Heterogeneous reconnaissance resource scheduling method and system for urban street combat precision perception, and electronic equipment

Abstract

The present application belongs to the field of military reconnaissance and intelligent scheduling, and aims at the problems of traditional reconnaissance system, such as coverage blind area, inaccurate identification, low efficiency of scheduling and weak robustness. The present application discloses a heterogeneous reconnaissance resource scheduling method and system for precise perception of urban street war, and an electronic device. Through multi-modal data acquisition, target threat dynamic grading determination, heterogeneous reconnaissance resource agent modeling, and resource scheduling based on cooperative game, an integrated closed loop is constructed. The system contains multi-modal acquisition, target threat grading, agent modeling, cooperative scheduling, and communication interaction modules, adopts multi-modal fusion, MAPPO algorithm and other technologies, realizes target precise identification, resource dynamic optimization scheduling and robust operation in strong confrontation environment, and provides support for precise perception and combat decision of urban street war.

Description

Technical Field

[0001] This invention belongs to the field of military reconnaissance and intelligent dispatch technology, specifically involving heterogeneous reconnaissance resource dispatch methods, systems and electronic equipment for precise perception in urban warfare. Background Technology

[0002] Modern urban warfare presents a highly complex and adversarial environment. Densely packed buildings, crisscrossing underground passages, and strong connectivity between indoor and outdoor spaces create an asymmetric combat space characterized by excellent concealment, varied maneuver routes, and a three-dimensional distribution of threats. The enemy frequently utilizes building cover, ruins for concealment, and underground pipelines for maneuver, conducting distributed operations, ambushes, and camouflage evasion, significantly increasing the difficulty of battlefield reconnaissance and situational awareness for our forces.

[0003] Traditional reconnaissance systems generally suffer from the following core flaws, making them unsuitable for the practical needs of urban warfare: Relying on a single reconnaissance platform (such as drones or ground reconnaissance vehicles) or independently operating sensor nodes is limited by field of view, mobility and deployment range, making it impossible to achieve all-round, blind-spot-free coverage of areas such as areas obscured by tall buildings, underground spaces, and complex indoor structures, which can easily create reconnaissance loopholes. In urban environments, targets are affected by changes in lighting, building obstruction, camouflage interference, etc., resulting in weak and dynamic characteristics. Traditional single-modal perception and simple data fusion methods are difficult to achieve accurate identification and stable tracking of concealed targets and high-speed maneuvering targets, which easily leads to missed detections and false alarms, affecting the accuracy of friend-or-foe determination. Existing scheduling methods are mostly static task allocation or scheduling based on fixed rules, lacking adaptability to dynamic changes in the battlefield situation. When sudden firefights occur in key areas, targets shift, or sensors are jammed, it is impossible to quickly adjust the deployment of reconnaissance resources, resulting in idle resources, over-concentration, or lagging reconnaissance in key areas, which is difficult to match the fast-paced combat requirements of urban warfare. The reconnaissance system relies on centralized command and fixed communication links. When some sensor nodes are physically destroyed, encounter electronic interference or communication suppression, the system is prone to functional paralysis and cannot maintain the continuous operation of key reconnaissance functions through autonomous reconstruction and collaborative replacement.

[0004] Therefore, there is an urgent need to build an integrated reconnaissance and perception system that is rapid in response, comprehensive in coverage, accurate in identification, intelligent in scheduling, and robust. This system should systematically integrate heterogeneous reconnaissance resources such as ground reconnaissance vehicles, aerial drones, drop sensors, and individual soldier reconnaissance terminals, and integrate multimodal perception, multi-agent collaboration, and game theory decision-making technologies to achieve closed-loop optimization from real-time intelligence acquisition, multi-source information fusion, accurate target determination to dynamic resource scheduling. This will meet the urgent military demand for high-precision, high-real-time, and high-reliability battlefield situation information in urban warfare. Summary of the Invention

[0005] To address the aforementioned problems, this invention aims to provide a heterogeneous reconnaissance resource scheduling method, system, and electronic equipment for precise perception in urban street fighting. Based on multimodal fusion target threat dynamic classification and judgment technology, heterogeneous resource modeling method of multi-agent system, and dynamic scheduling algorithm combining cooperative game and reinforcement learning, it constructs an integrated closed loop of perception-judgment-modeling-scheduling, solving the problems of blind spots, low recognition accuracy, inefficient scheduling, and weak robustness of traditional reconnaissance systems, and realizing precise coordination and efficient utilization of heterogeneous reconnaissance resources in urban street fighting scenarios.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] On the one hand, a heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting is proposed, the method including the following steps:

[0008] Multimodal data acquisition: Obtaining multimodal data of urban street fighting scenarios through heterogeneous reconnaissance resources;

[0009] The target threat dynamic classification and determination is based on multimodal fusion technology to process the collected multimodal data, determine the target threat level and value attributes, and screen enemy dynamic threat targets;

[0010] Heterogeneous reconnaissance resource intelligent agent modeling: Heterogeneous reconnaissance resources and dynamic enemy threat targets are abstracted into intelligent agents to construct a multi-agent system;

[0011] Collaborative game-based resource scheduling, based on a multi-agent system, transforms the heterogeneous reconnaissance resource scheduling problem into a stochastic game model. Reinforcement learning algorithms are then used to solve the stochastic game model to determine the scheduling and deployment scheme for heterogeneous reconnaissance resources.

[0012] Furthermore, based on multimodal fusion technology, the collected multimodal data is processed to determine the target threat level and value attributes, and to filter out dynamic enemy threat targets, specifically including:

[0013] Based on multimodal data, a dual-branch feature extraction network combining convolutional neural networks and Transformer architecture is used to detect and identify targets.

[0014] Identify friendly and enemy attributes of the identified targets and filter out dynamic threat targets.

[0015] Quantitatively assess the threat level and value attributes of dynamic enemy targets.

[0016] Furthermore, a dual-branch feature extraction network combining convolutional neural networks and the Transformer architecture is used to detect and identify targets, specifically including:

[0017] Construct a multimodal target data training set covering urban combat scenarios with different weather conditions, time periods, and types of obstruction;

[0018] A dual-branch feature extraction network based on the combination of convolutional neural network and Transformer architecture is constructed, and the dual-branch feature extraction network is trained using the training set;

[0019] The trained dual-branch feature extraction network identifies the target's model and status features.

[0020] Furthermore, the identification of friend or foe attributes of the identified targets can be performed using either cooperative or non-cooperative identification methods.

[0021] Furthermore, a quantitative assessment of the threat level and value attributes of dynamic enemy targets is conducted, specifically including:

[0022] Extract the target's static attributes, dynamic attributes, and battlefield context information to construct a comprehensive feature vector of the target;

[0023] Fuzzy set theory is used to process the uncertainty information in the comprehensive feature vector of the target, and combined with the deterministic information, a dynamic Bayesian network is used to complete the probabilistic inference of the threat level of multiple targets.

[0024] By combining the analytic hierarchy process (AHP) and the TOPSIS multi-criteria decision-making model, a threat score ranging from 0 to 100 and a three-level classification of high, medium, and low are quantitatively output, and the target value attributes are evaluated simultaneously.

[0025] Furthermore, heterogeneous reconnaissance resources and dynamic enemy threat targets are abstracted into intelligent agents to construct a multi-agent system, specifically including:

[0026] The intelligent agent classification abstracts ground vehicle-mounted multi-sensor nodes, scattered micro-sensor clusters, aerial drones, and distributed command and control units as functional intelligent agents, and enemy mobile targets as adversarial intelligent agents.

[0027] The internal structure design includes a state machine model and functional logic modules for each intelligent agent.

[0028] Environmental modeling involves constructing an environmental model that encompasses a three-dimensional mesh physical space and an information space that includes a shared intelligence database and a battlefield rule base, and dynamically defining the effective range of sensors and the coverage of communication networks.

[0029] The interaction mechanism is designed to achieve low-latency data exchange using a hierarchical networking communication architecture. It constructs a collaborative strategy through deep reinforcement learning in a multi-agent system and resolves conflicts in adversarial scenarios based on a Nash equilibrium model.

[0030] Furthermore, based on a multi-agent system, the heterogeneous reconnaissance resource scheduling problem is transformed into a stochastic game model. A reinforcement learning algorithm is used to solve the stochastic game model to determine the scheduling and deployment scheme for heterogeneous reconnaissance resources, specifically including:

[0031] Based on a multi-agent system, the heterogeneous reconnaissance resource scheduling problem is transformed into a stochastic game model;

[0032] Maximizing the expected cumulative reward of the random game model is the optimization objective;

[0033] A reinforcement learning algorithm is used to solve the optimization objective and determine the scheduling and deployment scheme of heterogeneous reconnaissance resources.

[0034] Furthermore, based on a multi-agent system, the heterogeneous reconnaissance resource scheduling problem is transformed into a stochastic game model, which is as follows:

[0035]

[0036] in, For the number of agents, For a set of states, For the first The action space of an intelligent agent For the first The reward function of an agent Let be the probability of environmental state transition. This is the loss factor.

[0037] The optimization objective is to maximize the expected cumulative reward of a stochastic game model, where the expected global cumulative reward V for n agents is... total The expression is:

[0038]

[0039] Where k is the number of functional intelligent agents;

[0040] No. The expression for the cumulative expected reward of an agent is:

[0041]

[0042] in, for Constant environmental conditions for Always in coordination For the first The strategy of an agent This is a joint strategy for other intelligent agent models.

[0043] On the other hand, a heterogeneous reconnaissance resource scheduling system for precise perception in urban street fighting is proposed to implement the aforementioned heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting. The system includes:

[0044] The multimodal data acquisition module acquires multimodal data from urban street fighting scenarios;

[0045] The target threat classification module processes the collected multimodal data based on multimodal fusion technology, determines the target threat level and value attributes, and filters out dynamic enemy threat targets.

[0046] The agent modeling module abstracts heterogeneous reconnaissance resources and dynamic enemy threat targets into agents to build a multi-agent system;

[0047] The collaborative scheduling module, based on a multi-agent system, transforms the heterogeneous reconnaissance resource scheduling problem into a stochastic game model. It then uses a reinforcement learning algorithm to solve the stochastic game model and determine the scheduling and deployment scheme for heterogeneous reconnaissance resources.

[0048] On the other hand, an electronic device includes at least one processor; and a memory communicatively connected to the processor; wherein the memory stores instructions executed by the processor to enable the processor to perform the above-described heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting.

[0049] The beneficial effects of this invention are:

[0050] 1. This invention utilizes complementary deployment of heterogeneous reconnaissance resources and multimodal data fusion technology to enable ground reconnaissance vehicles, aerial drones, scattered sensors, and individual soldier terminals to work collaboratively, achieving comprehensive coverage of complex areas such as areas obscured by tall buildings, underground passages, and ruins. Multimodal data is processed through a dual-branch network architecture of CNN and Transformer to extract and fuse features. Combined with self-supervised pre-training and meta-learning strategies, this significantly improves the accuracy of target recognition in complex environments, accurately distinguishing targets such as personnel, vehicles, and drones, as well as their models and states. Regarding friend-or-foe identification, the dual-path fusion scheme considers both cooperative and non-cooperative identification. The improved IFF protocol ensures the security of encrypted interactions, and the graph neural network enables feature inference for non-cooperative targets. After confidence-weighted fusion, the confidence level for friend-or-foe determination remains above 0.95 even in scenarios with strong noise and no response, effectively avoiding the risk of misjudgment. In the threat level and value attribute quantification process, a dynamic Bayesian network is used to process uncertain information. The combined analytic hierarchy process and TOPSIS method outputs accurate scores in the range of 0 to 100 and three-level classifications, providing a quantifiable and comparable scientific basis for scheduling decisions, replacing the traditional fuzzy qualitative assessment model.

[0051] 2. This invention employs multi-agent modeling and cooperative game scheduling techniques to abstract heterogeneous reconnaissance resources and enemy targets into functional and adversarial agents, clarifying their state transition rules and interaction mechanisms, and providing a rigorous mathematical framework for multi-agent adversarial scenarios. By capturing the core characteristics of multi-agent interaction and random environmental changes on the battlefield through a stochastic game model, and combining this with the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm to train the scheduling strategy, dynamic optimization of resource allocation is achieved. When targets shift, electromagnetic interference occurs, or resources fail, the system can adjust the scheduling scheme in real time, prioritizing resource allocation for high-threat targets and rationally allocating resources to low-threat targets to avoid idleness, thereby improving resource utilization and reconnaissance coverage of high-threat targets. During the scheduling process, the design of the KL divergence penalty term and adaptive penalty factor ensures stable and controllable policy updates, avoiding unjustified deviations from the initial scheme; the Nash equilibrium model effectively resolves multi-agent resource competition conflicts, ensuring orderly and efficient scheduling.

[0052] 3. The hierarchical networking communication architecture and emergency self-organizing network mechanism designed in this invention solve the problem of traditional centralized communication systems being paralyzed by interference. The backbone layer ensures high-bandwidth, low-latency data transmission for core modules and resources; the access layer adapts to lightweight data interaction for edge resources; and the emergency layer, in the event of electromagnetic interference or node failure, achieves data transfer through neighboring intelligent agents, improving the retention rate of the system's core functions. The state machine model and functional logic module design of the intelligent agents enable each resource to have autonomous obstacle avoidance, anti-interference, and fault recovery capabilities, significantly improving the system's resilience and continuous operation capability in highly confrontational environments, meeting the high-dynamic and high-risk practical requirements of urban warfare.

[0053] 4. After lightweighting and model distillation, the target detection and recognition model of this invention can be deployed on edge devices such as ground reconnaissance vehicles and individual soldier terminals to achieve real-time inference. The training process of the stochastic game model and the MAPPO algorithm is clear, and the convergence conditions are well-defined, allowing training to be completed using existing computing resources. The deployment logic and communication protocol of heterogeneous resources are standardized, facilitating the upgrade and adaptation of existing equipment. Furthermore, by adjusting threat quantification weights and expanding agent types, this invention adapts to urban warfare scenarios of different scales and terrains, eliminating the need for large-scale system reconstruction and significantly reducing the promotion cost of military applications. Attached Figure Description

[0054] Figure 1 This is a flowchart of the heterogeneous reconnaissance resource scheduling method in this invention.

[0055] Figure 2 This is a flowchart of the target recognition process in this invention.

[0056] Figure 3 This is a flowchart of the non-cooperative friend-or-foe identification process in this invention.

[0057] Figure 4This is a flowchart of the threat level assessment process in this invention.

[0058] Figure 5 This is a schematic diagram illustrating the modeling of the intelligent agent and the environment in this invention. Detailed Implementation

[0059] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] Example 1

[0061] This embodiment provides a heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting. It is based on multimodal fusion target threat dynamic classification and determination technology and multi-agent cooperative game resource scheduling technology. The core is a closed-loop process of multimodal data acquisition - target threat classification - agent modeling - cooperative scheduling, which mainly includes the following steps:

[0062] S1. Multimodal data acquisition: Acquire multimodal data such as visible light, infrared, electromagnetic, and motion trajectory in urban street fighting scenarios through heterogeneous reconnaissance resources.

[0063] In S1, the process is implemented through the following sub-steps: heterogeneous reconnaissance resource deployment → multimodal data acquisition → data preprocessing, as detailed below:

[0064] S101. Based on the perception requirements of urban warfare scenarios and combined with the functional characteristics of heterogeneous reconnaissance resources, ground reconnaissance vehicles, aerial drones, drop-type sensors, and individual soldier reconnaissance terminals are deployed in key areas within the scenario. Ground reconnaissance vehicles are deployed in open areas to cover the ground-based near-field perception range; aerial drones are deployed in different airspaces to achieve wide-area aerial coverage; drop-type sensors are deployed in concealed areas and at passageway entrances and exits for fixed-point monitoring; and individual soldier reconnaissance terminals are carried by forward combat personnel to obtain real-time near-field data, ensuring that the perception ranges of various resources are complementary and that there are no significant blind spots in reconnaissance.

[0065] S102. Through the aforementioned heterogeneous reconnaissance resources, multimodal perception data of urban street fighting scenarios are simultaneously collected. Ground reconnaissance vehicles and aerial drones collect visible light and infrared data, working together to acquire target appearance and thermal radiation characteristics from different perspectives; ground reconnaissance vehicles and individual soldier reconnaissance terminals simultaneously collect electromagnetic data to capture the target's electromagnetic radiation signals; all heterogeneous reconnaissance resources work together to collect motion trajectory data, recording the target's movement path within the scene; projectile sensors collect vibration and acoustic data to assist in identifying concealed or low-speed moving targets. During the collection process, ensure that all types of data are accompanied by a collection timestamp and resource identifier for easy subsequent correlation processing.

[0066] S103. Standardize the collected multimodal data. Specifically, denoise and normalize the size of visible light and infrared data to eliminate environmental noise and differences in image size acquired by different devices. Enhance infrared data under low-light conditions to improve target feature clarity. Filter electromagnetic data to retain effective frequency band signals relevant to the target, remove environmental clutter, and convert time-domain signals into frequency-domain feature forms. Denoise and extract features from vibration and acoustic data to retain key signal features reflecting target motion or activity. Smooth motion trajectory data to remove abnormal positioning points and ensure trajectory continuity. After preprocessing, convert all data into a unified format and store it in a shared data pool to provide standardized data support for subsequent dynamic target threat classification.

[0067] S2. Target Threat Dynamic Classification and Determination: Based on multimodal fusion technology, the collected multimodal data is processed to sequentially complete target detection and identification, friend-or-foe attribute identification, and threat level and value attribute quantitative evaluation, and output the target threat level and value attribute results.

[0068] In S2, the process is implemented through the following sub-steps: target detection and identification → friend-or-foe attribute identification → threat level and value attribute quantitative assessment.

[0069] S201. Target detection and recognition is achieved through a progressive approach of data-network-model-deployment, specifically including the following steps:

[0070] S2011. Construct a multimodal target dataset covering different environmental conditions in urban combat scenarios. Environmental conditions include different weather, different time periods, and different types of obstruction. Data types include visible light data, infrared data, electromagnetic data, and motion trajectory data acquired in the aforementioned multimodal data acquisition phase. The data is labeled, and the labeling content includes target category (personnel, vehicles, UAVs, etc.), target model, and target status (moving, stationary, concealed, etc.) to ensure that the labeling results accurately reflect the true attributes of the target and provide a reliable data foundation for model training.

[0071] S2012. Based on a constructed multimodal target dataset, a dual-branch feature extraction network combining CNN (Convolutional Neural Network) and Transformer architectures was designed: one branch processes visible light data to extract shallow texture and appearance features of the target; the other branch processes infrared data to extract thermal radiation features of the target. A cross-modal attention mechanism is used to achieve complementary fusion of features from the two branches, strengthening the correlation of target features across different modalities and improving feature representation capabilities in complex environments. During training, a self-supervised pre-training strategy is used to initially learn common modal features on unlabeled data, and then a meta-learning strategy is used to fine-tune the network on small sample labeled data, improving the model's generalization ability to unknown targets and complex occlusion scenarios.

[0072] S2013. The trained dual-branch feature extraction network is lightweighted. The model size is compressed through lightweight network structure design and model distillation technology. While ensuring the model's recognition accuracy, the computational complexity and storage requirements of the model are reduced, enabling it to be deployed on edge devices such as ground reconnaissance vehicles and individual soldier reconnaissance terminals. This enables real-time detection and recognition of multimodal targets, outputting key information such as target category, model, and status, providing a foundation for subsequent identification of friend or foe attributes.

[0073] S202. Dual-path fusion is used to achieve friend-or-foe attribute identification, as detailed below:

[0074] S2021. Based on an improved IFF (Identification Friend or Foe) protocol and dynamic coding mechanism, a collaborative friend or foe identification process is constructed: The friendly command and control unit sends an encrypted interrogation signal to the target. The pulse sequence and encryption factor of the interrogation signal dynamically change during the interaction process to prevent the signal from being cracked by the enemy. After receiving the interrogation signal, the friendly target carrying a transponder generates a response signal according to preset rules and sends it back. The command and control unit decodes and verifies the received response signal. If the verification is successful, the target is determined to be friendly, achieving accurate friend or foe identification in collaborative scenarios.

[0075] Specifically, an IFF signal typically consists of a series of pulses, each with a defined width, interval, and amplitude. Common signal parameters:

[0076] Pulse width ( ): The duration of the pulse.

[0077] Pulse interval ( ): The time interval between two adjacent pulses.

[0078] Pulse amplitude ( ): The amplitude of the pulse.

[0079] Assume the query signal is from It consists of 10 pulses, each pulse having a width of 1000. , interval is , amplitude The query signal can be represented as:

[0080]

[0081] in, For time, The pulse number. It is a rectangle function, defined as:

[0082]

[0083] Upon receiving an interrogation signal, the transponder generates a response signal according to a preset protocol. The response signal typically contains specific coded information to distinguish between friendly and enemy forces. The response signal can be represented as:

[0084]

[0085] in, It is the amplitude of the response signal. It is the number of pulses in the response signal.

[0086] After receiving the response signal, the interrogator decodes and analyzes the received signal using a signal processing algorithm.

[0087] Signal processing steps include:

[0088] Signal detection: Detecting whether a received signal exists.

[0089] Signal decoding: Extracting encoded information from a signal.

[0090] Target identification: Determine whether the target is friendly or enemy based on the decoded information.

[0091] S2022. Extract multi-dimensional features of the target, including motion features (target trajectory, velocity variation patterns), structural features (target outline, thermal radiation distribution patterns), and electromagnetic features (target electromagnetic signal frequency band, modulation method). Based on graph neural networks (GNNs), model the relationship between the target and other units in the battlefield environment. Through collaborative reasoning of target features and relationships, achieve target identification in non-cooperative scenarios, especially suitable for identifying enemy targets that are not carrying transponder devices or are camouflaged.

[0092] S2023. Integrate cooperative and non-cooperative friend-or-foe identification results using a confidence-weighted approach: Assign confidence weights to each type of identification result. Cooperative identification results, due to their reliance on explicit interactive verification, receive higher weights, while non-cooperative identification results receive lower weights. Calculate the weighted overall confidence score. If the overall confidence score reaches a preset threshold, a clear friend-or-foe attribute determination result (friendly or enemy) is output. If the overall confidence score is below the threshold, it is marked as "pending confirmation," triggering subsequent supplementary reconnaissance processes to resolve identification conflicts in scenarios with strong noise or missing local information, ensuring the accuracy of friend-or-foe attribute determination.

[0093] S203. Employ a progressive approach of feature-modeling-quantification to achieve quantitative assessment of threat level and value attributes, as detailed below:

[0094] S2031. Extract the target's static attributes, dynamic attributes, and battlefield context information: Static attributes include inherent attributes such as target type and equipment configuration. Dynamic attributes include real-time changing attributes such as target movement trends and activity frequency. Battlefield context information includes environmentally relevant attributes such as the target's geographical location (whether it is located in our key areas) and the target's historical attack behavior. Integrate the above information to construct a comprehensive target feature vector that fully reflects the target's threat potential and strategic value.

[0095] S2032. Based on actual battlefield conditions and the target's movement characteristics and attributes, an urban combat target threat assessment index system is established. This system consists of three levels: target layer, criterion layer, and index layer. The target layer represents the overall target threat level (T), the final quantitative output value used for target priority ranking. The criterion layer includes four core dimensions: target capabilities and attributes (C), target behavior and intent (B), spatial and situational relationships (S), and timeliness and mission relevance (M). The index layer comprises observable and judgmentable parameters under each criterion layer. A fuzzy method is used to divide continuous index data into three fuzzy sets: "low threat," "medium threat," and "high threat," and a membership function is designed. Discrete indicators are then classified and assigned corresponding threat scores. Using threat level and threat intent as parent nodes, a two-layer Bayesian network index system is established based on the dependencies between indicators. The top layer consists of parent nodes for threat level {extremely high, high, medium, low} and threat intent {present, absent}. The middle layer consists of child nodes composed of the criterion layer and the index layer. The bottom layer consists of observation nodes for original battlefield perception evidence. Substitute the fuzzified or graded indicator parameters into the fuzzy rule base, perform fuzzy inference using the "If-Then" rule, and then use the centroid method to defuzzify to obtain the probability vector of each target threat level, thus completing the probabilistic inference of multi-target threat levels.

[0096] S2033. The TOPSIS method, which constructs ideal solutions, is used to quantify the threat level probability vector inferred by fuzzy dynamic Bayesian reasoning. Combining the Analytic Hierarchy Process (AHP) and the TOPSIS multi-criteria decision model, the threat level probability output by the dynamic Bayesian network is quantified: the weights of each indicator in the target's comprehensive feature vector are determined using AHP. The weight allocation is based on the degree of influence of the indicator on threat assessment and can be dynamically adjusted according to different operational phases such as assault and clearing, and defensive holding. Based on the TOPSIS multi-criteria decision model, an initial decision matrix is ​​first constructed and standardized and weighted. Then, the ideal solution and negative ideal solution are determined, and the distance from each target to the ideal and negative ideal solutions is calculated to obtain the relative proximity score Ci. After normalization, the proximity score is converted into a threat score in the range of 0-100. Simultaneously, considering the strategic significance of the target to our operational mission, the target's value attribute is evaluated. Finally, the target's threat score, threat level (high / medium / low), and value attribute results are output, providing a basis for subsequent resource allocation.

[0097] S3. Modeling of Heterogeneous Reconnaissance Resources as Intelligent Agents: The heterogeneous reconnaissance resources and enemy dynamic threat targets are abstracted into intelligent agents. The internal structure, external environment model and interaction mechanism of the intelligent agents are designed to build a multi-agent system.

[0098] In S3, the implementation steps are as follows: agent segmentation → agent internal structure design → environment modeling → interaction mechanism design.

[0099] S301. In urban warfare scenarios, heterogeneous reconnaissance resources and dynamic enemy threat targets are abstracted into intelligent agents, and the intelligent agents are divided as follows:

[0100] S3011 abstracts ground reconnaissance vehicles, aerial drones, drop-type sensors, and distributed command and control units into functional intelligent agents: the ground reconnaissance vehicle intelligent agent has the function of ground multimodal data acquisition and target recognition; the aerial drone intelligent agent has the function of aerial large-area reconnaissance and target tracking; the drop-type sensor intelligent agent has the function of fixed-point area monitoring; the distributed command and control unit intelligent agent has the function of global data fusion, decision-making, and resource scheduling; and the individual soldier reconnaissance terminal is also abstracted into a functional intelligent agent, which has the function of forward close-range reconnaissance.

[0101] S3012. Enemy dynamic threat targets (such as enemy personnel, enemy vehicles, enemy drones, etc.) are abstracted into adversarial intelligent agents. The behavior patterns of such intelligent agents are set based on the enemy's combat characteristics, including covert maneuver, reconnaissance avoidance, and intermittent communication. Their state is dynamically updated with changes in the battlefield environment, providing adversarial scenario support for subsequent collaborative game scheduling.

[0102] S302. Design an internal structure for each agent that includes a state machine model and functional logic modules, as follows:

[0103] S3021. Define different state types and state transition rules for intelligent agents: State types include normal working states (such as reconnaissance, tracking, and monitoring), abnormal response states (such as obstacle avoidance, anti-interference, and fault recovery), and task switching states (such as switching from reconnaissance to tracking, and switching from tracking to return to base). State transition rules are triggered based on changes in the battlefield environment, adjustments to task requirements, and the agent's own state (such as remaining battery power and equipment failure), ensuring that the agent can dynamically adjust its working mode according to the actual situation.

[0104] S3022, the functional logic module includes a perception algorithm, a decision-making module, and a task priority evaluation unit: the perception algorithm processes multimodal data collected or received by the agent to extract target features or environmental information; the decision-making module outputs specific action strategies for the agent (such as movement direction and sensor parameter adjustment) based on the perception results and battlefield rules. The task priority evaluation unit prioritizes multiple tasks undertaken by the agent according to the target threat level, task urgency, and the agent's own capabilities, ensuring that high-priority tasks are executed first.

[0105] S303. Construct an environment model that encompasses both physical and information spaces, as detailed below:

[0106] S3031. The physical environment of urban street fighting scenarios is abstracted into a three-dimensional spatial model, including static elements such as buildings, obstacles, and passages, as well as dynamic elements such as electromagnetic interference and weather changes. The effective range of sensors for each functional intelligent agent is dynamically defined and updated in real time based on the agent's location and equipment performance. Simultaneously, the coverage area of ​​the communication network is defined, clarifying the communication reachability between different intelligent agents and providing physical spatial constraints for agent interaction.

[0107] S3032. Construct an information space consisting of a shared intelligence repository, a battlefield rule repository, and an adversarial knowledge repository: The shared intelligence repository stores multimodal data collection results, dynamic target threat classification results, and agent status information, enabling real-time information sharing. The battlefield rule repository includes rules such as operational mission requirements and resource scheduling constraints (e.g., prioritizing resource allocation for high-threat targets). The adversarial knowledge repository stores the behavioral patterns and response strategies of enemy targets, providing adversarial references for agent decision-making.

[0108] S304. Design a low-latency interaction mechanism between intelligent agents to support multi-agent collaborative work, as follows:

[0109] S3041. A layered networking approach is adopted to achieve data exchange between intelligent agents: The backbone layer connects distributed command and control unit intelligent agents with core functional intelligent agents such as ground reconnaissance vehicles and aerial drones, ensuring high-bandwidth, low-latency data transmission. The access layer connects edge intelligent agents such as scattered sensors and individual soldier reconnaissance terminals with backbone layer intelligent agents, employing communication methods suitable for short-range, low-power scenarios. The emergency layer is designed for communication interruption scenarios, supporting intelligent agent self-organizing networks and achieving data transmission through relay by neighboring intelligent agents, ensuring communication robustness.

[0110] S3042. Construct agent cooperation strategies through Multi-Agent Deep Reinforcement Learning (MADRL) to enable functional agents to autonomously adjust their actions based on the global situation, achieving task cooperation (such as the coordination between UAV tracking and ground reconnaissance vehicle close-in identification). Design a conflict resolution mechanism based on the Nash equilibrium model. When multiple agents compete for the same task resource or have conflicting actions, calculate the task benefits and resource consumption of each agent, prioritizing the allocation of tasks to agents with higher overall benefits, ensuring efficient conflict resolution during the cooperation process.

[0111] S4. Resource scheduling based on collaborative game theory: The resource scheduling problem is transformed into a stochastic game model. A reinforcement learning algorithm is used to train the scheduling strategy model. Combined with the target threat level and value attribute results, the heterogeneous reconnaissance resources are dynamically optimized and allocated to form the optimal reconnaissance resource scheduling scheme, so as to achieve accurate perception and efficient reconnaissance in urban street fighting scenarios.

[0112] In S4, the process is implemented through the sub-steps of building a stochastic game model → training a reinforcement learning algorithm → dynamic scheduling optimization, specifically including the following sub-steps:

[0113] S401. The resource scheduling problem is transformed into a stochastic game model, which is defined as follows:

[0114]

[0115] S4011, Definitions of parameters in the random game model:

[0116] (Number of agents): Includes all functional agents and adversarial agents. The number is determined based on the actual heterogeneous reconnaissance resources and the number of enemy targets.

[0117] (State set): encompasses battlefield environment state (such as electromagnetic interference intensity, weather), functional agent state (such as location, working state), enemy target state (such as location, movement trend), and target threat level results;

[0118] (Action Space): Contains the executable actions of each agent. Functional agent actions include reconnaissance, tracking, obstacle avoidance, and return to base, while adversarial agent actions include movement, concealment, and evasion.

[0119] T (State Transition Probability): Describes the probability that the battlefield environment state will transition from the current state to the next state after the agent performs an action. It is set based on the characteristics of the battlefield environment and the effect of the agent's action.

[0120] R i(Reward Function): Set reward rules for each functional intelligent agent, give positive rewards for completing effective actions such as reconnaissance of high-threat targets and efficient resource utilization, and give negative penalties for ineffective or negative actions such as idle resources and mission failure;

[0121] γ (loss factor): Balances the short-term gains of the agent's actions with the long-term cumulative gains, and is set according to the real-time and continuous requirements of the combat mission.

[0122] Specifically, the resource scheduling problem is transformed into a stochastic game model—constructing a unified high-dimensional game decision-making representation driven by dynamic threat and resource utility, and introducing a joint optimization logic of risk entropy and information gain. Risk entropy is used to quantify battlefield uncertainties (such as the probability of enemy target concealment, the intensity of electromagnetic interference, and perception blurring caused by terrain obstruction). By calculating the entropy value of the enemy target state distribution (the higher the entropy value, the stronger the uncertainty), high-risk areas prioritized for reconnaissance resource coverage are determined. Information gain is used to measure the value of acquiring effective intelligence from heterogeneous reconnaissance resources. Based on the target threat level and the distribution of reconnaissance blind spots, the intelligence increment under different resource deployment schemes is calculated (such as supplementary acquisition of high-threat target characteristics by UAV close-range reconnaissance, and the coverage gain of blind spots by distributed sensors), ensuring that resource scheduling is tilted towards tasks with high information gain.

[0123] By jointly optimizing risk entropy and information gain, the results of uncertainty quantification and intelligence value assessment are integrated into the state set of a stochastic game model. With reward function In the set of states The new dimensions of risk entropy and information gain potential are added to the reward function. The system provides positive rewards for scheduling actions that reduce entropy in high-risk areas and increase information gain of key targets, while negatively penalizing actions that allocate resources to areas with low information gain or fail to cover high-risk areas. Subsequently, a reinforcement learning algorithm is used to train the scheduling strategy model, and the heterogeneous reconnaissance resources are dynamically optimized and allocated based on the target threat level and value attribute results.

[0124] S4012. The optimization objective is to maximize the expected cumulative reward of the random game model, where the expected global cumulative reward V for n agents is... total The expression is:

[0125]

[0126] Where k is the number of functional intelligent agents;

[0127] No. The expected cumulative reward for a functional intelligent agent is:

[0128]

[0129] in, for Constant environmental conditions for Always in coordination For intelligent agents strategy, It is a joint strategy for other agents; the expectation reflects the long-term cumulative benefit that an agent can obtain by executing a certain strategy in a specific initial state, providing an optimization target for reinforcement learning training.

[0130] S402. The scheduling policy model is trained using the Multi-Agent Proximity Policy Optimization (MAPPO) algorithm, as detailed below:

[0131] S4021. Training parameter settings, specifically:

[0132] State input: The set of states in the stochastic game model As input to the algorithm, it covers key battlefield information observations;

[0133] Policy definition: A policy is a mapping from state to action probability distribution, that is, the probability of an agent choosing each action in a certain state;

[0134] Action space: The action space is defined using a stochastic game model. Define the range of scheduling actions that the intelligent agent can execute;

[0135] Reward Model: Based on Reward Function Construct a reward model, input the initial scheduling scheme (a scheme that pre-allocates resources based on the target threat level) and the actual scheduling actions (actions performed by the agent), and output a scalar reward value to quantify the effect of a single scheduling action.

[0136] S4022. The specific implementation of the training process is as follows:

[0137] Data collection: Allow the agent to interact in an environment constructed by a stochastic game model and collect trajectory data including state, action, and reward;

[0138] Reward Calculation: The scalar reward value for each trajectory data is calculated through the reward model. At the same time, KL divergence is used as a scheduling difference penalty to measure the difference between the initial scheduling and the actual scheduling, so as to avoid excessive policy update.

[0139] Penalty factor adjustment: The trust region is dynamically adjusted through an adaptive penalty factor β. When the KL divergence exceeds 1.5 times the target divergence, β is doubled, and when it is less than 2 / 3 times the target divergence, β is halved to ensure stable policy updates.

[0140] Policy iteration: Based on the collected trajectory data and the calculated rewards and penalties, the policy model is iteratively updated using the gradient ascent method to maximize the agent's cumulative reward expectation until the policy converges and a stable scheduling policy model is obtained.

[0141] Specifically, MAPPO is an extension of PPO that learns a policy distribution and a centered value function based on global states rather than local observations, thereby extending the single-agent PPO algorithm to multi-agent scenarios. The PPO algorithm is a reinforcement learning algorithm that maximizes the expected cumulative reward by alternately updating the policy. It aims to optimize the policy to maximize the expected cumulative reward while ensuring that the policy update magnitude is limited to stabilize the training process.

[0142] For a certain strategy Allow the agent to be in a state Select action Then gain profits. This yields the probability distribution of a trajectory sequence consisting of multiple states and actions:

[0143]

[0144] in, It is the set of states / actions of an agent under a certain policy. This is the strategy sequence number.

[0145] The corresponding reward is Therefore, the objective function can be expressed as:

[0146]

[0147] The PPO algorithm aims to maximize the objective and employs gradient ascent.

[0148]

[0149] In this invention, the scheduling strategy needs to be continuously optimized to maximize the reward. This process usually requires multiple iterations and model updates to gradually improve the model's performance.

[0150] Specifically, environmental parameters need to be included. Corresponding initial scheduling and actual scheduling As a data pair of [environmental parameters and scheduling parameters], it is passed to the reward model to obtain a scalar reward. Secondly, and The penalty term for comparing and calculating the differences is typically used as a metric. The final PPO formula is as follows:

[0151]

[0152] in, For scheduling that needs updating, This is the initial scheduling. Hyperparameters. This is an adaptive penalty factor. When KL is too large, exceeding the preset target divergence by more than 1.5 times, it is doubled. The penalty is increased by adjusting the value; conversely, if KL is too small, less than 2 / 3 times the preset target divergence, the penalty is halved. This reduces the severity of punishment and effectively expands the trust zone.

[0153] S403. Based on the trained scheduling strategy model, and combined with the output target threat level and value attribute results, dynamically optimize the allocation of heterogeneous reconnaissance resources, as follows:

[0154] S4031. Based on the target threat level and value attributes, prioritize the allocation of functional intelligent agent resources to high-threat and high-value targets, generate an initial scheduling plan, and ensure that the reconnaissance needs of key targets are met first.

[0155] S4032. Based on changes in the battlefield environment (such as enemy target movement, electromagnetic interference) and updates to the agent's status (such as equipment failure, changes in remaining power), the scheduling scheme is adjusted in real time through the scheduling strategy model: for high-threat targets, if the currently allocated reconnaissance resources are insufficient or invalid, resources are replenished or replaced in a timely manner; for low-threat targets, resources are reasonably adjusted to avoid idleness; for newly emerging targets, idle resources are quickly allocated for reconnaissance to ensure that the scheduling scheme adapts to the dynamic battlefield situation.

[0156] S4033. Verify the scheduling effect through indicators such as resource utilization rate, target reconnaissance coverage rate, and scheduling response delay: resource utilization rate reflects the efficient utilization of heterogeneous reconnaissance resources, target reconnaissance coverage rate ensures that no key targets are missed, and scheduling response delay ensures the real-time performance of the scheduling plan; based on the verification results, further fine-tune the scheduling strategy model, continuously optimize the resource scheduling effect, and ultimately achieve accurate perception and efficient reconnaissance in urban street fighting scenarios.

[0157] Example 2

[0158] This embodiment provides a heterogeneous reconnaissance resource scheduling system for precise perception in urban warfare, specifically including:

[0159] Multimodal data acquisition module: Composed of heterogeneous reconnaissance resources, including ground reconnaissance vehicles, aerial drones, drop sensors, and individual soldier reconnaissance terminals, used to collect multimodal data in urban warfare scenarios.

[0160] Specifically, the multimodal data acquisition module is the hardware execution unit for multimodal data acquisition, composed of heterogeneous reconnaissance resources. These include a ground reconnaissance vehicle unit equipped with visible light acquisition components, infrared acquisition components, and electromagnetic signal receiving components, enabling close-range multimodal data acquisition; an aerial UAV unit equipped with high-resolution visible light and infrared cameras, covering a wide range of aerial acquisition scenarios; a distributed sensor unit containing vibration and acoustic sensors, employing distributed deployment to achieve point-to-point acquisition in concealed areas; and a single-soldier reconnaissance terminal unit integrating a miniature visible light acquisition module and a short-range electromagnetic detection module to meet the needs of close-range data acquisition at the forefront. Each acquisition unit is deployed in key areas of urban warfare (open areas, airspace, concealed locations, etc.), and synchronous acquisition is triggered by command and control commands to ensure time alignment of visible light, infrared, electromagnetic, vibration, acoustic, and motion trajectory data. Each acquisition unit has a built-in lightweight preprocessing submodule that performs noise reduction and format conversion according to the "data standardization" requirements in the methodology. The standardized data is then transmitted to a shared data pool via a communication interaction module, serving as the input source for the target threat classification module.

[0161] Target Threat Classification Module: Used to perform target detection and identification, friend-or-foe attribute identification, and threat level and value attribute quantitative evaluation on the multimodal data, and output the target threat level and value attribute results.

[0162] Specifically, the target threat classification module is the functional unit responsible for the dynamic classification and determination of target threats. It includes a target detection and recognition submodule with a built-in dual-branch feature extraction network (CNN and Transformer) and a lightweight inference engine; a friend-or-foe attribute identification submodule containing a cooperative IFF protocol processing unit and a non-cooperative graph neural network inference unit and result fusion unit; and a threat and value quantification submodule integrating a dynamic Bayesian network (DBN) inference model and an AHP-TOPSIS quantization model. The target threat classification module receives standardized data from the multimodal data acquisition module, calls the dual-branch network to extract features and infer target category, model, and status, and outputs this information to the friend-or-foe attribute identification submodule. It obtains identification results through cooperative IFF verification and non-cooperative feature inference, and outputs friend-or-foe attributes after confidence-weighted fusion. Finally, it integrates target identification results, friend-or-foe attributes, and battlefield context information, calls the DBN and AHP-TOPSIS models, outputs a threat score in the range of 0-100, high / medium / low threat levels, and value attributes, and transmits this information to the collaborative scheduling module.

[0163] Intelligent agent modeling module: used to abstract the heterogeneous reconnaissance resources and enemy dynamic threat targets into intelligent agents, design the internal structure, external environment model and interaction mechanism of the intelligent agents, and build the intelligent agent system.

[0164] Specifically, the agent modeling module serves as the framework building unit for modeling heterogeneous reconnaissance resource agents. It includes an agent partitioning submodule responsible for abstracting heterogeneous reconnaissance resources (functional) and enemy targets (adversarial) into agents; an agent structure design submodule that designs the state machine model and functional logic modules of the agents; an environment modeling submodule that constructs an environment model integrating physical and information spaces; and an interaction mechanism design submodule that designs the communication architecture and collaborative strategies between agents. It abstracts ground reconnaissance vehicles and drones as functional agents and enemy targets as adversarial agents, configuring state machines (such as reconnaissance, tracking, and obstacle avoidance states) and functional logic (perception, decision-making, and task evaluation) for each agent. It models urban street fighting environments, designs a layered network communication architecture and a conflict resolution strategy based on Nash equilibrium, providing a multi-agent system framework for the collaborative scheduling module.

[0165] The collaborative scheduling module transforms the resource scheduling problem into a stochastic game model, trains the scheduling strategy model using reinforcement learning algorithms, and dynamically optimizes the allocation of heterogeneous reconnaissance resources. This module serves as the decision-making and execution unit for resource scheduling based on collaborative game theory. It includes a stochastic game model construction submodule that transforms the resource scheduling problem into a stochastic game model, a reinforcement learning training submodule with a built-in Multi-Agent Proximal Policy Optimization (MAPPO) algorithm, and a dynamic scheduling optimization submodule responsible for generating and adjusting resource allocation schemes. Based on the multi-agent system output from the agent modeling module, it defines stochastic game model parameters (number of agents, state set, reward function, etc.), calls the MAPPO algorithm to train the scheduling strategy model, and maximizes the expected cumulative reward of agents. Combining the threat results output from the target threat classification module, it generates an initial scheduling scheme (prioritizing resource allocation to high-threat targets), and optimizes the scheme in real time according to battlefield changes (such as enemy movement, resource failures) according to the "dynamic adjustment" logic in the method, outputting scheduling instructions to the multimodal data acquisition module.

[0166] The communication interaction module is used to realize low-latency data transmission and information sharing between the multimodal data acquisition module, target threat classification module, agent modeling module, collaborative scheduling module, and agents. This module serves as the information connection hub for the above four modules and agents in the system. It includes a layered communication submodule consisting of a backbone layer (connecting core modules and resources) and an access layer (connecting edge resources), a data interaction submodule responsible for data format conversion, encryption, and timing alignment, and an emergency communication submodule supporting agent self-organizing networks in case of communication interruption. During normal communication, large-capacity data (such as multimodal images and scheduling schemes) is transmitted through the backbone layer, while lightweight data (such as sensor signals and status commands) is transmitted through the access layer, ensuring real-time synchronization of information across modules. During emergency communication, an agent self-organizing network is triggered when communication is interrupted, relaying data through adjacent resources, prioritizing the reconnaissance and scheduling command flow for high-threat targets, and maintaining the core functions of the system.

[0167] Specific application examples

[0168] Street fighting broke out in an old urban area. The area is densely populated with buildings, interspersed with narrow alleys and an abandoned underground pipeline. The enemy used buildings for cover and ruins for concealment to carry out dispersed operations. Our side needs to quickly ascertain the enemy's deployment in the area to provide support for subsequent tactical operations. Therefore, the heterogeneous reconnaissance resource scheduling system of this invention was activated, as detailed below:

[0169] After system startup, the multimodal data acquisition module begins operation. Ground reconnaissance vehicles are deployed at open intersections on the outskirts of the block, simultaneously activating visible light and infrared acquisition components and electromagnetic detectors to cover the ground-based short-range sensing range. Three aerial drones take off and patrol the east, south, and north quadrants of the block, capturing target signals in building gaps and rooftop areas at low altitude. Ten drop-type sensors are precisely deployed at underground pipeline entrances and exits, around ruins, and at corners of major alleyways. Vibration sensors monitor vibration signals generated by the movement of personnel or vehicles, while acoustic sensors capture acoustic activity of concealed targets. Four individual reconnaissance terminals, carried by forward combat personnel, infiltrate observation points on the edge of the block to collect real-time short-range multimodal data. Each acquisition unit operates synchronously according to command and control instructions, attaching timestamps and resource identifiers to visible light images, infrared thermal images, electromagnetic signals, vibration data, acoustic data, and motion trajectory data. After lightweight preprocessing (denoising, format conversion, etc.), the data is transmitted to the shared data pool via the communication interaction module.

[0170] After receiving standardized data from the shared data pool, the target threat classification module initiates a progressive processing flow. The target detection and identification submodule calls a dual-branch feature extraction network based on a CNN and Transformer architecture to fuse features from the infrared thermal images transmitted by the UAV and the visible light images from the ground reconnaissance vehicle, accurately identifying three suspected targets (two clusters of personnel and one small vehicle). The friend-or-foe identification submodule first sends encrypted interrogation signals to the target area via an improved IFF protocol, but receives no valid response. It then extracts the target's movement trajectory (showing covert maneuvering characteristics), electromagnetic signal frequency band (matching known enemy communication frequency bands), and structural features (vehicle shape matches enemy reconnaissance vehicle parameters). Through graph neural network inference and confidence-weighted fusion, it determines that all three targets are enemy, with a confidence level of 0.95. The Threat and Value Quantification submodule constructs a comprehensive feature vector of the target, and combines dynamic Bayesian network inference and AHP-TOPSIS quantification model to output threat scores: personnel clusters are scored 82 and 76 respectively (both high threat), and small vehicles are scored 88 (extremely high threat). The value attribute is marked as "priority reconnaissance and monitoring target".

[0171] The intelligent agent modeling module then completes modeling based on collected resources and enemy targets: ground reconnaissance vehicles, drones, scattered sensors, individual soldier terminals, and command and control units are abstracted into functional intelligent agents, and the state machine model (such as the drone's cruise, tracking, and obstacle avoidance states) and functional logic of each intelligent agent are defined. Three enemy targets are abstracted into adversarial intelligent agents, and their stealthy maneuvering and reconnaissance evasion behavior patterns are defined. An environmental model covering the three-dimensional physical space of the street (marking the locations of buildings, alleys, and ruins) and information space (shared intelligence database, battlefield rule database, and adversarial knowledge database) is constructed, defining the effective sensor range and communication coverage of each intelligent agent. A hierarchical networking communication architecture and multi-agent collaborative strategy are designed, and conflict resolution rules are formulated based on the Nash equilibrium model, laying the framework for subsequent scheduling.

[0172] Based on the above results, the collaborative scheduling module initiates resource optimization allocation: First, an initial scheduling plan is generated, dispatching two drones to the high-threat personnel cluster area to perform tracking tasks, one drone to focus on monitoring the extremely high-threat small vehicle, a ground reconnaissance vehicle to approach the alleyway where the small vehicle is located for close-range detection, and a scattered sensor to enhance signal acquisition in the surrounding area of ​​the target, while the individual soldier terminal maintains edge vigilance. During the scheduling execution, the small vehicle is detected starting and moving towards the underground pipeline network, and electromagnetic interference signals appear in the surrounding area. The collaborative scheduling module immediately dynamically adjusts the plan: instructing one of the drones tracking the personnel cluster to turn around, forming a two-way encirclement with the drone originally monitoring the vehicle, maintaining continuous lock on the vehicle. The ground reconnaissance vehicle adjusts its route and blocks another exit of the underground pipeline network. The scattered sensor affected by electromagnetic interference starts an ad-hoc network through the emergency communication submodule, relaying data with the help of adjacent sensors. At the same time, the individual soldier terminal is instructed to detour to both sides of the vehicle's movement path to supplement close-range reconnaissance data.

[0173] The communication module ensures seamless data flow: Under normal conditions, the backbone layer transmits high-capacity images and dispatch commands from UAVs and reconnaissance vehicles, while the access layer transmits lightweight sensor signal data, with latency controlled within 100ms. When electromagnetic interference occurs in a localized area, it automatically switches to emergency communication mode, triggering an intelligent agent self-organizing network to ensure uninterrupted transmission of critical reconnaissance data and dispatch commands.

[0174] Ultimately, through the collaborative work of its various modules, the system continuously locked onto the positions and movement status of three enemy targets, and output battlefield situation reports in real time, providing precise and efficient support for our side to formulate tactical strike and containment plans.

[0175] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting, characterized in that, The method includes the following steps: Multimodal data acquisition: Obtaining multimodal data of urban street fighting scenarios through heterogeneous reconnaissance resources; The target threat dynamic classification and determination is based on multimodal fusion technology to process the collected multimodal data, determine the target threat level and value attributes, and screen enemy dynamic threat targets; Heterogeneous reconnaissance resource intelligent agent modeling: Heterogeneous reconnaissance resources and dynamic enemy threat targets are abstracted into intelligent agents to construct a multi-agent system; Collaborative game-based resource scheduling, based on a multi-agent system, transforms the heterogeneous reconnaissance resource scheduling problem into a stochastic game model. Reinforcement learning algorithms are then used to solve the stochastic game model to determine the scheduling and deployment scheme for heterogeneous reconnaissance resources.

2. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 1, characterized in that, Multimodal fusion technology is used to process the collected multimodal data, determine the target threat level and value attributes, and filter out dynamic enemy threat targets. Specifically, this includes: Based on multimodal data, a dual-branch feature extraction network combining convolutional neural networks and Transformer architecture is used to detect and identify targets. Identify friendly and enemy attributes of the identified targets and filter out dynamic threat targets. Quantitatively assess the threat level and value attributes of dynamic enemy targets.

3. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 2, characterized in that, A two-branch feature extraction network combining convolutional neural networks and the Transformer architecture is used for target detection and recognition, specifically including: Construct a multimodal target data training set covering urban combat scenarios with different weather conditions, time periods, and types of obstruction; A dual-branch feature extraction network based on the combination of convolutional neural network and Transformer architecture is constructed, and the dual-branch feature extraction network is trained using the training set; The trained dual-branch feature extraction network identifies the target's model and status features.

4. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 2, characterized in that, The identification of friend or foe attributes of the identified targets can be performed using either a cooperative or non-cooperative identification method.

5. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 2, characterized in that, Quantitatively assess the threat level and value attributes of dynamic enemy targets, specifically including: Extract the target's static attributes, dynamic attributes, and battlefield context information to construct a comprehensive feature vector of the target; Fuzzy set theory is used to process the uncertainty information in the comprehensive feature vector of the target, and combined with the deterministic information, a dynamic Bayesian network is used to complete the probabilistic inference of the threat level of multiple targets. By combining the analytic hierarchy process (AHP) and the TOPSIS multi-criteria decision-making model, a threat score ranging from 0 to 100 and a three-level classification of high, medium, and low are quantitatively output, and the target value attributes are evaluated simultaneously.

6. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 1, characterized in that, Heterogeneous reconnaissance resources and dynamic enemy threat targets are abstracted into intelligent agents to construct a multi-agent system, specifically including: The intelligent agent classification abstracts ground vehicle-mounted multi-sensor nodes, scattered micro-sensor clusters, aerial drones, and distributed command and control units as functional intelligent agents, and enemy mobile targets as adversarial intelligent agents. The internal structure design includes a state machine model and functional logic modules for each intelligent agent. Environmental modeling involves constructing an environmental model that encompasses a three-dimensional mesh physical space and an information space that includes a shared intelligence database and a battlefield rule base, and dynamically defining the effective range of sensors and the coverage of communication networks. The interaction mechanism is designed to achieve low-latency data exchange using a hierarchical networking communication architecture. It constructs a collaborative strategy through deep reinforcement learning in a multi-agent system and resolves conflicts in adversarial scenarios based on a Nash equilibrium model.

7. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 1, characterized in that, Based on a multi-agent system, the heterogeneous reconnaissance resource scheduling problem is transformed into a stochastic game model. A reinforcement learning algorithm is used to solve the stochastic game model to determine the scheduling and deployment scheme for heterogeneous reconnaissance resources, specifically including: Based on a multi-agent system, the heterogeneous reconnaissance resource scheduling problem is transformed into a stochastic game model; Maximizing the expected cumulative reward of the random game model is the optimization objective; A reinforcement learning algorithm is used to solve the optimization objective and determine the scheduling and deployment scheme of heterogeneous reconnaissance resources.

8. The heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting according to claim 7, characterized in that, Based on a multi-agent system, the heterogeneous reconnaissance resource scheduling problem is transformed into a stochastic game model, the expression of which is: ; in, For the number of agents, For the set of agent states, For the first The action space of an intelligent agent For the first The reward function of an agent Let be the probability of environmental state transition. Let V be the loss factor. The optimization objective is to maximize the expected cumulative reward of the stochastic game model, where V is the expected global cumulative reward for n agents. total The expression is: ; Where k is the number of functional intelligent agents; No. The expression for the cumulative expected reward of an agent is: ; in, for Constant environmental conditions for Always in coordination For the first The strategy of an agent This is a joint strategy for other intelligent agent models.

9. A heterogeneous reconnaissance resource scheduling system for precise perception in urban street fighting, used to implement the heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting as described in any one of claims 1 to 8, characterized in that, The system includes: The multimodal data acquisition module acquires multimodal data from urban street fighting scenarios; The target threat classification module processes the collected multimodal data based on multimodal fusion technology, determines the target threat level and value attributes, and filters out dynamic enemy threat targets. The agent modeling module abstracts heterogeneous reconnaissance resources and dynamic enemy threat targets into agents to build a multi-agent system; The collaborative scheduling module, based on a multi-agent system, transforms the heterogeneous reconnaissance resource scheduling problem into a stochastic game model. It then uses a reinforcement learning algorithm to solve the stochastic game model and determine the scheduling and deployment scheme for heterogeneous reconnaissance resources.

10. An electronic device, characterized in that, The electronic device includes at least one processor; and a memory communicatively connected to the processor; wherein the memory stores instructions that are executed by the processor, the instructions being executed by the processor to enable the processor to perform the heterogeneous reconnaissance resource scheduling method for precise perception in urban street fighting as described in any one of claims 1 to 8.

Images