Intelligent monitoring method for low-altitude UAV patrol based on flexible reconfigurable model

Abstract

This paper discloses a method for intelligent monitoring of low-altitude UAV patrols based on a flexible and reconfigurable model. The method first collects and preprocesses multi-source heterogeneous data; then constructs a multi-level prior knowledge base; then designs a Transformer-based backbone network; and finally, a gated multi-task network head architecture. Finally, multi-level target recognition and inference are achieved. This method effectively addresses key issues during patrol monitoring, such as large variations in target appearance and unstable feature representation.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicles (UAVs), and in particular relates to an intelligent monitoring method for low-altitude UAV patrols based on a flexible reconfigurable model. Background Art

[0002] Amidst the booming low-altitude economy, drone patrol monitoring is becoming a key technology for infrastructure maintenance. During the inspection of infrastructure such as roads, bridges, and power lines, drones are required to perform identification tasks such as detection, location, classification, and condition assessment of various targets. During patrols, target scale, posture, and environmental conditions vary significantly, and are severely affected by complex weather and lighting conditions. For example, when a certain type of inspection drone is lowered from an altitude of 200 meters to 50 meters, combined with optical zoom, the pixel size of a typical target (such as a bridge crack or a pothole) changes from 8 pixels per side to 96 pixels, a 12-fold change in scale. Furthermore, the target's geometry, texture, and radiometric properties evolve nonlinearly with changes in observation distance, angle, and environmental conditions. These multiple sources of interference lead to insufficient accuracy and poor generalization of airborne detection and recognition.

[0003] Existing UAV patrol target recognition methods mainly face the following challenges:

[0004] 1. Large differences in target appearance: The existing fixed network structure lacks the fusion modeling of prior knowledge and is difficult to adapt to the dynamic changes of various complex factors during the inspection process.

[0005] 2. Limited stability of the training process: The feature coupling effect between multi-level labels limits the model convergence efficiency. The lack of a label completeness perception mechanism makes it difficult for the model to maintain efficient convergence in scenarios with dynamic label loss.

[0006] 3. Difficulty in airborne embedded deployment: Accuracy decreases and hardware adaptation fails during model deployment. Especially when deployed on domestic platforms, insufficient adaptation of static compression methods leads to loss of key features and competition for multi-task resources.

[0007] Therefore, there is an urgent need for an intelligent monitoring target recognition technology for UAV patrols that can integrate prior knowledge, improve training stability, and optimize lightweight deployment. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this paper provides an intelligent monitoring method for low-altitude UAV patrols based on a flexible and reconfigurable model. The method first collects and preprocesses multi-source heterogeneous data; then constructs a multi-level prior knowledge base; then designs a Transformer-based backbone network; and finally, a gated multi-task network head architecture. Finally, multi-level target recognition and inference are achieved. This method effectively addresses key issues during patrol monitoring, such as large variations in target appearance and unstable feature representation.

[0009] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0010] Step 1: Multi-source heterogeneous data collection and preprocessing;

[0011] Step 2: Construction of multi-level prior knowledge base;

[0012] Step 3: Transformer-based backbone network design;

[0013] Step 4: Gated multi-task network head architecture;

[0014] Step 5: Multi-level object recognition and reasoning.

[0015] Preferably, the step 1 is specifically:

[0016] Step 1-1: Acquire multi-view, multi-scale, and multi-spectral image data collected by the UAV-mounted optoelectronic pod, and simultaneously record the corresponding flight parameters, meteorological conditions, and mission parameters;

[0017] Step 1-2: Perform image geometric correction based on flight parameters and camera parameters to eliminate geometric deformation caused by attitude changes and lens distortion;

[0018] Steps 1-3: Perform multispectral image registration and enhancement processing to improve image quality and target visibility.

[0019] Preferably, the flight parameters include altitude, speed, and attitude angle.

[0020] Preferably, the meteorological conditions include visibility, cloud cover, and temperature.

[0021] Preferably, the task parameters include inspection mode, distance, and angle.

[0022] Preferably, the step 2 is specifically as follows:

[0023] Constructing a four-layer three-dimensional prior knowledge base , respectively:

[0024] Physical layer prior ,include:

[0025] Target Material Property Mapping Table , including the reflectivity, thermal conductivity and temperature radiation characteristics of road, bridge and guardrail infrastructure materials; Respectively represent material type, material reflectivity, thermal conductivity, and temperature radiation characteristic coefficient;

[0026] Target spectral feature library , record the spectral response of different infrastructures at various wavelengths; 、 、 represent wavelength, target object ID, and spectral response function respectively;

[0027] Target-environment interaction model , describing the interaction effect between the target and the environment; They represent target type identifier, environmental factor, and interaction coefficient respectively;

[0028] Geometric layer prior ,include:

[0029] Target 3D structure template library , including parametric 3D geometric models of bridges, tunnels, and road infrastructure; Represent target categories and geometric model parameters respectively;

[0030] View-appearance mapping function , mapping the three-dimensional model into two-dimensional feature representation under different viewpoints; Represent the three-dimensional geometric model, pitch angle, and azimuth angle respectively; represents the two-dimensional projected eigenvector;

[0031] scale-detail relationship function , describes the set of features that can be identified at different distances; Represented at different observation distances The identifiable characteristic function below;

[0032] Proportion Constraint Rule Set , including the proportional relationship constraints between the target components; Respectively represent different geometric proportion constraint rules;

[0033] Semantic layer prior ,include:

[0034] Hierarchical Category Ontology ,construct a classification hierarchy system of road infrastructure objectives; They represent the semantic category set, hierarchical relationship matrix, and semantic relationship set respectively;

[0035] Attribute-category conditional probability table , describes the probability of an attribute appearing in a category; 、 Represent the attribute set and detection label set respectively; Represent the i-th attribute and j-th detection label respectively;

[0036] Functional feature knowledge graph , encoding the functional properties and relationships of the target; Respectively represent the knowledge graph entity set and the knowledge graph relationship set;

[0037] Contextual Association Rules , describes the co-occurrence probability of target categories; represent the co-occurrence probabilities of the i-th target category, the j-th target category, the i-th target category, and the j-th target category, respectively;

[0038] Task-level priors ,include:

[0039] Inspection Mode-Target Priority Mapping , describes the priority of targets under different inspection modes; Respectively represent the target type set and inspection mode set, They represent the i-th target type and the j-th inspection mode respectively;

[0040] Task-related decision rule base , They represent different sets of decision rules such as target priority sorting rules, resource allocation rules, and response strategy rules;

[0041] Attention Focus Guide , Respectively represent The coordinates of the region, Regional weights;

[0042] Target importance scoring function , calculate the importance based on target type, environmental conditions and time; Respectively represent the target type, current environmental conditions, and time information; Represents the field of real numbers.

[0043] Preferably, the step 3 is specifically as follows:

[0044] Step 3-1: Design three-way input of the backbone network;

[0045] Design a Transformer-based backbone network with three input branches to process image features respectively , prior knowledge and current parameters :

[0046]

[0047]

[0048]

[0049] in, is the input image, is multi-level prior knowledge, is the current inspection parameter; 、 and They are image feature encoding branch, prior knowledge encoding branch and current parameter encoding branch respectively;

[0050] Step 3-2: encoding structure of each branch;

[0051] The image feature encoding branch uses the CNN+Transformer structure to extract multi-scale spatial features;

[0052] The prior knowledge encoding branch converts structured knowledge into vector representation and encodes it through multiple layers of Transformer;

[0053] The current parameter encoding branch embeds various parameters into a unified feature space and establishes the association between parameters through the self-attention mechanism;

[0054] Step 3-3: Three-way feature fusion;

[0055]

[0056]

[0057]

[0058] in, It is a cross-attention mechanism to achieve interaction between different features; For the final feature fusion module, adaptive feature integration is achieved through gating mechanism and residual connection; It is the result of three-way feature fusion;

[0059] The calculation process of the cross attention mechanism is:

[0060]

[0061] in, To query features, is the key feature, is the value feature, is the square root of the feature dimension, represents the normalized exponential function, represents transpose;

[0062] Step 3-4: Feature enhancement and optimization:

[0063]

[0064] in, is a feedforward neural network, is the layer normalization operation; It is an enhanced feature representation that integrates image features, prior knowledge and current parameters.

[0065] Preferably, the step 4 is specifically as follows:

[0066] Step 4-1: Design of flexible gate control structure module;

[0067] The flexible door control structure module includes three components:

[0068] (1) Task parameter perception unit: collects and processes task-related parameters, including detection distance , pitch angle , azimuth , target characteristic scale ;

[0069] (2) Dynamic threshold generator based on the principle of “from far to near”: The activation threshold of each module is dynamically calculated according to the detection distance task parameters; the threshold calculation follows the principle of “from far to near”:

[0070]

[0071] in, For the The activation threshold of each module, To detect distance, 、 and is a learnable parameter;

[0072] (3) Gating activation decision unit: Generates gating signals based on the calculated threshold and feature correlation to control the activation state of each functional module:

[0073]

[0074] in, is the gate signal vector, is the Sigmoid function, is the learnable weight matrix, is the task parameter vector;

[0075] Gated activation determination:

[0076]

[0077] in, Indicates the The activation status of each module, Indicates the The gate signal value of each module, when When the module is activated ,when The module remains closed when ;

[0078] Step 4-2: Feature extraction module group;

[0079] The gate structure is followed by multiple feature extraction modules, including:

[0080] (1) Small target enhancement module: used to extract and enhance small-scale target features;

[0081] (2) Edge detail preservation module: retains target edge and detail information to improve segmentation accuracy;

[0082] (3) Occlusion processing module: handles partially occluded targets through contextual reasoning;

[0083] (4) Posture-invariant feature extraction module: extracts features that are insensitive to changes in the target posture;

[0084] Each module is selectively activated according to the gating signal, and the output features are adaptively fused:

[0085]

[0086] in, For the The feature output of each module, is the adaptive fusion weight, is in the activated state;

[0087] Step 4-3: Multi-task detection head design;

[0088] Based on the fusion features, four types of task-specific detection heads are designed:

[0089] (1) Object detection head: realizes object localization and bounding box regression;

[0090] Output: target location and bounding box ; Represents the x-coordinate of the target center point, the y-coordinate of the target center point, the width of the bounding box, the height of the bounding box, and the target confidence score respectively;

[0091] (2) Semantic segmentation head: realizes pixel-level semantic annotation;

[0092] Output: pixel-level semantic map ; Indicates a point Category label encoding;

[0093] (3) Target recognition head: realizes fine classification of targets;

[0094] Output: object category and its probability ; Represent target ID and classification confidence probability respectively; Indicates the object category;

[0095] (4) Pose estimation head: realizes the three-dimensional pose estimation of the target;

[0096] Output: target 3D pose parameters ; Represent the rotation parameters and translation parameters respectively;

[0097] Step 4-4: Integration of multi-task results;

[0098] Integrate the outputs of each detection head into a unified multi-task monitoring result set:

[0099]

[0100] in, is the final multi-task monitoring result set, is the bounding box parameter, For category information, is the segmentation result, is the pose parameter.

[0101] Preferably, the step 5 is specifically as follows:

[0102] Step 5-1: Results integration and optimization;

[0103]

[0104] in, Optimize the function for the result and use prior knowledge to filter, complete and correct the preliminary results;

[0105] Step 5-2: Design timing consistency constraints to improve the stability of target tracking and recognition;

[0106]

[0107] in, Indicates time and The target similarity, is the feature distance, Respectively indicate time The target feature vector, time The target feature vector of and All are adjustment parameters;

[0108] Step 5-3: Anomaly detection and early warning;

[0109] Abnormal pattern recognition based on historical data;

[0110] Multi-level early warning mechanism;

[0111] Early warning information push and visualization.

[0112] The beneficial effects of the present invention are as follows:

[0113] This method effectively addresses key issues such as large variations in target appearance and unstable feature representation during patrol monitoring, by constructing a multi-level prior knowledge base, a three-input Transformer-based backbone network, and a flexible, gated, multi-task network head architecture based on the "far-to-near" principle. Experimental results demonstrate that the proposed method significantly improves recognition accuracy and generalization capabilities, providing reliable technical support for intelligent patrol monitoring using drones. This method is suitable for patrol monitoring of various types of infrastructure, providing key technical support for the development of the low-altitude economy and possessing significant application value and promotional prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0115] The present invention will be further described below with reference to the accompanying drawings and examples.

[0116] like Figure 1 As shown in the figure, a low-altitude UAV patrol intelligent monitoring method based on a flexible reconfigurable model is described as follows:

[0117] Step 1: Multi-source heterogeneous data collection and preprocessing;

[0118] Input: Multi-view, multi-scale, and multi-spectral image data collected by the UAV-mounted optoelectronic pod;

[0119] Step 1-1: Acquire multi-view, multi-scale, and multi-spectral image data collected by the UAV-mounted optoelectronic pod, and simultaneously record the corresponding flight parameters, meteorological conditions, and mission parameters;

[0120] Step 1-2: Perform image geometric correction based on flight parameters and camera parameters to eliminate geometric deformation caused by attitude changes and lens distortion;

[0121] Steps 1-3: Perform multispectral image registration and enhancement processing to improve image quality and target visibility;

[0122] Output: rectified image data and related environment and task parameters;

[0123] Step 2: Construction of multi-level prior knowledge base;

[0124] Input: pre-processed image data, environment task parameters, and domain expert knowledge;

[0125] Constructing a four-layer three-dimensional prior knowledge base , respectively:

[0126] Physical layer prior ,include:

[0127] Target Material Property Mapping Table , including the reflectivity, thermal conductivity and temperature radiation characteristics of infrastructure materials such as roads, bridges and guardrails;

[0128] Target spectral feature library , record the spectral response of different infrastructures at various wavelengths;

[0129] Target-environment interaction model , describing the interaction effect between the target and the environment;

[0130] Geometric layer prior ,include:

[0131] Target 3D structure template library , including parametric 3D geometric models of infrastructure such as bridges, tunnels, and pavements;

[0132] View-appearance mapping function , mapping the three-dimensional model into two-dimensional feature representation under different viewpoints;

[0133] scale-detail relationship function , describes the set of features that can be identified at different observation distances;

[0134] Proportion Constraint Rule Set , including the proportional relationship constraints between the target components;

[0135] Semantic layer prior ,include:

[0136] Hierarchical Category Ontology ,construct a classification hierarchy system for infrastructure targets such as roads;

[0137] Attribute-category conditional probability table , describes the probability of an attribute appearing in a category;

[0138] Functional feature knowledge graph , encoding the functional properties and relationships of the target;

[0139] Contextual Association Rules , describes the co-occurrence probability of target categories;

[0140] Task-level priors ,include:

[0141] Inspection Mode-Target Priority Mapping , describes the priority of targets under different inspection modes;

[0142] Task-related decision rule base ;

[0143] Attention Focus Guide ;

[0144] Target importance scoring function , calculate the importance based on target type, environmental conditions and time;

[0145] Output: structured multi-level prior knowledge base;

[0146] Step 3: Transformer-based backbone network design;

[0147] Input: multi-level prior knowledge base, pre-processed image data, current inspection parameters;

[0148] Step 3-1: Design three-way input of the backbone network;

[0149] Design a Transformer-based backbone network with three input branches to process image features respectively , prior knowledge and current parameters :

[0150]

[0151]

[0152]

[0153] Step 3-2: encoding structure of each branch;

[0154] The image feature encoding branch uses the CNN+Transformer structure to extract multi-scale spatial features;

[0155] The prior knowledge encoding branch converts structured knowledge into vector representation and encodes it through multiple layers of Transformer;

[0156] The current parameter encoding branch embeds various parameters into a unified feature space and establishes the association between parameters through the self-attention mechanism;

[0157] Step 3-3: Three-way feature fusion;

[0158]

[0159]

[0160]

[0161] The calculation process of the cross attention mechanism is:

[0162]

[0163] Step 3-4: Feature enhancement and optimization:

[0164]

[0165] Output: Enhanced feature representation that combines image features, prior knowledge, and current parameters;

[0166] Step 4: Gated multi-task network head architecture;

[0167] Input: enhanced feature representation of the backbone network output, current inspection conditions, and computing resource constraints;

[0168] Step 4-1: Design of flexible gate control structure module;

[0169] The flexible door control structure module includes three components:

[0170] (1) Task parameter perception unit: collects and processes task-related parameters, including detection distance , pitch angle , azimuth , target characteristic scale ;

[0171] (2) Dynamic threshold generator based on the principle of “from far to near”: The activation threshold of each module is dynamically calculated according to the detection distance task parameters; the threshold calculation follows the principle of “from far to near”:

[0172]

[0173] (3) Gating activation decision unit: Generates gating signals based on the calculated threshold and feature correlation to control the activation state of each functional module:

[0174]

[0175] Gated activation determination:

[0176]

[0177] Step 4-2: Feature extraction module group;

[0178] The gate structure is followed by multiple feature extraction modules, including:

[0179] (1) Small target enhancement module: used to extract and enhance small-scale target features;

[0180] (2) Edge detail preservation module: retains target edge and detail information to improve segmentation accuracy;

[0181] (3) Occlusion processing module: handles partially occluded targets through contextual reasoning;

[0182] (4) Posture-invariant feature extraction module: extracts features that are insensitive to changes in the target posture;

[0183] Each module is selectively activated according to the gating signal, and the output features are adaptively fused:

[0184]

[0185] Step 4-3: Multi-task detection head design;

[0186] Based on the fusion features, four types of task-specific detection heads are designed:

[0187] (1) Object detection head: realizes object localization and bounding box regression;

[0188] Output: target location and bounding box ;

[0189] (2) Semantic segmentation head: realizes pixel-level semantic annotation;

[0190] Output: pixel-level semantic map ;

[0191] (3) Target recognition head: realizes fine classification of targets;

[0192] Output: object category and its probability ;

[0193] (4) Pose estimation head: realizes the three-dimensional pose estimation of the target;

[0194] Output: target 3D pose parameters ;

[0195] Step 4-4: Integration of multi-task results;

[0196] Integrate the outputs of each detection head into a unified multi-task monitoring result set:

[0197]

[0198] Output: comprehensive monitoring results after processing by the multi-task network head;

[0199] Step 5: Multi-level target recognition and reasoning;

[0200] Input: Results processed by the gated multi-task network head, knowledge-enhanced feature representation, and real-time image data;

[0201] Step 5-1: Results integration and optimization;

[0202]

[0203] in, Optimize the function for the result and use prior knowledge to filter, complete and correct the preliminary results;

[0204] Step 5-2: Design timing consistency constraints to improve the stability of target tracking and recognition;

[0205]

[0206] Step 5-3: Anomaly detection and early warning;

[0207] Abnormal pattern recognition based on historical data;

[0208] Multi-level early warning mechanism;

[0209] Early warning information push and visualization.

[0210] Output: target location, category, attributes, status and posture report.

[0211] Step 6: System integration and verification evaluation;

[0212] Input: target recognition results and performance evaluation requirements;

[0213] Step 6-1: Integrate the method of the present invention into the airborne processing unit and connect it with the optoelectronic pod, flight control system and other modules;

[0214] Step 6-2: Conduct system testing and evaluation in typical patrol scenarios, including recognition performance testing at different distances, angles, and weather conditions.

[0215] Step 6-3: Build a performance evaluation index system to comprehensively evaluate the following aspects:

[0216] •Recognition accuracy and recall;

[0217] •Real-time performance and resource usage;

[0218] • Ability to generalize across scenarios;

[0219] • Anomaly detection sensitivity and false alarm rate;

[0220] •System stability and reliability;

[0221] Output: Integrated system and performance report.

[0222] Experimental results:

[0223] Table 1

[0224] method Detection accuracy (AP) Classification accuracy (%) Status assessment accuracy (%) Inference time (ms) Memory usage (MB) Cross-scenario generalization YOLOv5 + ResNet 0.734 79.5 72.6 76 173 middle EfficientDet 0.762 81.8 76.2 83 145 middle Multimodal fusion network 0.805 85.4 81.3 128 226 middle Method of the present invention 0.936 89.7 86.4 82 185 high

[0225] The experiment compared the performance of four different drone inspection methods on multiple key performance indicators. The method of the present invention is significantly superior to the comparison methods in terms of detection accuracy and precision. Specifically, the detection accuracy (AP) of the method of the present invention reached 0.936, which is 27.5% higher than the 0.734 of the traditional YOLOv5+ResNet method and 22.8% higher than the 0.762 of EfficientDet. In terms of classification accuracy, the method of the present invention reached 89.7%, which is 4.3 percentage points higher than the 85.4% of the best comparison method, the multimodal fusion network. In terms of state assessment accuracy, the method of the present invention achieved an accuracy of 86.4%, which also surpassed all comparison methods.

[0226] In terms of computational efficiency, the proposed method demonstrates excellent real-time performance. Its inference time is 82ms, outperforming the 128ms achieved by a multimodal fusion network and comparable to the lightweight YOLOv5+ResNet approach (76ms). This demonstrates that the prior knowledge-guided architecture design effectively controls computational complexity while ensuring high accuracy. Its memory usage of 185MB is within a reasonable range, making it suitable for deployment in resource-constrained devices such as drones.

[0227] The innovation of the present invention is:

[0228] 1. Multi-level prior knowledge system construction method: An innovative four-layer prior knowledge base consisting of physical layer, geometric layer, semantic layer and task layer is constructed to achieve systematic encoding of the comprehensive characteristics of inspection targets, significantly improving the completeness and structure of knowledge representation.

[0229] 2. Transformer-based three-input backbone network: A three-input Transformer architecture is designed to process image features, prior knowledge, and inspection parameters. This achieves unified representation and fusion of multi-source heterogeneous information and solves feature alignment and interaction issues.

[0230] 3. Flexible gating mechanism based on the "from far to near" principle: A gating strategy is proposed to dynamically adjust the module activation threshold according to task parameters such as detection distance. This achieves precise allocation of computing resources and on-demand activation of modules, improving the system's adaptability in resource-constrained environments.

[0231] Application scenarios:

[0232] The present invention is applicable to various scenarios of UAV patrol and intelligent monitoring, including:

[0233] 1. Road infrastructure inspection: Identify road cracks, potholes, water damage and other defects;

[0234] 2. Bridge inspection: Identify abnormalities such as deformation, cracks, and rust in bridge structures;

[0235] 3. Traffic facility monitoring: Identify the intact status of traffic signs, guardrails, isolation belts and other facilities;

[0236] 4. Abnormal event detection: Identify abnormal events such as road obstacles, accumulated water, and accidents;

[0237] 5. Traffic flow monitoring: statistics of traffic parameters such as vehicle volume and vehicle type distribution.

Claims

1. A low-altitude UAV patrol intelligent monitoring method based on a flexible reconfigurable model, characterized in that: The steps include: Step 1: Multi-source heterogeneous data collection and preprocessing; Step 2: Build a multi-level prior knowledge base; Constructing a four-layer three-dimensional prior knowledge base They are: Physical layer prior include: Target material property mapping table; Target spectral feature library; target-environment interaction model; Geometric layer prior include: Target three-dimensional structure template library; View-appearance mapping function; scale-detail relationship function; Proportionality constraint rule set; Semantic layer prior include: Hierarchical category ontology; attribute-class conditional probability table; Functional feature knowledge graph; Context association rules; Task-level priors include: Inspection mode-target priority mapping; a database of task-related decision rules; Attention focus orientation map; Target importance scoring function; Step 3: Design a Transformer-based backbone network; Step 4: Gated multi-task network head architecture; Step 4-1: Design of flexible gate control structure module; The flexible door control structure module includes three components: (1) Task parameter perception unit: collects and processes task-related parameters, including detection distance d u , pitch angle θ, azimuth angle φ, target characteristic scale l; (2) Dynamic threshold generator based on the principle of from far to near: The activation threshold of each module is dynamically calculated according to the detection distance task parameters; the threshold calculation follows the principle of from far to near: Among them, τ i is the activation threshold of the i-th module, d u is the detection distance, α i , β i and γ i is a learnable parameter; (3) Gating activation decision unit: Generates gating signals based on the calculated threshold and feature correlation to control the activation state of each functional module: G=σ(W G ·[F enhanced ;T param ]); Where G=[g1,g2,...,g n ] is the gate signal vector, σ(.) is the Sigmoid function, W G is the learnable weight matrix, T param is the task parameter vector; g i Represents the gate signal value of the i-th module; Gated activation determination: Among them, A i Indicates the activation state of the i-th module. When g i >τ i When the module is activated, that is, A i =1, when g i ≤τ i When the module remains closed i =0; Step 4-2: Feature extraction module group; The gate structure is followed by multiple feature extraction modules, including: (1) Small target enhancement module: used to extract and enhance small-scale target features; (2) Edge detail preservation module: retains target edge and detail information to improve segmentation accuracy; (3) Occlusion processing module: handles partially occluded objects through contextual reasoning; (4) Posture-invariant feature extraction module: extracts features that are insensitive to changes in the target posture; Step 4-3: Multi-task detection head design; Based on the fusion features, four types of task-specific detection heads are designed: (1) Object detection head: realizes object localization and bounding box regression; (2) Semantic segmentation head: realizes pixel-level semantic annotation; (3) Target recognition head: realizes fine classification of targets; (4) Pose estimation head: realizes the three-dimensional pose estimation of the target; Step 4-4: Integration of multi-task results; Integrate the outputs of each detection head into a unified multi-task monitoring result set; Step 5: Multi-level object recognition and reasoning.

2. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1-1: Acquire multi-view, multi-scale, and multi-spectral image data collected by the UAV-mounted optoelectronic pod, and simultaneously record the corresponding flight parameters, meteorological conditions, and mission parameters; Step 1-2: Perform image geometric correction based on flight parameters and camera parameters to eliminate geometric deformation caused by attitude changes and lens distortion; Steps 1-3: Perform multispectral image registration and enhancement processing to improve image quality and target visibility.

3. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 2 is characterized in that: The flight parameters include altitude, speed, and attitude angle.

4. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 2 is characterized in that: The meteorological conditions include visibility, cloud cover, and temperature.

5. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 2 is characterized in that: The task parameters include inspection mode, distance, and angle.

6. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 2 is characterized in that: The step 2 is specifically as follows: Constructing a four-layer three-dimensional prior knowledge base They are: Physical layer prior include: Target material property mapping table M = {(m i ,r v ,c i ,t v )}, including the reflectivity, thermal conductivity and temperature radiation characteristics of road, bridge and guardrail infrastructure materials; m i 、r v 、c i , t v Respectively represent material type, material reflectivity, thermal conductivity, and temperature radiation characteristic coefficient; Target spectrum feature library S = {(o j ,s j (λ))}, recording the spectral response of different infrastructures at various wavelengths; λ, o j 、s j (.) represent wavelength, target object ID, and spectral response function respectively; Target environment interaction model E={(o k ,e k ,i k )}, describing the interaction effect between the target and the environment; o k 、e k 、i k They represent target type identifier, environmental factor, and interaction coefficient respectively; Geometric layer prior include: Target 3D structure template library T 3D ={(c l ,G l )}, including parametric 3D geometric models of bridges, tunnels, and road infrastructure; c l , G l Represent target categories and geometric model parameters respectively; View-appearance mapping function V(G v ,θ,φ)→I 2D , mapping the three-dimensional model into two-dimensional feature representation at different perspectives; G v , θ, φ represent the three-dimensional geometric model, pitch angle, and azimuth angle respectively; I 2D represents the two-dimensional projected eigenvector; Scale-detail relationship function D(d)={(f1(d),f2(d),...,f n (d))}, describing the set of features that can be identified at different distances; f1(d),f2(d),...,f n (d) represent the identifiable characteristic functions at different observation distances d; Proportion constraint rule set R prop ={r1,r2,...,r m }, including the proportional relationship constraints between the target components; r1, r2, ..., r m Respectively represent different geometric proportion constraint rules; Semantic layer prior include: Hierarchical category ontology O = (C,H,), constructs a classification hierarchy system for road infrastructure targets; C,H,represents the semantic category set, hierarchical relationship matrix, and semantic relationship set, respectively; Attribute-category conditional probability table P(A|L)={p(a i |l j )}, describing the probability of an attribute appearing in a category; A and L represent the attribute set and the detection label set respectively; a i 、l j Represent the i-th attribute and j-th detection label respectively; Functional feature knowledge graph KG=(E kg ,R kg ), encoding the functional properties and relationships of the target; E kg 、R kg Respectively represent the knowledge graph entity set and the knowledge graph relationship set; Context association rule Ctx={(ca i ,ca j ,p ij )}, describing the co-occurrence probability of target categories; ca i 、ca j 、p ij represent the co-occurrence probabilities of the i-th target category, the j-th target category, the i-th target category, and the j-th target category, respectively; Task-level priors include: Inspection mode-target priority mapping P(d v |m)={p(d i |m j )}, describing the priority of targets in different inspection modes; d v , m represent the target type set and inspection mode set respectively, d i 、m j They represent the i-th target type and the j-th inspection mode respectively; Task-related decision rule base Π={π1,π2,...,π k },π1,π2,...,π k They represent different sets of decision rules such as target priority sorting rules, resource allocation rules, and response strategy rules; Attention Focus Guide A f ={(r i ,w i )},r i 、w i Represent the coordinates of the i-th region and the weight of the i-th region respectively; Target importance scoring function Importance is calculated based on target type, environmental conditions, and time; ob, c, and t represent target type, current environmental conditions, and time information, respectively; Represents the field of real numbers.

7. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 6 is characterized in that: The step 3 is specifically as follows: Step 3-1: Design three-way input of the backbone network; Design a Transformer-based backbone network with three input branches to process image features F respectively. img , prior knowledge F know and the current parameter F param : F img =ImageEncoder(I); F know =KnowledgeEncoder(K5); F param =ParameterEncoder(P); Where I is the input image, K5 is the multi-level prior knowledge, and P is the current inspection parameter; ImageEncoder(.), KnowledgeEncoder(.), and ParameterEncoder(.) are the image feature encoding branch, the prior knowledge encoding branch, and the current parameter encoding branch, respectively; Step 3-2: encoding structure of each branch; The image feature encoding branch uses the CNN+Transformer structure to extract multi-scale spatial features; The prior knowledge encoding branch converts structured knowledge into vector representation and encodes it through multiple layers of Transformer; The current parameter encoding branch embeds various parameters into a unified feature space and establishes the association between parameters through the self-attention mechanism; Step 3-3: Three-way feature fusion; F cross1 =CrossAttention(F img ,F know ); F cross2 =CrossAttention(F cross1 ,F param ); F final =FusionModule(F cross1 ,F cross2 ); Among them, CrossAttention(.) is the cross attention mechanism to realize the interaction between different features; FusionModule(.) is the final feature fusion module, which realizes adaptive feature integration through gating mechanism and residual connection; F final It is the result of three-way feature fusion; The calculation process of the cross attention mechanism is: Among them, Q is the query feature, K6 is the key feature, V is the value feature, d k is the square root of the feature dimension, softmax(.) represents the normalized exponential function, and T represents transpose; Step 3-4: Feature enhancement and optimization: F enhanced =LayerNorm(F final +FFN(F final )); Among them, FFN(.) is a feedforward neural network, LayerNorm(.) is a layer normalization operation; F enhanced It is an enhanced feature representation that integrates image features, prior knowledge and current parameters.

8. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 7 is characterized in that: The step 4 is specifically as follows: Step 4-1: Design of flexible gate control structure module; Step 4-2: Feature extraction module group; Each module is selectively activated according to the gating signal, and the output features are adaptively fused: F fused =∑ i w ui ·F i / ∑ i w ui ·A i ; Among them, F i is the feature output of the i-th module, w ui is the adaptive fusion weight, A i is in the activated state; Step 4-3: Multi-task detection head design; Based on the fusion features, four types of task-specific detection heads are designed: (1) Object detection head: realizes object localization and bounding box regression; Output: target location and bounding box B = {(x i ,y i ,wd i ,h i ,s i )};x i 、y i 、wd i 、h i 、s i Represents the x-coordinate of the target center point, the y-coordinate of the target center point, the width of the bounding box, the height of the bounding box, and the target confidence score respectively; (2) Semantic segmentation head: realizes pixel-level semantic annotation; Output: pixel-level semantic map S = {(x i ,y i ,sc i )};sc i Represents a point (x i ,y i )’s category label encoding; (3) Target recognition head: realizes fine classification of targets; Output: object category and its probability C = {(id i ,obc i ,p i )};id i 、p i Respectively represent the target ID and classification confidence probability; obc i Indicates the object category; (4) Pose estimation head: realizes the three-dimensional pose estimation of the target; Output: Target 3D pose parameter P = {(id i ,ro i ,t vi )};ro i , t vi Represent the rotation parameters and translation parameters respectively; Step 4-4: Integration of multi-task results; Integrate the outputs of each detection head into a unified multi-task monitoring result set: R={(id i ,B i ,C i ,S i ,P i )}; Among them, R is the final multi-task monitoring result set, B i is the bounding box parameter, C i is the category information, S i is the segmentation result, P i is the pose parameter.

9. The method for intelligent monitoring of low-altitude UAV patrol based on a flexible reconfigurable model according to claim 8 is characterized in that: The step 5 is specifically as follows: Step 5-1: Results integration and optimization; R opt =OptimizeResults(R,F enhanced ,K); Among them, OptimizeResults(.) is the result optimization function, which uses prior knowledge to filter, complete and correct the preliminary results; Step 5-2: Design timing consistency constraints to improve the stability of target tracking and recognition; Among them, S(t i ,t j ) represents the time t i and t j The target similarity, d(f i ,f j ) is the characteristic distance, f i ,f j Respectively represent the time t i The target feature vector, time t j The target feature vector, σ v and τ v All are adjustment parameters; Step 5-3: Anomaly detection and early warning; Abnormal pattern recognition based on historical data; Multi-level early warning mechanism; Early warning information push and visualization.

Images