Non-cooperative target searching method based on identification-planning joint optimization

Abstract

The invention discloses a non-cooperative target searching method based on identification-planning joint optimization, and belongs to the technical field of unmanned aerial vehicles. A terrain perception multi-mode hybrid filter is constructed to accurately predict a trajectory after a target is separated from a visual field, high-reliability re-identification of the target is realized in combination with a fine-grained feature decoupling and dynamic fusion mechanism, and an identification-planning joint optimization adaptive search method is adopted to generate an optimal search path. The problems of long-time tracking and re-identification of the non-cooperative target in a complex environment are solved, and the continuous working capability of an unmanned aerial vehicle monitoring system is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicles (UAVs), and in particular relates to a non-cooperative target search method based on recognition-planning joint optimization. Background Art

[0002] In security applications such as counterterrorism, arrests, and border patrols, drones are crucial for long-term, continuous tracking of non-cooperative targets (such as suspicious vehicles or individuals). However, during actual tracking, the target may enter blind spots such as no-fly zones, underground parking lots, or inside buildings, resulting in loss of visual contact between the drone and the target. In these situations, the drone needs to predict the target's likely location based on on-site information, perform efficient search, and accurately re-identify the target to restore continuous surveillance.

[0003] Currently, the following key technical challenges exist in the long-term tracking and re-identification of drones:

[0004] 1. Trajectory prediction failure in complex scenarios: Traditional trajectory prediction methods, such as Kalman filtering or particle filtering, experience a sharp drop in prediction accuracy when the target enters complex terrain (such as urban environments, mountains, and hills), and are unable to effectively cope with the coupled motion patterns of the target and the environment.

[0005] 2. High false alarm rate in cross-domain feature re-identification: Existing target re-identification algorithms, such as single feature matching and twin networks, are prone to misidentification or missed identification when the target undergoes significant appearance changes or environmental conditions change over time, making it difficult to address the problem of temporal and spatial heterogeneity of features.

[0006] 3. Complexity of search planning under multiple constraints: Given the limited energy resources of drones and the vast search space, how can we generate an optimal search path under the multiple constraints of speed, terrain, and resources to achieve efficient target recapture?

[0007] 4. Disconnection between the recognition and planning systems: Existing methods typically treat target recognition and search planning as two independent modules, lacking deep integration and information sharing mechanisms. Key information from the recognition system, such as confidence and feature reliability, fails to effectively guide search strategy adjustments. Furthermore, the search system's environmental perception and resource constraints fail to provide feedback to optimize recognition thresholds. This results in poor overall system coordination and an inability to dynamically adjust search behavior based on recognition results.

[0008] Currently, the main solutions to the above problems include:

[0009] 1. Trajectory Prediction Methods: Traditional methods primarily employ state estimation methods such as Kalman filters, extended Kalman filters, and unscented Kalman filters, based on kinematic models for trajectory prediction. While these methods offer acceptable performance in simple scenarios, prediction accuracy significantly decreases when the target enters complex terrain or performs complex maneuvers, with average prediction errors exceeding 30% of the actual position.

[0010] 2. Object Re-ID Methods: Existing methods primarily extract appearance features based on deep convolutional neural networks, such as ResNet and DenseNet architectures, and use metric learning methods (such as Siamese networks or triplet loss) for feature matching. These methods lack modeling of environmental and temporal factors in feature representation, resulting in re-ID accuracy rates often below 70% in cross-domain scenarios.

[0011] 3. Search Planning Methods: Traditional search planning methods, such as heuristic search and probability map-based search, are typically optimized based on simple search strategies, such as maximizing coverage or minimizing information entropy reduction, using a single objective function. These methods lack unified consideration of multiple constraints and are difficult to achieve optimal search results with limited resources.

[0012] 4. Attempts at Recognition-Planning Integration: In recent years, some research has begun to attempt to integrate target recognition with path planning, such as attention-based region screening methods and perception-guided planning frameworks. However, these methods primarily employ simple serial processing models—recognition first, then planning—or employ fixed-threshold rule switching mechanisms. These methods lack the ability to model the uncertainty of recognition results and make dynamic decisions. For example, while the typical probabilistic roadmap method (PRM) considers the probability distribution of target occurrence, it cannot dynamically adjust the search strategy based on recognition confidence. Active search methods based on information gain, while able to maximize information acquisition, struggle to balance recognition needs with resource constraints.

[0013] The fundamental limitation of the above methods lies in the lack of deep integration and closed-loop feedback mechanism between the recognition system and the planning system. Specifically:

[0014] 1. One-way information flow: Most systems only implement one-way information transmission from recognition to planning, and the planning results cannot be fed back to optimize the recognition process.

[0015] 2. Static decision threshold: Traditional methods usually use fixed recognition thresholds and search ranges, which cannot be dynamically adjusted according to task progress and environmental changes.

[0016] 3. Split objective function: The recognition system and planning system use different objective functions for independent optimization, resulting in local optimality rather than global optimality.

[0017] 4. Unreasonable resource allocation: When the target location cannot be accurately determined, the system finds it difficult to strike a balance between search breadth and depth, resulting in wasted resources or inefficient search. Summary of the Invention

[0018] In order to overcome the shortcomings of the existing technology, the present invention provides a non-cooperative target search method based on recognition-planning joint optimization. By constructing a terrain-aware multi-mode hybrid filter, the trajectory of the target after it leaves the field of view is accurately predicted. The fine-grained feature decoupling and dynamic fusion mechanism are combined to achieve highly reliable target re-identification. The adaptive search method of recognition-planning joint optimization is used to generate the optimal search path, which solves the problem of long-term tracking and re-identification of non-cooperative targets in complex environments and significantly improves the continuous working capability of the UAV surveillance system.

[0019] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0020] Step 1: Terrain-aware multi-modal hybrid filtering prediction;

[0021] Step 2: Object re-identification with fine-grained feature decoupling and dynamic fusion;

[0022] Step 3: Adaptive search for joint optimization of identification and planning.

[0023] Preferably, the step 1 is specifically:

[0024] Step 1-1: Terrain semantic segmentation and environment modeling;

[0025] Step 1-1-1: Based on the high-resolution terrain imagery acquired by the drone, use the UNet++ semantic segmentation network to divide the environment into different categories: , , represents the total number of terrain categories, Indicates different terrain types; Represents a set of terrain categories;

[0026] Step 1-1-2: Build a terrain adjacency graph , where the vertex Indicates area, edge Indicates the connection relationship between regions;

[0027] Step 1-1-3: Extract key nodes: , Respectively represent the 1st to the key nodes, Indicates the total number of key nodes;

[0028] Step 1-1-4: Define the traffic characteristic function for each type of terrain: ,in 、 Represent the minimum and maximum travel speeds for each type of terrain, represents the velocity attenuation coefficient;

[0029] Step 1-2: Multimodal motion model construction;

[0030] Step 1-2-1: Define three basic motion models;

[0031] (1) Constant speed linear motion model CV: , express The state vector at time t, represents the constant velocity model state transfer matrix, express The state vector at time t, represents the constant velocity model process noise;

[0032] (2) Constant acceleration motion model CA: , represents the state transfer matrix of the constant acceleration model, represents the process noise of the constant acceleration model;

[0033] (3) Steering motion model CT: , represents the nonlinear state transfer function of the steering model, represents the steering model process noise;

[0034] Step 1-2-2: Each motion model is:

[0035]

[0036]

[0037] It is a nonlinear function including angular velocity parameters;

[0038] in, is the time step;

[0039] Steps 1-3: Terrain constraint model fusion;

[0040] Step 1-3-1: Construct the state transfer function under terrain constraints:

[0041]

[0042] in is the terrain influencing factor, according to the terrain type Adjust state transfer; represents the process noise vector, represents the state transition matrix of the motion model, represents the terrain constraint state transfer function;

[0043] Step 1-3-2: Position constraint;

[0044] Make sure the predicted location is in a traversable area:

[0045]

[0046] in Represents the state vector Extract location, is the set of traversable areas, is the projection function; represents the position constraint function;

[0047] Step 1-3-3: speed constraint;

[0048] Adjust speed range based on terrain type:

[0049]

[0050] in Represents the state vector extraction speed, is the speed adjustment function; represents the velocity constraint function, The minimum travel speed function representing the terrain type, The maximum travel speed function representing the terrain type;

[0051] Steps 1-4: Interactive multi-model spatiotemporal robust filtering;

[0052] Step 1-4-1: Build an interacting multi-model IMM architecture, including filter combinations: , Filters representing constant speed, constant acceleration, and steering motion models respectively;

[0053] Step 1-4-2: Define the model transition probability matrix: ,in Represents the model Transfer to Model probability;

[0054] Step 1-4-3: Model probability update:

[0055]

[0056] in For the moment Time Model The probability of For the model The likelihood function of is the observation data; Represents the k-1 moment model The probability of Represents the model at time k-1 The probability of Indicates the total number of models;

[0057] Step 1-4-4: state fusion prediction;

[0058]

[0059] in For the model Status prediction;

[0060] Steps 1-5: Multipath hypothesis generation;

[0061] Step 1-5-1: Consider multiple possible paths to the target and build a set of path hypotheses , Respectively represent items 1 to 2 The path hypothesis, represents the total number of path hypotheses;

[0062] Step 1-5-2: Assume for each path , , calculate its probability: in Indicates consistency with historical trajectory, Indicates compatibility with the terrain; Represents historical state trajectory data;

[0063] Step 1-5-3: Make predictions along each hypothetical path to generate possible locations;

[0064]

[0065] in is the prediction time window; represents the trajectory prediction function;

[0066] Step 1-5-4: Assign probability weights to each position;

[0067]

[0068] in is the time decay factor;

[0069] Step 1-5-5: Output location set: .

[0070] Preferably, the step 2 is specifically as follows:

[0071] Step 2-1: Feature decoupling representation learning;

[0072] Step 2-1-1: Design a three-way feature extraction network to extract features of different dimensions respectively;

[0073] Appearance Feature Network : Use the ResNet50-IBN backbone network to extract the target visual appearance features;

[0074] Motion Feature Network : Use the spatiotemporal graph convolutional network ST-GCN to extract target motion pattern features;

[0075] Contextual Feature Network : Use Transformer encoder to extract contextual features of the interaction between the target and the environment;

[0076] Step 2-1-2: Construct a multi-dimensional feature representation for the original target;

[0077] : 256-dimensional appearance feature vector;

[0078] : 128-dimensional motion feature vector;

[0079] : 192-dimensional context feature vector;

[0080] in, 、 、 Represent the image sequence, trajectory data and context data of the original target respectively;

[0081] Step 2-2: candidate target detection and feature extraction;

[0082] Step 2-2-1: Use the YOLOv5 object detector to detect candidate objects in the search area video: ; represents the target detection function, Indicates the search area video;

[0083] Step 2-2-2: Extract multidimensional features for each candidate target:

[0084]

[0085]

[0086]

[0087] in, Indicates the The appearance feature vector of candidate targets, Indicates the The motion feature vectors of candidate targets, Indicates the The context feature vector of candidate targets, Indicates the A sequence of candidate target images, Indicates the The trajectory data of candidate targets, Indicates the contextual data of candidate targets; , Indicates the total number of candidate targets;

[0088] Step 2-2-3: Construct candidate target feature set;

[0089]

[0090] Step 2-3: Feature reliability assessment under spatiotemporal conditions;

[0091] Step 2-3-1: Design a feature reliability evaluation function to measure the reliability of features in each dimension based on time intervals and environmental changes:

[0092]

[0093]

[0094]

[0095] in, The target loss duration, is a measure of environmental change, 、 are all weight parameters; represents the reliability evaluation function of appearance features, represents the motion feature reliability evaluation function, represents the context feature reliability evaluation function;

[0096] Step 2-3-2: Determine feature fusion weights;

[0097]

[0098]

[0099]

[0100] in, represents the appearance feature fusion weight, represents the motion feature fusion weight, represents the context feature fusion weight, 、 、 、 Respectively represent the reliability value of appearance feature, motion feature, context feature, and each dimension feature;

[0101] Step 2-4: Adaptive feature fusion matching;

[0102] Step 2-4-1: Multi-dimensional similarity calculation;

[0103]

[0104]

[0105]

[0106] in, represents the cosine similarity function, represents the appearance feature similarity function, represents the motion feature similarity function, represents the context feature similarity function;

[0107] Step 2-4-2: Dynamic fusion similarity:

[0108]

[0109] Step 2-4-3: Position prior enhancement;

[0110] The similarity is further adjusted based on the distance between the candidate target and the predicted position:

[0111]

[0112] in Represents the candidate target location and predicted position distance, is the position prior weight, is the distance attenuation parameter; represents the final similarity function;

[0113] Step 2-5: Cross-domain consistency verification and decision-making;

[0114] Step 2-5-1: Design an adaptive threshold function;

[0115]

[0116] in is the basic threshold, is the maximum increment, is the time coefficient;

[0117] Step 2-5-2: Preliminary screening meets candidate targets;

[0118] Step 2-5-3: Verify the cross-domain consistency of the initial screening results:

[0119] (1) Multi-angle observation verification: obtain matching scores under different perspectives;

[0120] (2) Temporal consistency verification: short-term tracking verifies the consistency of motion patterns;

[0121] Step 2-5-4: Final decision;

[0122] (1) When there are candidate targets that meet the conditions, select the one with the highest score: ;

[0123] (2) Otherwise, it is judged as “target not found”;

[0124] (3) Output re-identification results: , Represents the re-identification confidence.

[0125] Preferably, the step 3 is specifically:

[0126] Step 3-1: Search resource modeling and constraint definition;

[0127] Step 3-1-1: UAV resource constraint modeling;

[0128] (1) Energy constraints: ,in To search for energy needed for the mission, for available energy;

[0129] (2) Time constraints: ,in is the total search time, is the maximum allowed time;

[0130] Step 3-1-2: Define the UAV motion consumption model;

[0131] (1) Energy consumption: ,in For distance, Consumption for hovering observation; represents the energy consumption coefficient; Respectively represent and target search points;

[0132] (2) Time consumption: ,in is the average speed, is the observation time;

[0133] Step 3-2: Identify a confidence-driven search strategy:

[0134] Step 3-2-1: Based on the re-identification confidence , define three search modes:

[0135] (1) High confidence mode ;

[0136] (2) Medium confidence mode ;

[0137] (3) Low confidence mode or there are no candidate targets;

[0138] Step 3-2-2: Design different search parameters for each mode;

[0139] (1) High confidence mode: , ;

[0140] (2) Medium confidence mode: , ;

[0141] (3) Low confidence mode: , ;

[0142] in, represents the search radius, Indicates the single point observation time;

[0143] Step 3-3: candidate point clustering and hierarchical representation;

[0144] Step 3-3-1: Cluster the predicted locations using the density clustering algorithm DBSCAN according to the location aggregation: , T represents the clustering result set, Respectively represent the 1st to the clusters;

[0145] Step 3-3-2: Calculate the comprehensive weight of each cluster: , ;

[0146] Step 3-3-3: Build a hierarchical representation;

[0147] (1) Hot spots are high-weight clusters: ;

[0148] (2) Secondary areas, namely medium-weight clusters: ;

[0149] (3) Low probability areas are low weight clusters:

[0150] in , is the clustering weight threshold;

[0151] Step 3-4: Identify-plan joint optimization objective function:

[0152] Step 3-4-1: Define the basic search point utility function;

[0153]

[0154] in For the point The observation coverage value at represents weight;

[0155] Step 3-4-2: Integrate re-identification feedback to enhance the utility function:

[0156]

[0157] in To identify the enhancement coefficient, is the candidate target position indicator function; represents the utility function of fusion re-identification feedback;

[0158] Step 3-4-3: Define the search path cost function:

[0159]

[0160] in, Indicates the first Search points, Indicates the first Search points; Indicates the search path, Indicates the total number of search points in the search path;

[0161] Step 3-4-4: Define the search path time function;

[0162]

[0163] Step 3-4-5: Construct the identification-planning joint optimization objective function:

[0164]

[0165] in , is the balance parameter;

[0166] Steps 3-5: Hierarchical search strategy generation;

[0167] Step 3-5-1: Adjust the search strategy based on recognition confidence:

[0168] (1) High confidence mode: adopts a depth-first search strategy to prioritize exploring the target’s surroundings;

[0169] (2) Medium confidence mode: adopts a balanced search strategy that takes into account both depth and breadth;

[0170] (3) Low confidence mode: using a breadth-first search strategy to cover multiple possible areas;

[0171] Step 3-5-2: Constructing a search optimization problem:

[0172]

[0173] Constraints:

[0174] in, Indicates the maximum available energy, Indicates the maximum allowed search time;

[0175] Step 3-5-3: Use the improved branch and bound algorithm to solve;

[0176] (1) Initialization: Select the starting search area based on the recognition pattern;

[0177] (2) Branching strategy: adjust branching tendency according to recognition confidence;

[0178] (3) Pruning strategy: pruning when a path violates energy or time constraints;

[0179] Step 3-5-4: Output the optimal search path.

[0180] Preferably, the different terrain types include roads, buildings, and open areas.

[0181] Preferably, the key nodes include intersections and building entrances.

[0182] An electronic device comprises: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the above-mentioned non-cooperative target search method.

[0183] A computer-readable storage medium stores a computer program, which implements the above-mentioned non-cooperative target search method when executed by a processor.

[0184] A chip includes: a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the above-mentioned non-cooperative target search method.

[0185] A computer program product includes a computer storage medium storing a computer program, wherein the computer program includes instructions that can be executed by at least one processor, and when the instructions are executed by the at least one processor, the above-mentioned non-cooperative target search method is implemented.

[0186] The beneficial effects of the present invention are as follows:

[0187] 1. Terrain-aware multi-modal hybrid filtering prediction technology: This technology builds terrain semantic segmentation and environmental modeling, integrates environmental constraints into a multi-modal motion model, and uses an interactive multi-model spatiotemporal robust filter for target trajectory prediction. This addresses the issue of reduced accuracy in traditional trajectory prediction methods in complex terrain, improving the accuracy of target position prediction in blind spots to 82.7%, 43.5% higher than traditional Kalman filtering, achieving highly reliable target prediction in complex environments.

[0188] 2. Fine-grained feature decoupling and dynamic fusion technology: This paper designs a three-way feature extraction network to extract appearance, motion, and context features respectively, and introduces a feature reliability assessment mechanism under spatiotemporal conditions to achieve dynamic weight adjustment and fusion of feature dimensions. This solves the problem of low accuracy of traditional re-identification methods in cross-domain scenarios, and improves the re-identification accuracy to 91.3%, which is 22.8% higher than the single feature matching method, effectively reducing the false alarm rate and missed detection rate.

[0189] 3. Adaptive search technology for joint optimization of recognition and planning: This invention solves the problem of low coupling between traditional search methods and recognition results by constructing a multimodal search framework based on re-identification confidence, designing a joint optimization objective function for recognition and planning, and dynamically adjusting the search strategy according to different confidence levels. It shortens the target recovery time by 56.2% and improves energy utilization efficiency by 37.4%, significantly improving the system's search efficiency.

[0190] 4. Multi-scenario verification and performance evaluation method: This invention realizes comprehensive verification of the system in simulation environment and actual flight by constructing a test scenario matrix with multiple environments, multiple targets and multiple difficulties, and solves the problem of single-scenario evaluation and insufficient data in existing systems. The verification results show that the average completion rate of the system in different environments reaches 92.3%, which is 43.4% higher than the traditional method, proving the stability and adaptability of the system in complex scenarios. DETAILED DESCRIPTION

[0191] The present invention is further described below with reference to the embodiments.

[0192] The present invention proposes a non-cooperative target search method based on recognition-planning joint optimization. By constructing a terrain-aware multimodal motion model, designing a fine-grained feature decoupling and fusion mechanism, and a recognition-planning joint optimization strategy, it realizes bidirectional information flow and closed-loop feedback of target re-identification and search planning, enabling the two to work together and enhance each other, greatly improving the system's target recovery capability and resource utilization efficiency in complex environments.

[0193] The specific embodiment of the present invention comprises the following steps:

[0194] Step 1: Terrain-aware multi-modal hybrid filtering prediction;

[0195] Step 1-1: Terrain semantic segmentation and environment modeling;

[0196] Step 1-1-1: Based on the high-resolution terrain imagery acquired by the drone, use the UNet++ semantic segmentation network to divide the environment into different categories: , , represents the total number of terrain categories, Indicates different terrain types; Represents a set of terrain categories (such as roads, buildings, open areas, etc.);

[0197] Step 1-1-2: Build a terrain adjacency graph , where the vertex Indicates area, edge Indicates the connection relationship between regions;

[0198] Step 1-1-3: Extract key nodes (such as intersections, building entrances, etc.): , Respectively represent the 1st to the key nodes, Indicates the total number of key nodes;

[0199] Step 1-1-4: Define the traffic characteristic function for each type of terrain: ,in 、 Represent the minimum and maximum travel speeds for each type of terrain, represents the velocity attenuation coefficient;

[0200] Step 1-2: Multimodal motion model construction;

[0201] Step 1-2-1: Define three basic motion models;

[0202] (1) Constant speed linear motion model CV: , express The state vector at time t, represents the constant velocity model state transfer matrix, express The state vector at time t, represents the constant velocity model process noise;

[0203] (2) Constant acceleration motion model CA: , represents the state transfer matrix of the constant acceleration model, represents the process noise of the constant acceleration model;

[0204] (3) Steering motion model CT: , represents the nonlinear state transfer function of the steering model, represents the steering model process noise;

[0205] Step 1-2-2: Each motion model is:

[0206]

[0207]

[0208] It is a nonlinear function including angular velocity parameters;

[0209] in, is the time step;

[0210] Steps 1-3: Terrain constraint model fusion;

[0211] Step 1-3-1: Construct the state transfer function under terrain constraints:

[0212]

[0213] in is the terrain influencing factor, according to the terrain type Adjust state transfer; represents the process noise vector, represents the state transition matrix of the motion model, represents the terrain constraint state transfer function;

[0214] Step 1-3-2: Position constraint;

[0215] Make sure the predicted location is in a traversable area:

[0216]

[0217] in Represents the state vector Extract location, is the set of traversable areas, is the projection function; represents the position constraint function;

[0218] Step 1-3-3: speed constraint;

[0219] Adjust speed range based on terrain type:

[0220]

[0221] in Represents the state vector extraction speed, is the speed adjustment function; represents the velocity constraint function, The minimum travel speed function representing the terrain type, The maximum travel speed function representing the terrain type;

[0222] Steps 1-4: Interactive multi-model spatiotemporal robust filtering;

[0223] Step 1-4-1: Build an interacting multi-model IMM architecture, including filter combinations: , Filters representing constant speed, constant acceleration, and steering motion models respectively;

[0224] Step 1-4-2: Define the model transition probability matrix: ,in Represents the model Transfer to Model probability;

[0225] Step 1-4-3: Model probability update:

[0226]

[0227] in For the moment Time Model The probability of For the model The likelihood function of is the observation data; Represents the k-1 moment model The probability of Represents the model at time k-1 The probability of Indicates the total number of models;

[0228] Step 1-4-4: state fusion prediction;

[0229]

[0230] in For the model Status prediction;

[0231] Steps 1-5: Multipath hypothesis generation;

[0232] Step 1-5-1: Consider multiple possible paths to the target and build a set of path hypotheses , Respectively represent items 1 to 2 The path hypothesis, represents the total number of path hypotheses;

[0233] Step 1-5-2: Assume for each path , , calculate its probability: in Indicates consistency with historical trajectory, Indicates compatibility with the terrain; Represents historical state trajectory data;

[0234] Step 1-5-3: Make predictions along each hypothetical path to generate possible locations;

[0235]

[0236] in is the prediction time window; represents the trajectory prediction function;

[0237] Step 1-5-4: Assign probability weights to each position;

[0238]

[0239] in is the time decay factor;

[0240] Step 1-5-5: Output location set: .

[0241] The core innovation of this step lies in combining environmental terrain information with multimodal motion models to construct a spatiotemporal robust filter with terrain perception capabilities, which can accurately predict the possible motion trajectory of the target in a complex environment and provide precise guidance for subsequent search planning.

[0242] Step 2: Object re-identification with fine-grained feature decoupling and dynamic fusion;

[0243] Step 2-1: Feature decoupling representation learning;

[0244] Step 2-1-1: Design a three-way feature extraction network to extract features of different dimensions respectively;

[0245] Appearance Feature Network : Use the ResNet50-IBN backbone network to extract the target visual appearance features;

[0246] Motion Feature Network : Use the spatiotemporal graph convolutional network ST-GCN to extract target motion pattern features;

[0247] Contextual Feature Network : Use Transformer encoder to extract contextual features of the interaction between the target and the environment;

[0248] Step 2-1-2: Construct a multi-dimensional feature representation for the original target;

[0249] : 256-dimensional appearance feature vector;

[0250] : 128-dimensional motion feature vector;

[0251] : 192-dimensional context feature vector;

[0252] in, 、 、 Represent the image sequence, trajectory data and context data of the original target respectively;

[0253] Step 2-2: candidate target detection and feature extraction;

[0254] Step 2-2-1: Use the YOLOv5 object detector to detect candidate objects in the search area video: ; represents the target detection function, Indicates the search area video;

[0255] Step 2-2-2: Extract multidimensional features for each candidate target:

[0256]

[0257]

[0258]

[0259] in, Indicates the The appearance feature vector of candidate targets, Indicates the The motion feature vectors of candidate targets, Indicates the The context feature vector of candidate targets, Indicates the A sequence of candidate target images, Indicates the The trajectory data of candidate targets, Indicates the contextual data of candidate targets; , Indicates the total number of candidate targets;

[0260] Step 2-2-3: Construct candidate target feature set;

[0261]

[0262] Step 2-3: Feature reliability assessment under spatiotemporal conditions;

[0263] Step 2-3-1: Design a feature reliability evaluation function to measure the reliability of features in each dimension based on time intervals and environmental changes:

[0264]

[0265]

[0266]

[0267] in, The target loss duration, is a measure of environmental change, 、 are all weight parameters; represents the reliability evaluation function of appearance features, represents the motion feature reliability evaluation function, represents the context feature reliability evaluation function;

[0268] Step 2-3-2: Determine feature fusion weights;

[0269]

[0270]

[0271]

[0272] in, represents the appearance feature fusion weight, represents the motion feature fusion weight, represents the context feature fusion weight, 、 、 、 Respectively represent the reliability value of appearance feature, motion feature, context feature, and each dimension feature;

[0273] Step 2-4: Adaptive feature fusion matching;

[0274] Step 2-4-1: Multi-dimensional similarity calculation;

[0275]

[0276]

[0277]

[0278] in, represents the cosine similarity function, represents the appearance feature similarity function, represents the motion feature similarity function, represents the context feature similarity function;

[0279] Step 2-4-2: Dynamic fusion similarity:

[0280]

[0281] Step 2-4-3: Position prior enhancement;

[0282] The similarity is further adjusted based on the distance between the candidate target and the predicted position:

[0283]

[0284] in Represents the candidate target location and predicted position distance, is the position prior weight, is the distance attenuation parameter; represents the final similarity function;

[0285] Step 2-5: Cross-domain consistency verification and decision-making;

[0286] Step 2-5-1: Design an adaptive threshold function;

[0287]

[0288] in is the basic threshold, is the maximum increment, is the time coefficient;

[0289] Step 2-5-2: Preliminary screening meets candidate targets;

[0290] Step 2-5-3: Verify the cross-domain consistency of the initial screening results:

[0291] (1) Multi-angle observation verification: obtain matching scores under different perspectives;

[0292] (2) Temporal consistency verification: short-term tracking verifies the consistency of motion patterns;

[0293] Step 2-5-4: Final decision;

[0294] (1) When there are candidate targets that meet the conditions, select the one with the highest score: ;

[0295] (2) Otherwise, it is judged as “target not found”;

[0296] (3) Output re-identification results: , Re-identification confidence;

[0297] This step achieves highly reliable target re-identification through fine-grained feature decoupling and dynamic feature fusion. The core innovation lies in the introduction of feature reliability assessment under spatiotemporal conditions, enabling the system to dynamically adjust the weights of different feature dimensions based on the duration of target loss and environmental changes, effectively addressing the high false alarm rate in cross-domain feature re-identification.

[0298] Step 3: Adaptive search for joint optimization of identification and planning;

[0299] Step 3-1: Search resource modeling and constraint definition;

[0300] Step 3-1-1: UAV resource constraint modeling;

[0301] (1) Energy constraints: ,in To search for energy needed for the mission, for available energy;

[0302] (2) Time constraints: ,in is the total search time, is the maximum allowed time;

[0303] Step 3-1-2: Define the UAV motion consumption model;

[0304] (1) Energy consumption: ,in For distance, Consumption for hovering observation; represents the energy consumption coefficient; Respectively represent and target search points;

[0305] (2) Time consumption: ,in is the average speed, is the observation time;

[0306] This module clearly defines the boundaries of the drone's available resources and the resource consumption model for each activity. Energy constraints ensure that the planned path is within the drone's capacity, while time constraints ensure that the search is completed within the target's likely dwell time window. Furthermore, by establishing precise energy and time consumption models, it provides a computational foundation for subsequent path optimization.

[0307] Step 3-2: Identify a confidence-driven search strategy:

[0308] Step 3-2-1: Based on the re-identification confidence , define three search modes:

[0309] (1) High confidence mode ;Perform a refined verification search, focusing on the area around high-confidence targets;

[0310] (2) Medium confidence mode ;Perform a balanced search, focusing on candidate targets while not giving up other possible areas;

[0311] (3) Low confidence mode or no candidate targets; perform wide-area exploratory search, covering multiple high-probability areas;

[0312] Step 3-2-2: Design different search parameters for each mode;

[0313] (1) High confidence mode: , ;

[0314] (2) Medium confidence mode: , ;

[0315] (3) Low confidence mode: , ;

[0316] in, represents the search radius, Indicates the single point observation time;

[0317] The core function of this module is to directly convert the confidence level of the re-identification results into the basis for selecting a search strategy. In high-confidence mode, the system believes that the target's location is essentially certain, so the search range is small but the observation time is long to obtain more details to confirm the target's identity. In medium-confidence mode, the system has some certainty about the target's location but still has doubts, so the search range is moderate, balancing observation time and coverage. In low-confidence mode, the system is highly uncertain about the target's location and adopts a large-scale rapid scanning method to prioritize potential targets. This confidence-based search mode switching mechanism enables the system to adaptively adjust resource allocation strategies based on the recognition results.

[0318] Step 3-3: candidate point clustering and hierarchical representation;

[0319] Step 3-3-1: Cluster the predicted locations using the density clustering algorithm DBSCAN according to the location aggregation: , T represents the clustering result set, Respectively represent the 1st to the clusters;

[0320] Step 3-3-2: Calculate the comprehensive weight of each cluster: , ;

[0321] Step 3-3-3: Build a hierarchical representation;

[0322] (1) Hot spots are high-weight clusters: ;

[0323] (2) Secondary areas, namely medium-weight clusters: ;

[0324] (3) Low probability areas are low weight clusters:

[0325] in , is the clustering weight threshold;

[0326] This module organizes the set of possible locations generated in step 1 into a structured form. Using a density clustering algorithm, it groups spatially close prediction points and calculates the overall weight of each cluster. This hierarchical representation transforms the search space from a discrete set of points into a hierarchical set of regions, facilitating the formulation of subsequent search strategies. Hotspots represent the most likely locations for the target, secondary regions represent suboptimal possibilities, and low-probability regions serve as alternatives. This structured representation significantly simplifies the search space and improves search efficiency.

[0327] Step 3-4: Identify-plan joint optimization objective function:

[0328] Step 3-4-1: Define the basic search point utility function;

[0329]

[0330] in For the point The observation coverage value at represents weight;

[0331] Step 3-4-2: Integrate re-identification feedback to enhance the utility function:

[0332]

[0333] in To identify the enhancement coefficient, is the candidate target position indicator function; represents the utility function of fusion re-identification feedback;

[0334] Step 3-4-3: Define the search path cost function:

[0335]

[0336] in, Indicates the first Search points, Indicates the first Search points; Indicates the search path, Indicates the total number of search points in the search path;

[0337] Step 3-4-4: Define the search path time function;

[0338]

[0339] Step 3-4-5: Construct the identification-planning joint optimization objective function:

[0340]

[0341] in , is the balance parameter;

[0342] This module is the core innovation of this step, which realizes the deep integration of recognition results and planning decisions. Basic search point utility function Considering the prior probability and observation value of the location; the utility function of re-identification enhancement The confidence of the re-identification result is further integrated into the utility calculation, and additional rewards are given to the identified candidate target locations, and the reward coefficient is proportional to the confidence. and time function Calculate the energy consumption and time cost of the path separately. The final joint optimization objective function is The three key factors are balanced: search point utility (including recognition feedback), energy consumption and time cost, through the weight parameter and This joint optimization objective function ensures that the search decision not only considers the location prior probability, but also fully utilizes the re-identification feedback, while taking into account resource constraints, thus achieving the best balance between utility and cost.

[0343] Steps 3-5: Hierarchical search strategy generation;

[0344] Step 3-5-1: Adjust the search strategy based on recognition confidence:

[0345] (1) High confidence mode: adopts a depth-first search strategy to prioritize exploring the target’s surroundings;

[0346] (2) Medium confidence mode: adopts a balanced search strategy that takes into account both depth and breadth;

[0347] (3) Low confidence mode: using a breadth-first search strategy to cover multiple possible areas;

[0348] Step 3-5-2: Constructing a search optimization problem:

[0349]

[0350] Constraints:

[0351] in, Indicates the maximum available energy, Indicates the maximum allowed search time;

[0352] Step 3-5-3: Use the improved branch and bound algorithm to solve;

[0353] (1) Initialization: Select the starting search area based on the recognition pattern;

[0354] (2) Branching strategy: adjust branching tendency according to recognition confidence;

[0355] (3) Pruning strategy: pruning when a path violates energy or time constraints;

[0356] Step 3-5-4: Output the optimal search path.

[0357] This module is responsible for converting the aforementioned objective function and constraints into a specific search path.

[0358] The working mechanisms of the three search modes in actual applications are as follows:

[0359] In high-confidence mode, the system primarily performs a depth-first search within a 200-meter radius around the candidate target. This means the algorithm prioritizes fully exploring one area before moving on to the next. Single-point observation time is set to 30 seconds to obtain high-quality images and behavioral features. In the branch-and-bound algorithm, high-confidence mode favors branches closer to the candidate target, with a higher pruning threshold, allowing for a more detailed local search.

[0360] In medium-confidence mode, the system conducts a balanced search within a 500-meter radius around the candidate target. This strategy prioritizes the candidate target area while also not neglecting other areas with high probability. The single-point observation time is set to 20 seconds, ensuring sufficient identification information without excessively delaying the search process. In the branch-and-bound algorithm, medium-confidence mode balances node utility and distance, using a moderate pruning strategy to achieve a balance between depth and breadth.

[0361] In low-confidence mode, the system performs a breadth-first search within a 1000-meter radius. This strategy prioritizes coverage of all high-probability areas over in-depth exploration of a single region. Single-point observation time is shortened to 15 seconds, sacrificing some observation quality in exchange for wider coverage. In the branch-and-bound algorithm, low-confidence mode favors branch selection over unexplored areas, and the pruning strategy is more relaxed, prioritizing search breadth.

[0362] Through this hierarchical search strategy generation mechanism, the system can dynamically adjust the search behavior according to the confidence of the re-identification results, maximizing the probability of target recovery under limited resources.

[0363] This step implements an adaptive search strategy based on re-identification confidence by constructing a joint recognition-planning optimization framework. Its core innovation lies in directly integrating target recognition results into the search planning decision-making process, designing differentiated search modes for different confidence levels, and significantly improving search efficiency and target recovery rates. This approach not only optimizes resource utilization but also establishes a bidirectional feedback mechanism between the recognition and planning systems, enabling them to work together and reinforce each other, forming a true closed-loop system.

[0364] Step 4: System Verification and Performance Evaluation

[0365] Step 4-1: Simulation environment construction and parameter setting;

[0366] Step 4-1-1: Construct three typical simulation environments;

[0367] (1) Urban environment: high-density buildings, complex road networks, and multiple no-fly zones;

[0368] (2) Mountainous environment: undulating terrain, vegetation cover, and areas with limited vision;

[0369] (3) Open terrain: flat terrain, sparse obstacles, and good visibility;

[0370] Step 4-1-2: Configure simulation parameters;

[0371] (1) UAV parameters: maximum speed 30 m / s, maximum flight time 60 minutes, sensing range 500 m;

[0372] (2) Target parameters: Personnel moving speed 1-3 m / s, vehicle moving speed 5-20 m / s;

[0373] (3) Environmental parameters: map size 5 km × 5 km, no-fly zone ratio 15%, blind spot ratio 25%;

[0374] Step 4-2: Simulation test scenario design;

[0375] Step 4-2-1: Design a test scenario matrix;

[0376] (1) Target type: personnel tracking, vehicle tracking, mixed target tracking;

[0377] (2) Type of vision loss: entering a building, passing through a tunnel, entering an underground parking lot;

[0378] (3) Re-identification difficulty: slight change (same target), moderate change (clothing change / speed change), significant change (transfer / deformation);

[0379] Step 4-2-2: Run simulation test;

[0380] (1) Each scenario was repeated 20 times, with randomized initial conditions;

[0381] (2) Record key performance indicators: prediction accuracy, re-identification accuracy, target reacquisition time, energy efficiency, and task completion rate;

[0382] Step 4-3: Actual flight test verification;

[0383] Step 4-3-1: Conduct field testing using a multi-rotor drone platform;

[0384] (1) Configuration: Six-axis multi-rotor platform, equipped with a high-definition visible light camera (4K / 60fps), an infrared camera (640×512 / 30fps), and a small laser radar (16 lines, 30m range);

[0385] (2) Computing platform: edge computing unit (8-core CPU, GPU acceleration, 8GB RAM);

[0386] Step 4-3-2: Conduct tests in three real-world environments;

[0387] (1) Suburban areas: medium building density, regular road network, and some no-fly zones;

[0388] (2) Mountainous and hilly areas: The terrain is undulating, covered with vegetation, and there are many areas with restricted vision;

[0389] (3) Open space: flat terrain, few obstacles, and good visibility;

[0390] Step 4-3-3: Perform typical tracking tasks;

[0391] (1) Personnel tracking: The target enters and exits a building and passes through an obstructed area;

[0392] (2) Vehicle tracking: The target passes through the tunnel and enters and exits the parking lot;

[0393] (3) Complex scenes: multiple targets mixed together, frequently entering and exiting occluded areas;

[0394] Step 4-3-4: Collect measured data: flight records, video streams, sensor data, and decision logs;

[0395] Step 4-4: Performance evaluation and comparative analysis;

[0396] Step 4-4-1: Evaluate key performance indicators;

[0397] (1) Target position prediction accuracy: the deviation between the predicted position and the actual position;

[0398] (2) Re-identification accuracy: the success rate of correctly identifying the target;

[0399] (3) False alarm rate of re-identification: the error rate of misidentification as a target;

[0400] (4) Average target reacquisition time: the average time from loss to reacquisition of the target;

[0401] (5) Energy efficiency: search coverage area per unit energy consumption;

[0402] (6) System completion rate: the proportion of tracking tasks successfully completed;

[0403] Step 4-4-2: Comparison with existing technology methods;

[0404] (1) Traditional Kalman filter + single feature matching method;

[0405] (2) Probabilistic map search + deep re-identification method;

[0406] (3) Improved particle filtering + multi-feature fusion method;

[0407] Step 4-4-3: Analyze system performance in different scenarios;

[0408] (1) Adaptability to different environmental conditions;

[0409] (2) Tracking success rate of different target types;

[0410] (3) System robustness in scenarios of varying difficulty;

[0411] Step 4-4-4: Generate a comprehensive evaluation report to verify the system's long-term tracking and re-identification capabilities in complex environments.

[0412] This step comprehensively evaluated the system's performance in various scenarios through simulation testing and real-world flight verification. The test results demonstrated that the non-cooperative target search method, based on joint recognition-planning optimization, demonstrated excellent prediction accuracy, re-identification capabilities, and search efficiency in complex environments and with diverse targets, validating the system's effectiveness and practical value.

[0413] Example:

[0414] To verify the effectiveness of this invention, the system was tested in three typical scenarios (urban, mountainous, and open terrain). The test tasks included vehicle tracking, person tracking, and mixed target tracking. The experiments employed six multirotor drones equipped with high-definition visible light cameras, infrared cameras, and a small LiDAR. Key evaluation metrics included target prediction accuracy, re-identification accuracy, search efficiency, and overall system completion rate. The test results are shown in Table 1.

[0415] Table 1 Test results

[0416] Performance indicators Method of the present invention Traditional Kalman filtering + single feature matching Probabilistic Map Search + Deep Re-ID Improved particle filtering + multi-feature fusion Target position prediction accuracy (%) 82.7 39.2 53.6 65.3 Re-identification accuracy (%) 91.3 68.5 76.2 83.9 Re-identification false alarm rate (%) 4.2 23.7 16.4 8.9 Average target recovery time (s) 156 356 248 195 Energy efficiency (search area / energy consumption) 1.68 0.95 1.22 1.37 System completion rate - urban environment (%) 92.5 45.2 62.8 74.3 System completion rate - mountainous environment (%) 89.6 38.4 58.1 69.2 System Completion Rate - Open Terrain (%) 94.7 63.2 77.5 85.1 Average system completion rate (%) 92.3 48.9 66.1 76.2

[0417] Experimental results show that the proposed method significantly outperforms existing methods across all key metrics. In terms of target position prediction accuracy, this method achieved 82.7%, 43.5 percentage points higher than the traditional Kalman filter method. In terms of re-identification accuracy, this method achieved 91.3%, 22.8 percentage points higher than the single feature matching method. In terms of average target reacquisition time, this method only required 156 seconds, a 56.2% reduction compared to traditional methods.

[0418] Particularly noteworthy is the high completion rate achieved by this method across diverse environments, reaching 89.6% in the most challenging mountainous environment, demonstrating the system's strong adaptability. Furthermore, the method achieved an energy efficiency of 1.68, an improvement of 37.4% to 76.8% over traditional methods, significantly extending the drone's effective operating time.

[0419] Comprehensive evaluation results show that the non-cooperative target search method based on recognition-planning joint optimization proposed in this invention has significant advantages in the continuous monitoring of non-cooperative targets in complex environments, and provides efficient and reliable technical support for key areas such as security monitoring and border patrols.

Claims

1. A non-cooperative target search method based on joint optimization of recognition and planning, characterized in that: The steps include: Step 1: Terrain-aware multi-modal hybrid filtering prediction; Step 2: Object re-identification with fine-grained feature decoupling and dynamic fusion; Step 3: Adaptive search for joint optimization of identification and planning.

2. The non-cooperative target search method based on identification-planning joint optimization according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1-1: Terrain semantic segmentation and environment modeling; Step 1-1-1: Based on the high-resolution terrain imagery acquired by the drone, use the UNet++ semantic segmentation network to divide the environment into different categories: , , represents the total number of terrain categories, Indicates different terrain types; Represents a set of terrain categories; Step 1-1-2: Build a terrain adjacency graph , where the vertex Indicates area, edge Indicates the connection relationship between regions; Step 1-1-3: Extract key nodes: , Respectively represent the 1st to the key nodes, Indicates the total number of key nodes; Step 1-1-4: Define the traffic characteristic function for each type of terrain: ,in 、 Represent the minimum and maximum travel speeds for each type of terrain, represents the velocity attenuation coefficient; Step 1-2: Multimodal motion model construction; Step 1-2-1: Define three basic motion models; (1) Constant speed linear motion model CV: , express The state vector at time t, represents the constant velocity model state transfer matrix, express The state vector at time t, represents the constant velocity model process noise; (2) Constant acceleration motion model CA: , represents the state transfer matrix of the constant acceleration model, represents the process noise of the constant acceleration model; (3) Steering motion model CT: , represents the nonlinear state transfer function of the steering model, represents the steering model process noise; Step 1-2-2: Each motion model is: ； ； It is a nonlinear function including angular velocity parameters; in, is the time step; Steps 1-3: Terrain constraint model fusion; Step 1-3-1: Construct the state transfer function under terrain constraints: ； in is the terrain influencing factor, according to the terrain type Adjust state transfer; represents the process noise vector, represents the state transition matrix of the motion model, represents the terrain constraint state transfer function; Step 1-3-2: Position constraint; Make sure the predicted location is in a traversable area: ； in Represents the state vector Extract location, is the set of traversable areas, is the projection function; represents the position constraint function; Step 1-3-3: speed constraint; Adjust speed range based on terrain type: ； in Represents the state vector extraction speed, is the speed adjustment function; represents the velocity constraint function, The minimum travel speed function representing the terrain type, The maximum travel speed function representing the terrain type; Steps 1-4: Interactive multi-model spatiotemporal robust filtering; Step 1-4-1: Build an interacting multi-model IMM architecture, including filter combinations: , Filters representing constant speed, constant acceleration, and steering motion models respectively; Step 1-4-2: Define the model transition probability matrix: ,in Represents the model Transfer to Model probability; Step 1-4-3: Model probability update: ； in For the moment Time Model The probability of For the model The likelihood function of is the observation data; Represents the k-1 moment model The probability of Represents the model at time k-1 The probability of Indicates the total number of models; Step 1-4-4: state fusion prediction; ； in For the model Status prediction; Steps 1-5: Multipath hypothesis generation; Step 1-5-1: Consider multiple possible paths to the target and build a set of path hypotheses , Respectively represent items 1 to 2 The path hypothesis, represents the total number of path hypotheses; Step 1-5-2: Assume for each path , , calculate its probability: in Indicates consistency with historical trajectory, Indicates compatibility with the terrain; Represents historical state trajectory data; Step 1-5-3: Make predictions along each hypothetical path to generate possible locations; ； in is the prediction time window; represents the trajectory prediction function; Step 1-5-4: Assign probability weights to each position; ； in is the time decay factor; Step 1-5-5: Output location set: .

3. The non-cooperative target search method based on identification-planning joint optimization according to claim 2, characterized in that: The step 2 is specifically as follows: Step 2-1: Feature decoupling representation learning; Step 2-1-1: Design a three-way feature extraction network to extract features of different dimensions respectively; Appearance Feature Network : Use the ResNet50-IBN backbone network to extract the target visual appearance features; Motion Feature Network : Use the spatiotemporal graph convolutional network ST-GCN to extract target motion pattern features; Contextual Feature Network : Use Transformer encoder to extract contextual features of the interaction between the target and the environment; Step 2-1-2: Construct a multi-dimensional feature representation for the original target; : 256-dimensional appearance feature vector; : 128-dimensional motion feature vector; : 192-dimensional context feature vector; in, 、 、 Represent the image sequence, trajectory data and context data of the original target respectively; Step 2-2: candidate target detection and feature extraction; Step 2-2-1: Use the YOLOv5 object detector to detect candidate objects in the search area video: ; represents the target detection function, Indicates the search area video; Step 2-2-2: Extract multidimensional features for each candidate target: ； ； ； in, Indicates the The appearance feature vector of candidate targets, Indicates the The motion feature vectors of candidate targets, Indicates the The context feature vector of candidate targets, Indicates the A sequence of candidate target images, Indicates the The trajectory data of candidate targets, Indicates the contextual data of candidate targets; , Indicates the total number of candidate targets; Step 2-2-3: Construct candidate target feature set; ； Step 2-3: Feature reliability assessment under spatiotemporal conditions; Step 2-3-1: Design a feature reliability evaluation function to measure the reliability of features in each dimension based on time intervals and environmental changes: ； ； ； in, The target loss duration, is a measure of environmental change, 、 are all weight parameters; represents the reliability evaluation function of appearance features, represents the motion feature reliability evaluation function, represents the context feature reliability evaluation function; Step 2-3-2: Determine feature fusion weights; ； ； ； in, represents the appearance feature fusion weight, represents the motion feature fusion weight, represents the context feature fusion weight, 、 、 、 Respectively represent the reliability value of appearance feature, motion feature, context feature, and each dimension feature; Step 2-4: Adaptive feature fusion matching; Step 2-4-1: Multi-dimensional similarity calculation; ； ； ； in, represents the cosine similarity function, represents the appearance feature similarity function, represents the motion feature similarity function, represents the context feature similarity function; Step 2-4-2: Dynamic fusion similarity: ； Step 2-4-3: Position prior enhancement; The similarity is further adjusted based on the distance between the candidate target and the predicted position: ； in Represents the candidate target location and predicted position distance, is the position prior weight, is the distance attenuation parameter; represents the final similarity function; Step 2-5: Cross-domain consistency verification and decision-making; Step 2-5-1: Design an adaptive threshold function; ； in is the basic threshold, is the maximum increment, is the time coefficient; Step 2-5-2: Preliminary screening meets candidate targets; Step 2-5-3: Verify the cross-domain consistency of the initial screening results: (1) Multi-angle observation verification: obtain matching scores under different perspectives; (2) Temporal consistency verification: short-term tracking verifies the consistency of motion patterns; Step 2-5-4: Final decision; (1) When there are candidate targets that meet the conditions, select the one with the highest score: ; (2) Otherwise, it is judged as "target not found"; (3) Output re-identification results: , Represents the re-identification confidence.

4. The non-cooperative target search method based on identification-planning joint optimization according to claim 3, characterized in that: The step 3 is specifically as follows: Step 3-1: Search resource modeling and constraint definition; Step 3-1-1: UAV resource constraint modeling; (1) Energy constraints: ,in To search for energy needed for the mission, for available energy; (2) Time constraints: ,in is the total search time, is the maximum allowed time; Step 3-1-2: Define the UAV motion consumption model; (1) Energy consumption: ,in For distance, Consumption for hovering observation; represents the energy consumption coefficient; Respectively represent and target search points; (2) Time consumption: ,in is the average speed, is the observation time; Step 3-2: Identify a confidence-driven search strategy: Step 3-2-1: Based on the re-identification confidence , define three search modes: (1) High confidence mode ; (2) Medium confidence mode ; (3) Low confidence mode or there are no candidate targets; Step 3-2-2: Design different search parameters for each mode; (1) High confidence mode: , ; (2) Medium confidence mode: , ; (3) Low confidence mode: , ; in, represents the search radius, Indicates the single point observation time; Step 3-3: candidate point clustering and hierarchical representation; Step 3-3-1: Cluster the predicted locations using the density clustering algorithm DBSCAN according to the location aggregation: , T represents the clustering result set, Respectively represent the 1st to the clusters; Step 3-3-2: Calculate the comprehensive weight of each cluster: , ; Step 3-3-3: Build a hierarchical representation; (1) Hot spots are high-weight clusters: ; (2) Secondary areas, namely medium-weight clusters: ; (3) Low probability areas are low weight clusters: in , is the clustering weight threshold; Step 3-4: Identify-plan joint optimization objective function: Step 3-4-1: Define the basic search point utility function; ； in For the point The observation coverage value at represents weight; Step 3-4-2: Integrate re-identification feedback to enhance the utility function: ； in To identify the enhancement coefficient, is the candidate target position indicator function; represents the utility function of fusion re-identification feedback; Step 3-4-3: Define the search path cost function: ； in, Indicates the first Search points, Indicates the first Search points; Indicates the search path, Indicates the total number of search points in the search path; Step 3-4-4: Define the search path time function; ； Step 3-4-5: Construct the identification-planning joint optimization objective function: ； in , is the balance parameter; Steps 3-5: Hierarchical search strategy generation; Step 3-5-1: Adjust the search strategy based on recognition confidence: (1) High confidence mode: adopts a depth-first search strategy to prioritize exploring the target’s surroundings; (2) Medium confidence mode: adopts a balanced search strategy that takes into account both depth and breadth; (3) Low confidence mode: using a breadth-first search strategy to cover multiple possible areas; Step 3-5-2: Constructing a search optimization problem: Constraints: ; in, Indicates the maximum available energy, Indicates the maximum allowed search time; Step 3-5-3: Use the improved branch and bound algorithm to solve; (1) Initialization: Select the starting search area based on the recognition pattern; (2) Branching strategy: adjust branching tendency according to recognition confidence; (3) Pruning strategy: pruning when a path violates energy or time constraints; Step 3-5-4: Output the optimal search path.

5. The non-cooperative target search method based on identification-planning joint optimization according to claim 4, characterized in that: The different terrain types include roads, buildings, and open areas.

6. The non-cooperative target search method based on identification-planning joint optimization according to claim 5, characterized in that: The key nodes include intersections and building entrances.

7. An electronic device, characterized in that: include: processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A chip, characterized in that: include: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product comprises a computer storage medium storing a computer program, wherein the computer program comprises instructions executable by at least one processor, and when the instructions are executed by the at least one processor, the method according to any one of claims 1 to 6 is implemented.

Images