Unmanned aerial vehicle countering method and system based on image recognition

Abstract

The invention relates to the technical field of unmanned aerial vehicle countering, and discloses an unmanned aerial vehicle countering method and system based on image recognition. The method comprises the steps of collecting multi-source image data of a real-time monitoring airspace, and performing preprocessing to generate a standardized image stream; dividing spatial domain dynamic grids for the image flow, determining monitoring sub-regions and establishing a spatial association relationship; through cross-regional collaborative analysis, identifying a motion track and a cluster mode of the unmanned aerial vehicle, and generating a preliminary threat signal set; constructing an adaptive threat evaluation model by combining the standardized image stream and the preliminary threat signal set, and optimizing a model discrimination threshold by using a collaborative analysis result; inputting real-time multi-source image data, identifying an abnormal flight behavior, marking a violation target, and feeding back the violation target to the airspace management database; and outputting a deep query database according to the model, analyzing the threat level and the intrusion intention, and generating a threat traceability report. The method can effectively cope with a complex airspace environment, and improves the recognition and countering efficiency of the unmanned aerial vehicle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle countermeasures, in particular to an unmanned aerial vehicle countermeasure method and system based on image recognition. BACKGROUND

[0002] Currently, the phenomenon of illegal flight of unmanned aerial vehicles occurs frequently, which not only may interfere with the normal operation of civil aviation flights, but also may pose a threat to important infrastructure and the safety of large event sites. To address these problems, various unmanned aerial vehicle countermeasures have emerged, among which image recognition-based countermeasures have become one of the key research and application directions due to their non-contact and long-distance monitoring characteristics.

[0003] However, the existing image recognition-based unmanned aerial vehicle countermeasures still have many shortcomings in practical application. In the image data acquisition and processing link, most methods rely only on a single type of image acquisition device, which is difficult to cope with complex and variable airspace environments. For example, in scenes such as sudden changes in light, bad weather (such as rain, fog, and sand), and complex backgrounds (such as urban building groups and dense forests), the quality of the collected image data is poor, and problems such as noise interference and target outline blurring easily occur, which affects the subsequent accuracy of unmanned aerial vehicle target recognition. At the same time, the pre-processing of the collected image data is relatively simple, usually only basic noise reduction and grayscale processing, and the image data is not standardized and optimized according to different airspace environment characteristics, resulting in insufficient compatibility and reliability of the generated image stream in the subsequent analysis process.

[0004] In terms of airspace monitoring and target analysis, most existing methods use a whole airspace monitoring mode, which does not divide the monitored airspace reasonably and is difficult to achieve precise tracking and dynamic analysis of unmanned aerial vehicle targets in different regions. When the monitoring airspace range is large or multiple unmanned aerial vehicle targets are active at the same time, target missed detection, false detection, and motion trajectory tracking confusion may occur, especially in the face of unmanned aerial vehicle cluster flight scenarios, the cluster pattern characteristics cannot be effectively identified, and potential threats cannot be predicted in advance. In addition, there is a lack of effective spatial correlation mechanism between different monitoring regions, which cannot realize cross-regional collaborative analysis, resulting in insufficient overall motion situation grasp of unmanned aerial vehicle targets and difficulty in forming a comprehensive and coherent monitoring data chain.

[0005] In the threat assessment and feedback link, the threat assessment model used by the existing method is mostly a fixed threshold model, which cannot be adaptively adjusted according to real-time spatial monitoring data and changes in the behavior of unmanned aerial vehicle targets. This fixed threshold model greatly reduces the accuracy of the assessment when facing new unmanned aerial vehicle models, complex flight trajectories, and changing threat scenarios, and is prone to misjudgment or omission of threat levels. At the same time, the linkage mechanism between the monitoring results and the airspace management database is imperfect, and the information of the identified illegal flight targets cannot be timely and accurately fed back to the database, nor can it be deeply queried and analyzed through the database, so that the potential threat level and invasion intention cannot be quickly analyzed, resulting in a lack of effective data support for the development and implementation of subsequent countermeasures, affecting the efficiency and effectiveness of the countermeasures.

[0006] In the threat tracing aspect, the existing method can only record the identified illegal flight targets, and cannot integrate multi-dimensional monitoring data (such as flight time, flight trajectory, model characteristics, and flight area of the unmanned aerial vehicle target) for comprehensive analysis, making it difficult to generate a complete threat tracing report. This makes it difficult to quickly trace the source of the intrusion, analyze the intrusion path and intention after a drone intrusion event, which is not conducive to the subsequent responsibility identification and optimization of preventive measures, further reducing the overall and systematic nature of the drone countermeasures. SUMMARY

[0007] The purpose of the present application is to provide a method and system for counteracting unmanned aerial vehicles based on image recognition to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides a method for counteracting unmanned aerial vehicles based on image recognition, which comprises:

[0009] Collecting multi-source image data of the real-time monitoring airspace, pre-processing the multi-source image data, and generating a standardized image stream;

[0010] Dividing the standardized image stream into a dynamic grid of airspace, determining a plurality of monitoring sub-regions, and establishing a spatial correlation relationship between the monitoring sub-regions;

[0011] Based on the spatial correlation relationship, the standardized image stream is analyzed in a cross-regional collaborative manner to identify the motion trajectory and cluster pattern of the unmanned aerial vehicle target, and a preliminary threat signal set is generated;

[0012] An adaptive threat assessment model is constructed based on the standardized image stream and the preliminary threat signal set, and the discriminant threshold of the adaptive threat assessment model is optimized using the cross-regional collaborative analysis results;

[0013] Input the real-time collected multi-source image data into the adaptive threat assessment model, identify abnormal flight behaviors, mark the illegal flight targets, and feed back the identification results to the airspace management database;

[0014] Based on the output results of the adaptive threat assessment model, perform a deep query on the airspace management database, analyze the potential threat level and invasion intention, and generate a complete threat tracing report.

[0015] Preferably, the multi-source image data collected in real-time monitoring of the airspace includes:

[0016] A plurality of image acquisition devices deployed in the monitored airspace synchronously acquire visible light image and infrared image data;

[0017] The acquired visible light image and infrared image data are time-stamped and spatially registered;

[0018] The registered multi-modal image data is integrated into a time-series image sequence to generate a standardized image stream.

[0019] Preferably, the airspace dynamic grid division of the standardized image stream includes:

[0020] Extracting airspace geographical features and real-time meteorological parameters from the standardized image stream;

[0021] Adjusting the grid division granularity dynamically according to the airspace geographical features and real-time meteorological parameters;

[0022] Assigning a unique region identifier to each monitoring sub-region and recording the spatial coordinate range of each monitoring sub-region.

[0023] Preferably, the cross-regional collaborative analysis of the standardized image stream based on the spatial correlation relationship includes:

[0024] Establishing a spatio-temporal correlation matrix between monitoring sub-regions to analyze the cross-regional movement patterns of UAV targets;

[0025] Identifying the formation pattern and cooperative flight characteristics of the UAV cluster;

[0026] Generating a preliminary threat signal set according to the motion trajectory characteristics, and assigning a confidence weight to each threat signal.

[0027] Preferably, the adaptive threat assessment model includes:

[0028] Extracting the shape features and behavior features of the UAV target from the standardized image stream;

[0029] Training a deep learning classifier using the preliminary threat signal set to establish a discrimination benchmark for normal flight patterns and abnormal flight patterns;

[0030] adjusting sensitivity parameters of the adaptive threat assessment model according to the cross-region collaborative analysis result.

[0031] Preferably, the identifying abnormal flight behavior comprises:

[0032] comparing real-time image data with normal flight patterns in the adaptive threat assessment model;

[0033] detecting abnormal deviations in flight altitude, speed and heading;

[0034] classifying the identified abnormal flight targets by threat level and generating a list of irregular flight targets.

[0035] Preferably, the deep querying of the airspace management database comprises:

[0036] retrieving historical flight records and registration information according to the list of irregular flight targets;

[0037] associating multi-dimensional auxiliary information such as weather data and airspace control announcements;

[0038] analyzing the flight intent and potential risk level of the UAV target.

[0039] Preferably, the generating of the complete threat traceability report comprises:

[0040] integrating the list of irregular flight targets, historical flight records and multi-dimensional auxiliary information;

[0041] reconstructing the complete flight path and behavior pattern of the UAV target;

[0042] outputting a structured report containing threat level assessment and traceability analysis results.

[0043] Preferably, the method further comprises:

[0044] updating the discriminant rule of the adaptive threat assessment model according to the complete threat traceability report;

[0045] optimizing airspace dynamic grid division strategy and cross-region collaborative analysis algorithm;

[0046] adjusting parameter configuration and deployment scheme of image acquisition equipment.

[0047] Preferably, the present application further comprises an image recognition-based UAV countermeasure system, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of an image recognition-based UAV countermeasure method as described above.

[0048] Compared with the prior art, the present application has the beneficial effects that:

[0049] In terms of image data acquisition and processing, the method adopts a multi-source image data acquisition method, which can integrate the advantages of different types of image acquisition devices such as visible light cameras, infrared cameras, and high-definition video cameras, and still obtain relatively comprehensive and high-quality image data in complex airspace environments. Whether in insufficient lighting, bad weather, or complex background scenes, the fusion acquisition of multi-source image data can effectively compensate for the limitations of single devices, reduce noise interference and environmental factors affecting image quality. At the same time, the collected multi-source image data is subjected to targeted preprocessing, and according to different airspace environment characteristics and image data types, a differentiated standardization processing strategy is adopted to generate standardized image streams of unified format and high quality, greatly improving the compatibility and reliability of the subsequent image analysis link, and laying a solid foundation for accurate identification of unmanned aerial vehicle targets.

[0050] In the airspace monitoring and target analysis link, the airspace is divided into multiple monitoring sub-regions through dynamic grid division of the standardized image stream, which can achieve fine monitoring and management of unmanned aerial vehicle targets in different regions. This dynamic grid division method can flexibly adjust the size and number of sub-regions according to real-time airspace conditions such as unmanned aerial vehicle target distribution density and flight activity, avoiding the target tracking confusion problem in the overall airspace monitoring mode. At the same time, the spatial relationship between each monitoring sub-region is established, making cross-regional collaborative analysis possible, which can integrate the unmanned aerial vehicle target information monitored in different sub-regions to form a complete target motion data chain, effectively improving the coherence and accuracy of unmanned aerial vehicle target motion trajectory tracking. Especially in the face of unmanned aerial vehicle cluster flight scenarios, through cross-regional collaborative analysis, the cluster pattern characteristics can be clearly identified, including cluster size, flight formation, motion speed and direction, etc., which can provide strong support for early detection of abnormal cluster behavior and potential threats.

[0051] In the threat assessment link, the adaptive threat assessment model constructed by the method breaks through the limitations of traditional fixed threshold models and can adjust the model discrimination threshold in real time according to the standardized image stream and the preliminary threat signal set. In actual application, the model can continuously learn the unmanned aerial vehicle target features (such as model, flight speed, flight trajectory, flight area, etc.) and behavior patterns in real-time monitoring data, dynamically optimize the discrimination standard, so that it can still maintain high threat assessment accuracy when facing new unmanned aerial vehicle models, complex flight trajectories, and variable threat scenarios. Whether it is a single unmanned aerial vehicle violation or a cluster of abnormal activities, the adaptive threat assessment model can quickly and accurately identify abnormal flight behavior and accurately mark the violation target, effectively reducing the risk of threat level misjudgment or omission, and providing a reliable basis for the development of subsequent countermeasures.

[0052] In terms of feedback of monitoring results and linkage with airspace management, the method realizes efficient linkage between real-time monitoring data and airspace management database. The identified illegal flight target information is fed back to the airspace management database in a timely and accurate manner, which not only updates the airspace safety information in the database, but also integrates historical monitoring data and real-time data through deep query of the database for multi-dimensional analysis. Through deep query and analysis, the potential threat level can be quickly analyzed, such as the importance of the flight area of the unmanned aerial vehicle target, the degree of flight trajectory deviation, whether it carries suspicious articles, etc. to comprehensively judge the threat severity; at the same time, the invasion intention of the unmanned aerial vehicle can also be analyzed in depth, such as whether it is targeted to approach sensitive areas, whether there is a behavior of deliberately avoiding monitoring, etc. This efficient linkage mechanism and deep analysis capability enable the airspace management department to timely grasp the airspace safety dynamics, quickly develop and implement corresponding countermeasures, and greatly improve the response speed and efficiency of airspace safety management.

[0053] In terms of threat tracing, the method can generate a complete threat tracing report by integrating the output results of the adaptive threat assessment model and the deep query data of the airspace management database. The report not only contains basic information of the illegal flight target (such as model, flight time, flight trajectory, type of illegal behavior, etc.), but also presents key contents such as potential threat level analysis, invasion intention analysis, and possible invasion source speculation. The complete threat tracing report provides a comprehensive and accurate basis for subsequent responsibility identification, and also provides an important reference for the optimization and improvement of airspace safety prevention measures. Through analysis of the threat tracing report, the characteristics and rules of different types of unmanned aerial vehicle threat events can be summarized, and the monitoring equipment layout can be improved, the countermeasures can be optimized, and the threat assessment model can be updated, further improving the stability and effectiveness of the overall unmanned aerial vehicle countermeasure system, and providing continuous protection for long-term airspace safety management. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The working principle diagram of the unmanned aerial vehicle countermeasure method based on image recognition described in the present application;

[0055] Figure 2 The flowchart of multi-source image data acquisition and preprocessing;

[0056] Figure 3 The flowchart of cross-region collaborative analysis. DETAILED DESCRIPTION

[0057] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0058] Please refer to Figure 1 The present application provides an image recognition-based anti-drone method, which comprises: multiple image acquisition devices deployed in a monitoring airspace start synchronous work to collect multi-source image data of the monitoring airspace in real time, which usually includes visible light images and infrared images. The collected raw multi-source image data first enters a preprocessing link to perform timestamp alignment and spatial registration operations to eliminate the differences in time and space of different sensors, and finally generate a standardized image stream with unified format and time-space synchronization. Subsequently, the system performs airspace dynamic grid division on the generated standardized image stream, which dynamically determines the grid division granularity by considering the geographical features and real-time meteorological parameters of the current airspace, thereby dividing the entire monitoring airspace into multiple logically independent monitoring sub-regions and assigning a unique identifier to each sub-region, while establishing spatial correlation between the monitoring sub-regions to form an airspace topology.

[0059] After completing the airspace division, the system performs cross-region collaborative analysis on the standardized image stream based on the established spatial correlation. This analysis aims to identify the movement trajectory of the drone target between different monitoring sub-regions, analyze its movement pattern, and further identify whether there is a cluster flight behavior of multiple drones and its formation mode, and generate a preliminary threat signal set containing target position, speed, heading and confidence weight. Next, the system constructs an adaptive threat assessment model based on the standardized image stream and the preliminary threat signal set, which is trained by the drone shape and behavior features extracted from the image stream, and uses the global movement pattern information obtained by cross-region collaborative analysis to dynamically optimize the internal discriminant threshold and sensitivity parameters of the model, so that the model can adapt to changes in different airspace environments and threat situations.

[0060] After the model is put into real-time operation, the system continuously inputs newly acquired multi-source image data into the optimized adaptive threat assessment model. The model compares the real-time image data with learned normal flight patterns, identifies abnormal deviations in flight altitude, speed, heading, etc., thereby marking illegal flight targets, and feeding back the identification results of these targets to the airspace management database for storage in real time. Based on the model's output, the system further performs deep queries on the airspace management database, linking the illegal targets' historical flight records, registration information, and multi-dimensional auxiliary information such as meteorological and air traffic control notices, to analyze the potential threat level and intrusion intent. Finally, the system integrates all information, reconstructs the target's complete flight path, and generates a structured and complete threat attribution report containing threat level assessment and attribution analysis results, providing a comprehensive basis for subsequent countermeasure decisions. The entire system has self-learning capabilities, able to update the model's discrimination rules based on the generated threat attribution report, and optimize grid partitioning and collaborative analysis strategies to achieve continuous performance improvement.

[0061] Example 1: See Figure 2 Image acquisition devices deployed in different geographical locations within the monitoring area need to operate under strict synchronization. These devices typically include high-resolution visible light cameras and infrared thermal imagers with thermal sensing capabilities. They achieve simultaneous exposure and acquisition within a millisecond-level error range through precise time synchronization devices. This hardware-level synchronization mechanism provides a fundamental guarantee for subsequent multi-source data fusion. The acquisition devices transmit the raw visible light and infrared image data to the central processing server via a dedicated fiber optic network. Due to the different physical principles underlying visible light and infrared imaging and the differences in device installation locations, the raw data exhibits spatiotemporal inconsistencies and must undergo a rigorous preprocessing process to be transformed into a usable standardized image stream. The preprocessing stage first performs timestamp alignment. Although the acquisition devices have achieved hardware synchronization, slight delays during data transmission may cause timestamp drift. The system uses a sliding window-based time-series matching algorithm to recalibrate the data stream, ensuring that each set of visible light and infrared image data strictly corresponds to the same acquisition time. Spatial registration is the core step in preprocessing. This system adopts a registration method based on feature point matching. By automatically detecting corner points or edge information with significant features on visible light and infrared images, and using pre-calibrated camera parameters to calculate the transformation matrix between the two sensor coordinate systems, the infrared image pixels are finally accurately mapped to the visible light image coordinate system, achieving pixel-level alignment between images of different modalities.

[0062] The multi-modal image data completed registration is integrated into a time series of images, the system adds an index mark to each set of registered image data in chronological order, forming a standardized image stream with a unified space-time reference system, which contains both visible light detail information and infrared thermal radiation information. The generation of the standardized image stream enables the subsequent analysis module to process multi-source information in a consistent data format, creating conditions for accurate identification and tracking of unmanned aerial vehicle targets. The spatial dynamic grid division module then analyzes the standardized image stream. This division process is not static and unchanging, but dynamically adjusted according to the real-time perceived spatial environmental characteristics. The division module real-time parses geographical features such as building height and terrain changes from the image stream, and real-time weather parameters such as wind speed, wind direction, and visibility from the meteorological sensor.

[0063] The geographical features and weather parameters jointly determine the granularity strategy of grid division. In densely built urban environments or low-visibility smog weather, the system will automatically adopt smaller grid cell sizes. This fine-grained division helps improve the detection accuracy of low-altitude, slow-moving small targets. In open areas or sunny weather conditions, the system will use larger grid cells to optimize the allocation of computing resources and improve processing efficiency. The dynamic grid division algorithm calculates the optimal grid size based on the current environmental parameters, divides the entire monitoring airspace into several logically independent monitoring sub-regions, and each sub-region has a clear boundary range in three-dimensional space. The system assigns each generated monitoring sub-region a globally unique digital identifier, which remains unchanged throughout the system's life cycle, and accurately records the spatial coordinate range of each sub-region, including latitude and longitude boundaries and height intervals. The spatial coordinate information of the monitoring sub-region is stored in the system's topology database, and these coordinate data are used to establish the adjacency relationship and spatial association between sub-regions, providing topological structure support for subsequent cross-region target tracking. After the entire spatial dynamic grid division process is completed, a virtual grid airspace model is constructed in the system memory, which can dynamically update as environmental parameters change, providing an adaptive spatial framework for the subsequent analysis stage.

[0064] Example 2: see Figure 3, a model that can quantitatively monitor the space-time relationship between sub-regions is constructed, which exists in the form of a space-time correlation matrix. The construction of the matrix not only considers geographical adjacency, but also incorporates kinematic constraints. The row and column indices of the matrix correspond to each uniquely identified monitoring sub-region, and the matrix element values are assigned by calculating the shortest reachable path length between the sub-regional centroids and combining the typical UAV flight speed range, thereby reflecting the time probability and difficulty of target transfer between sub-regions. The space-time correlation matrix serves as the core data structure and computational basis for collaborative analysis, providing a theoretical basis for subsequent trajectory prediction and correlation. The system uses the matrix to analyze the possible cross-regional movement patterns of UAV targets. When an image processing thread in a monitoring sub-region identifies a potential UAV target, it immediately encapsulates the target's feature vector, time stamp, and location information into an event signal. The collaborative analysis engine receives the event signal and activates the processing units of multiple adjacent sub-regions connected in the space-time correlation matrix. Based on the target's current motion vector direction and speed, combined with the precomputed transfer probability in the correlation matrix, the engine predicts the most likely candidate sub-region sequence of the target in the next few time slices. This prediction allows the system to allocate computational resources to high-probability areas in advance, enabling proactive target search and matching. The key step in cross-regional collaborative analysis is the association and trajectory reconstruction of targets between consecutive frames. The system needs to ensure that the same physical target appearing continuously in the video stream is correctly identified as the same logical entity in different sub-regions at different times. This association process is achieved by comparing the multi-modal features of the target, including visual appearance features such as color distribution, texture features, and contour shape extracted from visible light images, thermal radiation features extracted from infrared images, and kinematic features such as instantaneous speed, acceleration, and heading angle. The system maintains a dynamically updated feature template for each tracked target. When a candidate target is found in a new monitoring sub-region, the system calculates the feature similarity between the candidate target and the existing target templates, and uses a multi-hypothesis tracking algorithm to handle complex scenarios such as target occlusion, intersection, and temporary disappearance.

[0065] On the basis of successfully associating cross-regional targets and forming continuous trajectories, the system further analyzes whether there is a cooperative relationship between multiple trajectories to identify the formation of a UAV cluster. The cluster identification algorithm monitors whether multiple moving targets in the airspace within a specific time window show a high degree of spatial correlation and time synchronization. The algorithm analyzes the consistency of the movement patterns of any two target trajectories by calculating the cross-correlation function between them, such as checking whether their speed changes are synchronized, whether their heading adjustments are consistent, and whether their relative distances remain stable. For multiple targets determined to belong to the same cluster, the system analyzes their spatial distribution patterns by calculating parameters such as the centroid, principal axis direction, and dispersion of the target group to identify whether they present typical formation patterns such as linear, triangular, or ring-shaped. The identification of formation patterns has reference value for judging the intent and threat level of the cluster. The cooperative analysis module also needs to extract deeper behavioral features from the movement trajectories, including the smoothness of the trajectory, the change rule of the curvature, and whether there are specific flight patterns such as hovering, circling, and rapid advance. These behavior patterns are important indicators for assessing the intent of the target. The system performs multi-scale analysis on each trajectory, observing both the overall path direction and local micro-maneuver actions to extract feature indicators that can distinguish between normal patrol, aerial photography, and potential malicious flight. All the information extracted from cross-regional cooperative analysis, including the complete trajectory of the target, kinematic parameters, cluster number, formation pattern features, and identified special flight behavior patterns, are systematically integrated to generate a structured preliminary threat signal set.

[0066] Each threat signal corresponds to a tracked entity target or cluster, and the signal contains the unique tracking ID of the target, the current best estimate position, the velocity vector, the height information, the confidence level, and the initial threat score based on behavior analysis. The confidence level calculation considers multiple factors such as the clarity of target detection, the continuity of trajectory tracking, and the similarity score of feature matching. It is a dynamic value that reflects the system's assessment of the reliability of the target information. The preliminary threat signal set constitutes the system's situational awareness summary of the current airspace activity targets. This signal set will be passed to the subsequent adaptive threat assessment module to provide rich input features for more detailed machine learning-based threat judgment. The entire cross-regional cooperative analysis process is a complex computing process of multi-data stream fusion and multi-hypothesis verification, which effectively elevates isolated, instantaneous target detection events to continuous, semantically meaningful trajectory and behavior pattern understanding, significantly enhancing the perception ability of cooperative and intelligent UAV activities.

[0067] Embodiment 3: The adaptive threat assessment model is the intelligent discrimination core of the entire system, and its performance directly determines the accuracy and timeliness of countermeasures. The first step in model construction is multi-level feature extraction. The system extracts static and dynamic attributes from the standardized image stream continuously input and the preliminary threat signal set generated by the cross-region collaborative analysis module to describe the UAV target. Static features are mainly obtained from single or multiple images, including shape complexity, aspect ratio, surface symmetry, infrared thermal radiation distribution pattern, and other morphological indicators. Dynamic features are calculated from time series data, such as the magnitude and direction of instantaneous velocity vectors, tangential and normal acceleration components, heading angle change rate, flight height variance, and trajectory curvature statistical properties. The feature extraction process needs to consider the complementarity of different sensor modalities. Visible light images provide texture details, and infrared images reflect thermal characteristics, which are fused into a unified feature vector. Cluster association information and formation pattern indicators from the collaborative analysis module are also considered as context features. These raw features usually have high dimensions and redundancy. The system uses principal component analysis and other feature dimension reduction methods to compress the feature vector dimension while retaining most of the information, improving the efficiency of subsequent model training and inference. After completing the feature engineering, the system uses accumulated historical data, especially confirmed threat cases and normal flight records, to supervise the training of deep learning classifiers. The training process aims to enable the model to distinguish between normal flight activities and various abnormal flight patterns.

[0068] The classifier usually adopts a hybrid architecture that combines convolutional neural networks and long short-term memory networks. The convolution branch is responsible for processing morphological features extracted from images, while the long short-term memory network is good at modeling time-series behavior features. The outputs of the two branches are fused in the deep layers of the network, and finally a probability score representing the target belonging to the "abnormal" category is output through the fully connected layer and the Softmax activation function. The adaptive nature of the model is reflected in its parameters, which can be dynamically adjusted according to external feedback and environmental changes. The system continuously monitors the results of cross-region collaborative analysis, and when the analysis indicates the emergence of new cluster tactics or abnormal motion patterns in the airspace, these information will be fed back to the model as feedback signals.

[0069] An adaptive mechanism for adjusting the model discrimination threshold can be represented by the following relationship:

[0070]

[0071] Where: represents the adaptive discrimination threshold currently used by the system, which is a dynamically changing value. is the basic discrimination threshold obtained by statistical analysis of historical data in advance. is an indicator reflecting the intensity of suspicious swarm activity identified in the current airspace, whose value increases with the number of swarms and the degree of coordination. represents the confidence level of the overall situation in the current time period, which decreases when weather conditions are bad or sensor data quality is generally low. and are two weight coefficients greater than zero, which control the degree of influence of swarm activity intensity and situation confidence on the final threshold. This mechanism makes the model more sensitive when detecting potential coordinated threats, and tends to be conservative when the overall perception uncertainty is high.

[0072] After the adaptive threat assessment model is trained and parameterized, it is put into real-time operation to identify abnormal flight behavior for each new frame of incoming data. The identification process involves inputting the real-time extracted target feature vector into the deployed model for forward propagation calculation. The model compares the real-time features with the large number of normal patterns learned internally and calculates an abnormal score that deviates from the normal pattern. The comparison process is multi-dimensional, with the system checking in parallel whether the flight altitude significantly deviates from the historical conventional altitude layer of the airspace, whether the flight speed shows abnormal acceleration or deceleration that does not conform to the performance characteristics of the type of UAV, and whether the heading angle changes show unprovoked violent jitter or continuous pointing to the protected key area.

[0073] For the identified abnormal targets, the system will perform a fine classification of threat levels, not only based on the absolute value of the abnormal score output by the model, but also considering factors such as the size of the target, flight altitude, distance from sensitive areas, and whether it belongs to a swarm. The threat level is finally divided into multiple gradients, such as "observation level", "warning level", and "high threat level". Each labeled target generates a detailed record, including target ID, timestamp, geographic coordinates, identified abnormal behavior type, calculated threat level, and key feature items as the basis for decision-making. All these records are aggregated to generate a structured list of violating flight targets. This list is a direct trigger condition for subsequent deep query and countermeasures decision-making by the system, and its accuracy and detail directly affect the effectiveness of the entire countermeasure action. The entire adaptive threat assessment and anomaly identification process constitutes a key transition from perception to cognition, which converts low-level data streams into high-level, actionable threat intelligence.

[0074] Example 4: Assume that at 3:25 PM on a weekday, the adaptive threat assessment model flags a rogue flight target, temporarily numbered UAV-734, as having an abnormally low flight altitude and a persistent heading towards a critical infrastructure, initially rated as a "warning level" threat. The system then triggers a deep query process on the airspace management database, which is a relational database storing multi-dimensional structured data. The deep query is not a simple single-condition search, but a complex operation of multi-table association and spatio-temporal range matching. The query engine first takes the core features of the target UAV-734 (such as the image feature hash value first captured, initial position, timestamp) as the primary key. It performs a fuzzy match and exact query in the "aircraft registration information table" to try to confirm whether the target corresponds to a registered drone serial number. The query result may show no matching record, which itself is an important risk indicator, indicating that the target may belong to an unreported "black flight" aircraft. Next, the system performs a spatio-temporal range query in the "historical flight trajectory record table" to search for similar feature hash values within the same geographical area or adjacent airspace within the past 72 hours, as shown in Table 1.

[0075] Table 1: Target UAV-734 association query results

[0076]

[0077] The query engine will associate the "airspace control status table", which dynamically updates the latest announcements from the air traffic control department. The system will perform a geospatial inclusion analysis of the target UAV-734's real-time latitude and longitude with the polygon control areas in the table to confirm whether it is currently in a no-fly zone, restricted flight zone, or temporary control airspace. The association with the "real-time weather data table" is to assess the behavior rationality. The system obtains the accurate weather data of the target's activity period in that area, including wind speed, wind direction, temperature, precipitation, and visibility, to analyze whether the target's abnormal flight behavior (such as sudden altitude drop, trajectory jitter) is likely caused by sudden wind shear, turbulence, and other adverse weather conditions. If the weather data is stable and the flight behavior is abnormal, the possibility of human malicious operation increases significantly. The system will perform an association query on the "sensitive facility geographic information table", which stores the geographic fence information of all protected critical infrastructure (such as energy facilities, transportation hubs). By calculating the target's real-time position and heading, and performing ray casting analysis along its velocity vector direction, it can determine whether its flight path is directly pointing at or likely passing through these sensitive areas. Combined with the query results of the "historical flight record table", if the same feature target is found to approach the same sensitive facility with a similar path at different times, it constitutes a pattern, greatly increasing the credibility of the threat. The results of all these association queries are temporarily integrated into a multi-dimensional data view for use by the subsequent intent analysis module.

[0078] The intent analysis module uses a rule-based reasoning and weighted scoring method to analyze the potential flight intent and risk level of the UAV target based on all the fragmented information obtained from the deep query. The module has a pre-installed rule base for multiple typical threat scenarios. For example, the rule can be defined as: "IF target is not registered AND flight path repeatedly passes through sensitive area boundary AND weather conditions are good THEN intent is likely to be'reconnaissance'." The system assigns a confidence score to each inferred intent hypothesis, and the score depends on the completeness and strength of the evidence chain supporting the intent. After completing the deep query and intent analysis, the system starts the generation of the complete threat trace report. The report generator is a template-driven document construction module. It first integrates the basic information in the illegal flight target list, and then embeds the key findings such as historical flight patterns, airspace control status, weather impact analysis, and sensitive facility correlation obtained from the deep query. The core part of the report is the reconstruction of the flight path and behavior pattern. The system uses the location points and timestamps of the target in all monitoring sub-areas to reconstruct its complete flight trajectory from entering the monitoring airspace to being identified using curve fitting algorithms, and marks the speed change points, height anomaly points, and heading mutation points on the trajectory line. The final complete threat trace report is a structured document, usually containing an executive summary, target basic information, complete flight trajectory backtracking graph, abnormal behavior detailed analysis, multi-dimensional correlation analysis result summary, analysis of potential invasion intent, and final assessment of comprehensive threat level. The report is output in a standardized format and can be directly provided to airspace management personnel as a key decision-making basis for taking further warning, driving away, or capturing countermeasures. The entire process embodies the information refinement and value sublimation from low-level sensor data to high-level tactical intelligence.

[0079] Example 5: The system's closed-loop process of self-optimization based on operational feedback enables the entire countermeasure system to continuously learn from actual handling cases and evolve its performance. Suppose that the system generates a complete threat traceability report after handling a complex incident that occurred in the southeast sector, detailing the complete process of a small drone swarm penetrating using low-altitude complex terrain, although the final target was successfully identified and disposed of, the report also points out that the system had a recognition delay in the edge grid area where the target first appeared, and the confidence level of the judgment on the new coordination communication mode between the swarm was initially low. This report containing both successful experience and shortcomings immediately becomes the core input for the system to optimize parameters and strategies. The optimization process first acts on the adaptive threat assessment model, and the system adds the complete time series data recorded in this incident, including all image frames from the appearance of the target to its identification, motion trajectory, swarm interaction features, and finally confirmed threat labels, as a new training sample to the incremental learning queue of the model. Instead of complete retraining, the model uses online learning or small batch update strategies, focusing on adjusting the classification boundaries that performed poorly in this incident, such as the model needs to learn to mark this low-altitude penetration mode using terrain cover earlier as abnormal, the learning process is achieved by adjusting the weights of specific nodes in the neural network, so that the model's sensitivity to similar features is improved. At the same time, the model's internal rule library is also updated, adding a judgment rule about the "multi-target low-altitude dispersed marching, intermittent merging" behavior pattern, and giving it a higher threat weight, which means that in the future, when similar patterns are detected again, the system can make high-level judgments more quickly.

[0080] The optimization of the spatial dynamic grid division strategy is another important aspect. The analysis report shows that the target was first captured in the southwestern edge mountainous grid during this intrusion event. Due to the large terrain undulations and high vegetation coverage, the originally set grid granularity was relatively coarse to reduce the computational load. Post-event analysis shows that the relatively coarse grid division leads to insufficient early detection capability for small targets. The system starts the optimization algorithm of the grid division strategy based on this experience. The algorithm considers the historical event frequency, terrain complexity, and this performance bottleneck data in the region to calculate that the benefits of using finer grid granularity in this edge mountainous region will be higher than the increase in computational cost. The optimization algorithm may decide to redivide the region into smaller monitoring sub-regions, although this will increase the data processing amount of individual grids, it can improve the discovery probability of low, slow, and small targets. The optimized grid configuration is updated to the system's airspace topology model. The cross-region collaborative analysis algorithm also receives feedback from this cluster event for tuning. The analysis report points out that the cluster target shows a non-typical coordination mode during movement, they do not maintain a fixed formation, but disperse for a short time at a specific geographic location, and then reassemble at the next scheduled location. The original coordination analysis algorithm has low recognition efficiency for this loose, time-based coordination mode. System algorithm engineers will adjust the judgment logic of the time-space correlation matrix related to target dispersion-aggregation behavior based on the data of this event, such as reducing the weight of formation maintenance and increasing the evaluation coefficient of target time synchronization at the preset path point, so that the algorithm can better understand and recognize this more covert cluster tactics. The parameter configuration and deployment scheme of the image acquisition device are also included in the optimization review range. The system background will generate an evaluation report on the performance of all relevant acquisition devices during this event. The report may show that the infrared thermal imager located in the southeast direction has an unsatisfactory environmental background temperature setting during the target intrusion stage, resulting in insufficient thermal contrast between the target and the background, affecting early identification. The system may automatically adjust the temperature sensitivity range or image enhancement parameters of the device, or it may suggest that the maintenance personnel fine-tune its installation pitch angle during the next maintenance period to eliminate a small blind area. More macroscopically, if multiple event analyses show that there is a weak link in the defense depth of a certain direction, the system may generate a proposal to demonstrate the necessity and expected effect of adding an auxiliary acquisition point or replacing a higher resolution device in that region.

[0081] The whole optimization process constitutes a complete "perception-decision-action-evaluation" closed loop, and each time of threat handling is not only a real combat response, but also a stress test and data collection opportunity for the system itself. Through continuous learning from actual operation, the system's identification accuracy, response speed and ability to adapt to new threats can be continuously evolved, so as to effectively cope with the challenges brought by the rapid development of unmanned aerial vehicle technology. The self-optimization capability ensures the long-term effectiveness of the system life cycle.

[0082] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0083] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A method for countering unmanned aerial vehicles (UAVs) based on image recognition, characterized in that, The method includes the following steps: Collect multi-source image data of the real-time monitoring airspace, preprocess the multi-source image data, and generate a standardized image stream; The standardized image stream is dynamically divided into spatial grids to determine multiple monitoring sub-regions, and spatial relationships between the monitoring sub-regions are established. Based on the spatial correlation, cross-regional collaborative analysis is performed on the standardized image stream to identify the motion trajectory and swarming pattern of UAV targets and generate a preliminary threat signal set; An adaptive threat assessment model is constructed based on the standardized image stream and the preliminary threat signal set, and the discrimination threshold of the adaptive threat assessment model is optimized using the cross-regional collaborative analysis results; The multi-source image data collected in real time is input into the adaptive threat assessment model to identify abnormal flight behavior, mark illegal flight targets, and feed the identification results back to the airspace management database. Based on the output of the adaptive threat assessment model, a deep query is performed on the airspace management database to analyze potential threat levels and intrusion intentions, and a complete threat tracing report is generated.

2. The method for countering unmanned aerial vehicles based on image recognition according to claim 1, characterized in that, The multi-source image data collected in the real-time monitoring airspace includes: Multiple image acquisition devices deployed in the monitored airspace simultaneously acquire visible light and infrared image data; The acquired visible light and infrared image data are time-stamp aligned and spatially registered. The registered multimodal image data are integrated into a time-series image sequence to generate a standardized image stream.

3. The method for countering unmanned aerial vehicles based on image recognition according to claim 2, characterized in that, The process of performing spatial dynamic mesh partitioning on the standardized image stream includes: Extract the spatial geographic features and real-time meteorological parameters from the standardized image stream; The granularity of grid division is dynamically adjusted based on airspace geographic characteristics and real-time meteorological parameters. Assign a unique area identifier to each monitoring sub-region and record the spatial coordinate range of each monitoring sub-region.

4. The method for countering unmanned aerial vehicles based on image recognition according to claim 3, characterized in that, The cross-regional collaborative analysis of the standardized image stream based on the spatial correlation includes: Establish a spatiotemporal correlation matrix between monitoring sub-regions to analyze the cross-regional movement patterns of UAV targets; Identify the formation patterns and cooperative flight characteristics of drone swarms; A preliminary set of threat signals is generated based on the characteristics of the movement trajectory, and a confidence weight is assigned to each threat signal.

5. The method for countering unmanned aerial vehicles based on image recognition according to claim 4, characterized in that, The construction of the adaptive threat assessment model includes: Extract morphological and behavioral features of the UAV target from the standardized image stream; A deep learning classifier is trained using the aforementioned preliminary threat signal set to establish a discrimination criterion between normal and abnormal flight modes; The sensitivity parameters of the adaptive threat assessment model are dynamically adjusted based on the results of cross-regional collaborative analysis.

6. The method for countering unmanned aerial vehicles based on image recognition according to claim 5, characterized in that, The identification of abnormal flight behavior includes: The real-time image data is compared with the normal flight pattern in the adaptive threat assessment model; Detect abnormal deviations in flight altitude, speed, and heading; The identified abnormal flight targets are classified by threat level, and a list of illegal flight targets is generated.

7. The method for countering unmanned aerial vehicles based on image recognition according to claim 6, characterized in that, The deep query of the airspace management database includes: Historical flight records and registration information were retrieved from the list of prohibited flight targets. It integrates meteorological data, airspace control notices, and other multi-dimensional auxiliary information; Analyze the flight intentions and potential risk levels of drone targets.

8. The method for countering unmanned aerial vehicles based on image recognition according to claim 7, characterized in that, The generation of a complete threat attribution report includes: Integrate the list of illegal flight targets, historical flight records, and multi-dimensional auxiliary information; Reconstruct the complete flight path and behavior patterns of drone targets; The output includes a structured report containing threat level assessment and source tracing analysis results.

9. A method for countering unmanned aerial vehicles based on image recognition according to claim 8, characterized in that, The method further includes: The discrimination rules of the adaptive threat assessment model are updated based on the complete threat attribution report. Optimize the dynamic spatial grid partitioning strategy and cross-regional collaborative analysis algorithm; Adjust the parameter configuration and deployment plan of the image acquisition equipment.

10. A drone countermeasure system based on image recognition, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the image recognition-based drone countermeasure method as described in any one of claims 1 to 9.

Images