uploading high-value information to the central terminal, effectively supporting global recognition decisions.

[0030] The feedback-driven mechanism is one of the core innovations of this system. The central control terminal constructs a feedback strategy based on the fusion evaluation results of multi-source data (such as data reliability scores, feature information missing indicators, etc.), and then generates dynamic instructions to guide the UAV to perform supplementary shooting, re-identification, or patrol of specific areas. Finally, it realizes closed-loop control of data collection, analysis and task adjustment, further ensuring the pertinence and efficiency of the identification task.

[0031] The identity recognition algorithm is based on deep learning and combines multi-dimensional information such as facial features, clothing attributes, and behavioral trajectories to support cross-view and cross-time identity verification and abnormal behavior analysis. Multi-view data collected collaboratively by multiple drones is fused through a central fusion algorithm to improve overall recognition accuracy and robustness.

[0032] By integrating UAV self-organizing collaboration, edge intelligent processing, central feedback control, and intelligent recognition algorithms, this proposal can achieve efficient perception, dynamic positioning, and intelligent recognition of target personnel in complex situations such as unregistered or missing identity information. It is particularly suitable for various mission scenarios such as security patrol, military identity reconnaissance, and disaster relief in complex environments.

[0033] like Figure 1 The diagram shown is an architecture diagram of a feedback-driven multi-UAV cooperative identity recognition system provided in an embodiment of the present invention. The system includes: A drone swarm is used to collect data and extract semantic features from the collected video data.

[0034] In this system, the UAV swarm serves as the data acquisition front-end, equipped with a high-performance embedded recognition chip and a high-speed communication module, enabling target detection and semantic feature extraction at the edge recognition end. Let the UAVs... At time step The captured video frames are The edge recognition model parameters are The features extracted by the drone are represented as

[0035] in It includes facial information, clothing texture, motion trajectory, spatial location, and RF signal features. Subsequently, the feature vector... Input lightweight classification network Perform identity recognition and calculate the recognition confidence level using a soft maximum function. :

[0036] Here, confidence level The reliability assessment of the current target identification result by UAV i directly affects subsequent closed-loop scheduling decisions. When the assessment is insufficient to support reliable decisions, it will trigger subsequent scheduling and re-identification. To enhance robustness in complex environments, the edge feature extraction network introduces a channel attention mechanism. Key features are weighted, and edge deployment efficiency is optimized by combining pruning and quantization-aware training (QAT).

[0037] in It is a weighted and optimized feature vector, which not only improves the effectiveness of feature representation but also provides reliable input for semantic-level fusion. Cross-modal features Then input into the semantic generation module Generate text description :

[0038] This module fine-tunes the visual-semantic task through a deep pre-trained BERT architecture, ensuring that the semantic description is consistent with the environmental context, and providing constraints and decision-making basis for subsequent feedback scheduling.

[0039] Features generated by edge recognition end Confidence level Semantic description and UAV state vector

[0040] Uploaded to the central control terminal, where , Indicates location, Indicates speed, Indicates the remaining energy. Indicates the status of the communication link.

[0041] The central control terminal is used to acquire global status information and semantic features from the UAV and generate action decisions.

[0042] The feedback scheduling control module is used to perform path planning, error control, and communication link modeling based on action decisions, and to generate feedback control commands. The system includes an edge recognition terminal and a distributed online learning module. The edge recognition terminal is used to incrementally train the data collected by the UAV swarm and update the local model parameters. The central control terminal is used to aggregate the updates from the UAV swarm, realize global model synchronization, and distribute them to the edge training units. Through this edge-central collaborative online learning, the UAV can continuously optimize the recognition model and scheduling strategy in a dynamic environment, thereby enhancing the system's adaptability, robustness, and long-term performance.

[0043] In this system, the central control terminal utilizes a deep reinforcement learning policy network based on the global state and cross-modal features of multiple UAVs. Output scheduling action :

[0044] The strategy optimization objective function comprehensively considers identification confidence and task completion time. Energy consumption :

[0045] in , , These are the weighting coefficients for confidence, time, and energy consumption, respectively, reflecting the core role of recognition accuracy in task optimization.

[0046] The feedback scheduling control module will take action Converted into drone control commands :

[0047] in This is a feedback controller function used to dynamically adjust path selection, communication resource allocation, and task priority. The UAV path planning combines a dynamic window method with heuristic search, and corrects flight errors using PID control.

[0048] This enables precise flight control and mission execution. Communication link performance is based on the signal-to-noise ratio between UAVs. Bandwidth calculation with Shannon formula With delay :

[0049] When a decrease in recognition confidence or semantic loss is detected, the closed-loop mechanism automatically triggers supplementary sampling or path replanning to ensure that the recognition reliability is not reduced due to environmental changes.

[0050] The distributed online learning mechanism is tightly coupled with edge feature extraction and the central feedback loop in the system. The edge recognition end performs incremental training on the collected high-value samples to update the local model parameters.

[0051] in For learning rate, The loss function is used for edge training. The central control terminal aggregates updates from multiple drones, achieving global model synchronization and distributing them to the edge recognition terminals, thus completing edge-center collaborative optimization. Simultaneously, the policy network π is fine-tuned online based on task performance.

[0052] Where γ is the policy network learning rate. To reinforce the learning objective function, this adaptive optimization not only improves the recognition model but also continuously optimizes the scheduling strategy, thereby feeding back into the feedback loop and cross-modal fusion decision-making, enabling the system to form a continuously iterative adaptive closed loop.

[0053] This invention proposes a feedback-driven multi-UAV collaborative identity recognition system, combining UAV communication, edge computing, and intelligent scheduling strategies. The system uses a central control terminal to perform reliability assessments and detect missing important information on data uploaded from the edge recognition terminals, establishing a dynamic feedback mechanism to guide UAVs to collect and transmit data more effectively. This enables high-precision detection and real-time response to unidentified individuals, effectively improving recognition accuracy and overall system performance. Specifically, it includes the following key technical points: 1. Adaptive Link Based on Identification Confidence and Semantic Missing The feedback mechanism proposed in this invention uses the confidence level and semantic integrity of the identity recognition result as triggering conditions. When the recognition confidence level falls below a threshold or semantic missingness is detected, the central terminal immediately generates a scheduling command to drive the UAV to perform supplementary data collection, path replanning, or task priority adjustment, and re-initiates the recognition process. This forms a dynamic closed loop with recognition reliability as the core, no longer relying on traditional geometric deviation or coverage indicators, but directly embedding recognition quality into the task chain.

[0054] 2. Semantic-level multimodal fusion method After extracting multi-source features such as images, faces, clothing attributes, and behavioral trajectories from the drone's edge, semantic modeling is performed to generate an identity description before being uploaded to the central control terminal. This semantic-level fusion not only reduces bandwidth overhead and enhances feature robustness but also provides semantic support for subsequent confidence-triggered mechanisms. Compared to existing low-level feature concatenation, this semantically driven fusion approach is more suitable for maintaining stable recognition performance in complex environments.

[0055] 3. Edge-Center Dual-Layer Collaborative Recognition Architecture The system adopts a two-layer architecture of "lightweight edge recognition + central global fusion". The edge recognition end quickly completes preliminary detection and semantic feature extraction, ensuring real-time performance; the central control terminal performs cross-UAV feature alignment and identity fusion, improving consistency across multiple perspectives. This two-layer collaboration not only solves the bottleneck of centralized processing, but also works in conjunction with the confidence triggering mechanism to ensure the continuous operation of the recognition link in dynamic environments.

[0056] 4. Distributed online learning and continuous optimization mechanism During operation, the edge recognition terminal performs incremental training and selects high-value samples to upload to the central control terminal, which then performs global synchronous updates. This distributed online learning enables the recognition model and scheduling strategy to dynamically adapt to the environment and tasks, continuously feeding back into the aforementioned feature fusion and scheduling chain, thereby improving the overall robustness and evolutionary capability of the system.

[0057] 5. Scheduling optimization objective with recognition accuracy as the core The system's proposed scheduling objective function prioritizes the reliable completion of identity recognition, comprehensively considering confidence level, task latency, and energy consumption distribution. Unlike traditional strategies oriented towards navigation efficiency, this objective function unifies the optimization of scheduling, recognition, and learning around the core indicator of "recognition accuracy," thereby achieving true dynamic adaptive scheduling and resource optimization.

[0058] Compared to traditional solutions, this invention has significant technical advantages. It proposes a feedback-driven multi-UAV collaborative identity recognition system, whose core innovations include cross-modal semantic fusion, confidence-driven feedback closed loop, and distributed online learning, emphasizing the close dependency and closed-loop logic among these three elements. 1. Cross-modal semantic fusion provides a reliable foundation for decision-making information. By performing unified semantic modeling on multimodal features such as faces, clothing, motion trajectories, and RF signals, the system can still accurately identify targets even in situations where information is scarce or the environment is complex.

[0059] 2. The confidence-driven feedback loop uses the fusion results and semantic integrity as triggering conditions to guide the UAV to dynamically adjust its acquisition strategy, flight path and task priority, thereby realizing a real-time closed loop of identification-scheduling-control.

[0060] 3. Distributed online learning continuously transforms newly collected data into model optimization through incremental training at the edge recognition end and global synchronization at the central control terminal. This enables the system to continuously adapt and adjust its strategies during task execution, and to feed back feedback loops and cross-modal fusion decision-making.

[0061] The three elements work closely together in the system: cross-modal fusion provides the basis for closed-loop decision-making, the feedback loop transforms the recognition results into task scheduling instructions, and the online learning mechanism optimizes model parameters and strategies, enabling the closed loop to continuously adapt and iterate, thereby maintaining high accuracy, high efficiency, and high robustness in complex dynamic environments. Through this inherently coupled closed-loop design, this proposal overcomes the limitations of existing technologies such as single-modal recognition, static scheduling, and fixed communication strategies, achieving real-time optimization of UAV collaborative perception.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feedback-driven multi-UAV cooperative identity recognition system, characterized in that, The system includes: A drone swarm, which is used to collect data and extract semantic features from the collected video data; A central control terminal is used to acquire global state information and semantic features from the UAV and generate action decisions. The feedback scheduling control module is used to perform path planning, error control, and communication link modeling based on action decisions, and to generate feedback control commands. It also includes an edge recognition terminal and a distributed online learning module for continuous optimization. The edge recognition terminal is used to incrementally train the data collected by the drone swarm and update the local model parameters.

2. The feedback-driven multi-UAV cooperative identity recognition system according to claim 1, characterized in that, The drone swarm collects data and extracts semantic features from the collected video data. In time The captured video frames are The edge recognition model parameters are The features extracted from the drone are represented as follows: ； in Including facial information, clothing texture, motion trajectory, spatial location, and RF signal features, the feature vector is... Input lightweight classification network Perform identity recognition and calculate the recognition confidence level using a soft maximum function. : 。 3. The feedback-driven multi-UAV cooperative identity recognition system according to claim 1, characterized in that, The edge feature extraction network introduces a channel attention mechanism. Key features are weighted, and edge deployment efficiency is optimized by combining pruning and quantization-aware training. ； in It is the feature vector after weighted optimization; The drone swarm includes a language generation module and cross-modal features. Then input into the semantic generation module Generate text description : ； Among them, language model parameters We employ a deep pre-trained BERT architecture and fine-tune it to adapt to the visual semantic translation task. Features generated by edge feature extraction network Confidence level and text description The data is transmitted to the central control terminal via a wireless communication network, and the current global status information of each UAV is uploaded simultaneously.

4. The feedback-driven multi-UAV cooperative identity recognition system according to claim 1, characterized in that, The global state information is the global state vector of the UAV. : ； in, For spatial location information, For speed, Represents remaining energy. This indicates the current status of the communication link.

5. The feedback-driven multi-UAV cooperative identity recognition system according to claim 1, characterized in that, The central control terminal employs a deep reinforcement learning policy network, and the parameters of the deep reinforcement learning policy network are set as follows: Input status With semantic information Output action decision : ； The objective function for this deep reinforcement learning policy network is: ； in, Corresponding task completion time, As an energy consumption indicator, , , These are the weighting coefficients for each indicator.

6. The feedback-driven multi-UAV cooperative identity recognition system according to claim 1, characterized in that, The feedback scheduling control module includes a path control submodule, an error control submodule, and a communication link modeling submodule. The path control submodule integrates an artificial intelligence path search algorithm with a dynamic window method to generate a dynamic obstacle avoidance path based on the current flight status and target position. The feedback scheduling control module will take action Converted into drone control commands : ； in This is the feedback controller function.

7. The feedback-driven multi-UAV cooperative identity recognition system according to claim 6, characterized in that, The error control submodule is adjusted in real time by a PID control law, and the error calculation is expressed as follows: ； Where t represents the discrete time step number of the current control cycle. The tracking error at the current moment is defined as the target position. Compared with the actual location of the drone The difference, , , These are the proportional, integral, and differential gains, respectively.

8. The feedback-driven multi-UAV cooperative identity recognition system according to claim 6, characterized in that, The communication link modeling submodule defines the signal-to-noise ratio based on the wireless channel characteristics between UAVs. Characterizing the first frame and the first Link quality between drones; bandwidth According to Shannon's formula, the result is: ； Corresponding transmission delay The calculation is as follows: ； in, The allocation of communication resources is dynamically adjusted according to the size of the data to be transmitted, so as to avoid link congestion.

9. The feedback-driven multi-UAV cooperative identity recognition system according to claim 6, characterized in that, The edge recognition device performs incremental training on the collected high-value samples to update the local model parameters: ； in For learning rate, The loss function is trained for the edge. The policy network π is fine-tuned online based on task performance: ； Where γ is the policy network learning rate. To reinforce the learning objective function.

10. The feedback-driven multi-UAV cooperative identity recognition system according to claim 1, characterized in that, The training and deployment of the system are divided into three phases, including: In the offline phase, the recognition and language generation models are trained based on publicly available multimodal datasets, and the parameters are compressed and optimized to adapt to the operating requirements of embedded devices. During the simulation phase, a multi-objective dynamic scenario is constructed based on the simulation platform, and a scheduling policy network is trained to optimize the overall system performance. During the online deployment phase, model parameters are distributed to the drone cluster and central terminal, enabling the system to achieve adaptive closed-loop operation with simultaneous data collection and feedback in the actual environment.

Images