A low-altitude security-oriented unmanned aerial vehicle identification method and system

Abstract

The application discloses a kind of unmanned aerial vehicle identification method and system for low altitude security, belong to low altitude security, artificial intelligence, computer vision and intelligent monitoring technical field.First, the continuous video image frame of low altitude monitoring area is acquired and candidate air target area is extracted, the low altitude scene distinguishability of candidate target level is calculated, and the global image enhancement parameter and candidate area enhancement parameter are determined in linkage;Through deep learning target detection model, detection confidence and target visual features are obtained, and rotary wing-rigid body collaborative features are calculated by combining continuous frame target correlation;Generate bird, kite and cloud layer interference evidence, and output enhancement, correlation, frequency band, weight and determination threshold through reinforcement learning strategy model adaptive output, calculate interference counterproof determination margin and output identification early warning.The application solves the problems of low distinguishability of low altitude small target, high interference false alarm and difficult parameter adaptation, improves the accuracy, stability and real-time early warning capability of unmanned aerial vehicle identification.

Description

Technical Field

[0001] This invention relates to the fields of next-generation information technology, artificial intelligence, computer vision, and low-altitude security technology, and in particular to a method and system for identifying unmanned aerial vehicles (UAVs) for low-altitude security. Background Technology

[0002] With the increasing demands for low-altitude security, perimeter protection of key areas, airport airspace protection, security for major events, and urban low-altitude airspace management, the automatic identification and timely warning of "low, small, and slow" aerial targets within low-altitude monitoring areas has become a crucial link in low-altitude security systems. In low-altitude UAV identification scenarios based on video images, monitoring equipment typically needs to continuously observe distant, small-scale, and highly dynamic aerial targets within a large field of view, and complete target detection, tracking, and category determination even in the presence of interference factors such as birds, kites, clouds, floating objects, background disturbances, and changes in lighting. However, in real-world applications, UAV targets often exhibit characteristics such as small scale, weak texture, unstable edges, significant motion blur, and insufficient contrast with the background in images. Furthermore, different interfering targets may share similarities with UAVs in terms of local appearance, motion trajectory, or short-term morphological changes, leading to problems such as false detections, missed detections, tracking interruptions, or delayed warnings in existing identification methods. Therefore, there is an urgent need for a UAV identification technology solution that can adapt to complex low-altitude scenarios and balance image recognizability, target motion patterns, and interference counter-evidence.

[0003] To address the aforementioned issues, those skilled in the art typically employ techniques such as image enhancement, candidate target extraction, deep learning target detection, continuous frame target association, motion feature analysis, and multi-feature fusion judgment. Image enhancement is generally used to improve target visibility under low contrast, noise, and blur conditions; deep learning target detection models are typically used to extract candidate target positions, confidence levels, and visual features from video frames; target association and tracking algorithms are typically used to maintain the consistency of the identity of the same aerial target across consecutive frames; and motion feature analysis is typically used to distinguish between drones and birds based on speed, acceleration, trajectory fluctuations, or morphological changes. However, these conventional techniques are mostly based on fixed preprocessing parameters, single detection models, or static fusion rules. In low-altitude scenes where identifiability changes rapidly, target scale continuously changes, and multiple types of interfering targets coexist, they still struggle to fundamentally solve the problems of disconnect between image enhancement and recognition judgment, lack of dynamic adjustment of target association parameters, insufficient utilization of interfering evidence, and difficulty in adaptive judgment thresholds.

[0004] A search revealed numerous existing solutions for visual detection, recognition, and tracking of low-altitude unmanned aerial vehicles (UAVs). For example, Chinese invention patent application CN116109950A, entitled "A Visual Detection, Recognition, and Tracking Method for Low-Altitude Anti-UAVs," discloses a technical solution that uses an improved YOLOv5s target detection model for UAV detection, an improved fine-grained target recognition model for UAV model identification, and a TCTrack target tracking model for UAV tracking. This solution primarily focuses on improving the target detection model, the fine-grained recognition model, and the tracking model itself. While this solution can handle low-altitude UAV detection, recognition, and tracking, it does not disclose a mechanism for evaluating the discernibility of low-altitude scenes based on candidate target-level image sharpness, local contrast, noise, background disturbance, and target scale. Furthermore, it does not involve closed-loop adjustment of global image enhancement, candidate region enhancement, target association windows, and judgment thresholds based on feedback from the previous recognition cycle. Therefore, in continuous monitoring scenarios with complex lighting, motion blur, background disturbance, and significant target scale variations, it is difficult to achieve coordinated adaptation between image enhancement, detection / tracking, and recognition / judgment.

[0005] Chinese invention patent CN114255407B, entitled "A High-Resolution Anti-UAV Multi-Target Recognition and Tracking Video Detection Method," discloses a scheme that uses a 4K high-definition surveillance camera for photoelectric data acquisition, constructs a multi-scale depth target detection network, and obtains UAV position information by integrating the image recognition results before and after segmentation. This scheme mainly improves the ability to capture long-range UAV targets through high-resolution acquisition and multi-scale detection. However, its technical focus is on image segmentation, multi-scale detection, and target locking, without establishing a quantifiable evaluation method for candidate targets, nor disclosing a mechanism for generating counter-evidence and adaptively weighting different interference sources such as birds, kites, and clouds. Especially when UAV target imaging has few pixels and contour and texture features are easily lost, relying solely on high-resolution acquisition and detection networks may still be insufficient to fully distinguish UAVs from interfering targets such as birds, insects, reflections, or clouds.

[0006] Chinese invention patent CN116363412B, entitled "A Method and System for Constructing a Classification Model of Birds and UAVs Based on Center-of-Center Motion Features," discloses a scheme that uses the center-of-centimeter motion trajectories of birds and UAVs to calculate velocity, acceleration, and trajectory fluctuation features in different frequency ranges, and then weights and integrates the prediction results of multiple classifiers to establish a classification model for birds and UAVs. While this scheme can classify birds and UAVs using center-of-centimeter motion features, its core objective focuses on modeling the differences in motion trajectories between birds and UAVs. It does not address the impact of low-altitude video image quality changes on detection confidence and subsequent judgment, nor does it disclose a processing method that combines the micro-periodic response of the rotor or propeller region, rigid body attitude stability, and center-of-centimeter trajectory changes to form collaborative discriminative features. Therefore, in situations where the target's local texture is weak, the trajectory is significantly affected by wind disturbance or camera shake, and the short-term motion characteristics of birds and UAVs are similar, relying solely on center-of-centimeter motion features may still result in insufficient discriminative basis and low robustness.

[0007] Furthermore, the Chinese invention patent CN111241931B, entitled "A Method for Target Recognition and Tracking of Aerial UAVs Based on YOLOv3," discloses a scheme that improves candidate box parameters based on K-means clustering and trains a YOLOv3 model, then combines target template similarity scores for target selection and template updating. While this scheme can improve the target recognition and tracking performance of aerial UAVs to some extent, it primarily relies on detection boxes, target templates, and similarity evaluation to maintain targets across consecutive frames. It does not disclose a reinforcement learning strategy adjustment mechanism for different recognition cycles, nor does it provide a processing link for jointly adaptively outputting image enhancement parameters, target association parameters, micro-periodic response bands, interference evidence weights, and decision margin thresholds. Therefore, when illumination, occlusion, scale, background motion disturbances, or changes in target appearance cause tracking deviations, this type of scheme still struggles to achieve dynamic compensation of parameters at each stage of the recognition link.

[0008] In summary, overcoming the problems existing in the current technology, such as insufficient identifiability of low-altitude small target images, lack of closed-loop linkage between image enhancement and recognition judgment, insufficient utilization of interference evidence such as birds, kites and clouds, failure to coordinate the modeling of micro-periodic features and rigid body motion features of rotor-type targets, and difficulty in adaptively adjusting judgment thresholds and tracking parameters, and proposing a more accurate, stable, robust and suitable UAV recognition method and system for deployment in complex low-altitude security scenarios, has become a technical challenge that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] To address the challenges of small target size, low image discernibility, high similarity of interfering targets such as birds, kites, and clouds in low-altitude security scenarios, and the difficulty of adapting fixed recognition parameters to complex environmental changes, this invention proposes a drone recognition method and system for low-altitude security. By constructing candidate target-level low-altitude scene discernibility, rotor-rigid body collaborative features, and multiple types of interference evidence, and utilizing a reinforcement learning strategy model to adaptively adjust image enhancement, target association, micro-periodic response band, interference evidence weights, and decision margin thresholds, this invention solves the problems of insufficient accuracy, high false positive and false negative rates, weak anti-interference capabilities, and lack of dynamic compensation in the recognition link of existing low-altitude drone recognition technologies.

[0010] A drone identification method for low-altitude security includes the following steps: S1: Acquire continuous video image frames of the low-altitude monitoring area, take a preset number of continuous video image frames as a recognition period, and extract candidate aerial target areas from the continuous video image frames of the current recognition period. S2: According to the first The first frame The image sharpness, local contrast, noise intensity, background motion disturbance, and target scale of each candidate aerial target region are used to calculate the low-altitude scene discernibility of the candidate targets. And based on the candidate target level low-altitude scene identifiability corresponding to each candidate aerial target area. In addition to the enhancement adjustment parameters output from the previous recognition cycle, the global image enhancement parameters and candidate region enhancement parameters for the current recognition cycle are determined, and global enhancement is performed on continuous video image frames, while local enhancement is performed on candidate aerial target regions. S3: Input the enhanced image frame into the deep learning object detection model to obtain the detection confidence of candidate aerial targets. and target visual features; S4: Based on the target association parameters output from the previous recognition cycle, perform target association on the same candidate aerial target in consecutive frames, and calculate the rotor-rigid body cooperative features based on the temporal changes of the edge gradient of the candidate aerial target region, the changes of the main body attitude, and the changes of the center of mass trajectory. ; S5: Generate evidence of bird interference based on target visual features and continuous frame motion features. Evidence of kite interference Evidence of cloud interference ; S6: Improve the identifiability of candidate target-level low-altitude scenes Detection confidence Rotor-rigid body cooperative characteristics Evidence of interference from birds Evidence of kite interference Evidence of cloud interference Input the reinforcement learning policy model to obtain the decision adjustment parameters, interference evidence weights and decision margin thresholds for the current recognition cycle, as well as the enhancement adjustment parameters, target association parameters and micro-cycle response frequency band adjustment parameters for the next recognition cycle; S7: Calculate the interference rebuttal judgment margin for candidate aerial targets based on the judgment adjustment parameters and interference evidence weights. and in the margin of interference rebuttal judgment When the threshold for judgment is exceeded, the corresponding candidate aerial target is judged as a drone target, and the drone identification result and warning information are output.

[0011] A drone identification system for low-altitude security includes the following modules: T1: Image sensing module, used to acquire continuous video image frames of the low-altitude monitoring area; T2: Candidate target extraction module, used to extract candidate aerial target regions from consecutive video image frames; T3: Scene identifiability calculation module, used to calculate the identifiability of candidate target-level low-altitude scenes; T4: Image enhancement module, used to perform global enhancement on continuous video image frames and local enhancement on candidate aerial target regions based on the identifiability of candidate target-level low-altitude scenes and the enhancement adjustment parameters output from the previous recognition cycle. T5: Deep learning detection module, used to output the detection confidence and visual features of candidate aerial targets; T6: Rotor-rigid body cooperative analysis module, used to calculate the rotor-rigid body cooperative characteristics of candidate aerial targets; T7: Interference Evidence Generation Module, used to generate interference evidence from birds, kites, and clouds; T8: Reinforcement learning strategy module, used to output the decision adjustment parameters, interference evidence weights and decision margin thresholds for the current recognition cycle, as well as the enhancement adjustment parameters, target association parameters and micro-cycle response frequency band adjustment parameters for the next recognition cycle; T9: Fusion judgment and early warning module, used to calculate the interference counter-evidence judgment margin, and output UAV identification results and early warning information; The image enhancement module is deployed in a programmable logic processing unit or an image signal processor, while the deep learning detection module, rotor-rigid body collaborative analysis module, and reinforcement learning strategy module are deployed in an intelligent computing unit. The programmable logic processing unit includes an FPGA or CPLD, and the intelligent computing unit includes an AI chip, a SOC chip, a GPU, an NPU, or an ASIC.

[0012] Compared with the relevant prior art, the beneficial effects of the present invention are: To improve the effective detection rate of small-sized UAV targets when the quality of low-altitude images fluctuates greatly: Since the discernibility of candidate targets in low-altitude scenes participates in image enhancement and judgment reliability adjustment, the system can perform more targeted enhancement for candidate targets that are far away, have low contrast, motion blur, or strong background disturbance, reducing the target inundation and edge artifact problems caused by uniform enhancement of the whole frame. Reduce false alarms caused by non-drone targets such as birds, kites, and clouds: The rotor-rigid body cooperative feature provides a positive discrimination basis for drones that is different from ordinary trajectory features, while the interference counter-evidence judgment margin suppresses the judgment result when the confidence of the interference target detection is high. Therefore, it can reduce false detections caused by relying solely on the confidence of the detection box or the centroid motion trajectory. Improve recognition stability in continuous monitoring scenarios: The reinforcement learning strategy model adjusts image enhancement, target association, micro-periodic response band and judgment threshold across cycles based on recognition feedback, so that the system can maintain relatively stable recognition results without manual parameter reset when there are sudden changes in illumination, camera shake, target scale changes or interference type changes. Reduce short-term missed detections and tracking interruptions: Since the target association parameters and micro-cycle response frequency band of the next identification cycle are adaptively adjusted according to the state of the previous cycle, when the UAV is temporarily blurred, changes its attitude, or has incomplete local features, the system can still maintain target identification and early warning output by combining continuous frame information. Suitable for deployment in low-altitude security edge terminals: global enhancement, local enhancement, deep learning detection and policy adjustment can be executed collaboratively by hardware such as FPGA, AI chip, SOC chip, GPU, NPU or ASIC, which is conducive to realizing real-time drone identification and early warning in scenarios such as airports, parks, borders and urban low-altitude monitoring. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the drone identification method for low-altitude security provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a drone identification system for low-altitude security provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the candidate target-level low-altitude scene discernibility calculation and image enhancement processing provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the construction of low-altitude monitoring simulation video data provided in an embodiment of the present invention; Figure 5 A schematic diagram showing the experimental comparison before and after enhancing the discernibility of candidate target-level low-altitude scenes according to an embodiment of the present invention; Figure 6 This is a comparison chart of detection confidence before and after adaptive enhancement in a simulation scenario provided by an embodiment of the present invention; Figure 7 A simulation comparison chart of the number of false alarms for different identification methods provided in the embodiments of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0015] Example 1: Refer to Figures 1 to 3 This embodiment provides a drone identification method and system for low-altitude security, applicable to low-altitude security scenarios such as airport airspace protection, perimeter protection of key industrial parks, low-altitude patrols at ports and wharves, security areas for major events, and low-altitude airspace monitoring in border areas and cities. The low-altitude monitoring area described in this embodiment refers to the airspace image area covered by fixed photoelectric monitoring equipment, PTZ camera equipment, array camera equipment, or edge intelligent monitoring terminals. This airspace image area may contain targets or interfering objects such as drones, birds, kites, clouds, floating objects, reflective points, flying insects, and background vibrations.

[0016] In this embodiment, the UAV identification system includes an image sensing module, a candidate target extraction module, a scene identifiability calculation module, an image enhancement module, a deep learning detection module, a rotor-rigid body collaborative analysis module, an interference evidence generation module, a reinforcement learning strategy module, and a fusion judgment and early warning module. The image sensing module acquires continuous video image frames of the low-altitude monitoring area; the candidate target extraction module extracts candidate aerial target regions from the continuous video image frames; the scene identifiability calculation module calculates the low-altitude scene identifiability of candidate targets; and the image enhancement module calculates the low-altitude scene identifiability of candidate targets based on the low-altitude scene identifiability of the candidate targets. The enhancement adjustment parameters output from the previous recognition cycle are used to perform global image enhancement and local enhancement of candidate regions; the deep learning detection module is used to output the detection confidence of candidate aerial targets. And target visual features; the rotor-rigid body cooperative analysis module is used to calculate rotor-rigid body cooperative features based on the same candidate target region in consecutive frames. The interference evidence generation module is used to generate interference evidence from flying birds. Evidence of kite interference Evidence of cloud interference The reinforcement learning strategy module is used to output the decision adjustment parameters, interference evidence weights, and decision margin thresholds for the current recognition period based on the state variables in the current recognition period. It outputs the enhancement adjustment parameters, target correlation parameters, and micro-period response frequency band adjustment parameters for the next identification cycle; the fusion judgment and early warning module is used to calculate the interference counter-evidence judgment margin. And determine the margin based on the evidence of interference. Output drone identification results and early warning information.

[0017] In one specific deployment method, the image sensing module employs a visible light camera, a telephoto camera, a low-light camera, or an infrared / visible light composite camera, with an output resolution of [missing information]. , , Alternatively, a resolution suitable for low-altitude monitoring can be used, with a video frame rate of 25, 30, 50, or 60 frames per second. The image enhancement module can be deployed in a programmable logic processing unit (FPGA) or an image signal processor, where the FPGA includes an FPGA or CPLD, and the image signal processor includes an ISP processor. The deep learning detection module, rotor-rigid body collaborative analysis module, interference evidence generation module, and reinforcement learning strategy module can be deployed in an intelligent computing unit, which includes an AI chip, a SOC chip, a GPU, an NPU, or an ASIC.

[0018] The method in this embodiment includes the following steps: S1: Acquire continuous video image frames of the low-altitude monitoring area and extract candidate aerial target areas. The image sensing module continuously acquires video images of the low-altitude monitoring area at a preset frame rate, obtaining a continuous video image frame sequence: ;in, This is the starting frame number of the current recognition period. This refers to the number of video image frames contained within one recognition period. The value of K can be determined based on the video frame rate, target motion speed, and computing resources. For example, K=16, K=24, or K=32 can be used for a 30 fps video, while K=32 or K=48 can be used for a 60 fps video. One recognition cycle can partially overlap with the next recognition cycle, for example, by sliding 8 or 16 frames between adjacent recognition cycles to improve the stability of continuous recognition.

[0019] The candidate target extraction module first extracts the sky region or low-altitude background region from the video image frames. Specifically, it can use color space thresholding, semantic segmentation models, horizon detection, preset monitoring area masks, background modeling, or combinations thereof to obtain low-altitude candidate detection regions. Subsequently, the candidate target extraction module performs inter-frame differencing, background differencing, sparse optical flow clustering, motion saliency detection, edge connectivity analysis, or lightweight target candidate box generation within the low-altitude candidate detection regions. Frame Image Multiple candidate aerial target areas were obtained: ;in, Indicates the first The first frame There are candidate aerial target regions. These can be represented as rectangular candidate boxes, rotated candidate boxes, target masks, or candidate target image patches. In a preferred implementation, the candidate target extraction module expands the candidate boxes by a preset ratio, such as 10% to 30%, to preserve the edges, rotor blur bands, wakes, or background transition areas around the target, thereby facilitating subsequent calculations of local contrast, high-frequency structure preservation, and temporal changes in edge gradients.

[0020] S2: Calculate the identifiability of candidate targets in low-altitude scenes. And based on the identifiability of candidate target-level low-altitude scenes Perform global image enhancement and local enhancement of candidate regions.

[0021] Reference Figure 3 For the first The first frame Candidate aerial target areas The scene recognizability calculation module calculates the local contrast respectively. High-frequency structure retention Image noise intensity Background motion disturbance intensity and candidate target region scale Based on the above quantities, the identifiability of candidate target-level low-altitude scenes is obtained. .

[0022] In one specific implementation, local contrast Based on candidate air target areas The grayscale difference between the target area and its surrounding background ring region is determined. Let the candidate aerial target region be defined. The average gray level is The average gray level of the local background ring region formed after the candidate aerial target region expands outward is The initial value of local contrast can then be calculated using the following formula. : ; in, To prevent constants with a denominator of zero.

[0023] Will According to the preset upper limit Normalizing to [0,1], we get .

[0024] In other implementations, local contrast It can also be calculated from the gray standard deviation, local Michelson contrast, local RMS contrast, or local histogram difference within the candidate region.

[0025] High-frequency structural retention Used to characterize the edge sharpness and distinguishable structural strength of candidate target regions. For candidate target regions Execute the Sobel, Scharr, or Laplacian operator to obtain the gradient magnitude map or second-order edge response map, and calculate the mean gradient energy or Laplacian variance within the candidate region. Let the i-th gradient energy within the candidate region be... The gradient magnitude of each pixel is The number of pixels in the candidate region is Then it can be calculated according to the following formula: ; And Normalized to [0,1] according to the preset gradient energy range, we get When candidate targets experience edge blurring due to long-range imaging, motion blur, or atmospheric disturbances, The corresponding reduction occurs when there are obvious blurred boundaries of the fuselage arms and rotors, fuselage outlines, or other high-frequency structures in the candidate region. The corresponding increase.

[0026] Image noise intensity Used to characterize the noise level in the overall or local sky background of the current frame. In one implementation, it can be derived from the first... frame In the low-altitude background region, a background sub-region without moving targets is selected. This sub-region is then subjected to median filtering or Gaussian filtering to obtain a smoothed image. The mean square value of the residual between the original image and the smoothed image is then calculated as the initial value of the noise intensity. Let the first sub-region of the background region be... The original gray level of each pixel is The smoothed grayscale is The number of pixels in the background sub-region is The initial value of noise intensity It can be represented as: ; Will Normalize to [0,1] to obtain the image noise intensity When images exhibit significant grainy noise under low light, high gain, rain / fog, or high ISO conditions, Increase.

[0027] Background motion disturbance intensity This is used to characterize the impact of camera shake, cloud drift, treetop swaying, thermal disturbance, or large-scale background motion on target recognition. In one implementation, the candidate target extraction module or scene discriminability calculation module first estimates a global motion compensation model based on background feature points, such as an affine transformation model or a homography transformation model; then, it calculates the variance of the compensated background region optical flow residual as the initial value of the background motion disturbance intensity. Let the background region be the first... The optical flow residual vector at each feature point is: The number of background feature points is Then it can be calculated according to the following formula: ; Will Normalize to [0,1] to obtain the background motion disturbance intensity When the camera pan / tilt shakes, strong winds cause localized background movement, or clouds drift rapidly, Increase.

[0028] Candidate target region scale It can be determined based on the candidate bounding box area, the candidate target mask area, the length of the long side of the candidate bounding box, the length of the short side, or the relative image area. In one implementation, if the candidate aerial target region... The candidate box width is Height is The scale of the candidate target region It can be represented as: ; To preset a reference scale, for example, we can take... Pixels A pixel or a reference small target area corresponding to the current camera focal length and monitoring distance.

[0029] In obtaining , , , and Then, the identifiability of candidate target-level low-altitude scenes is calculated according to the following formula. : ; in, , , , and All are dimensionless quantities and normalized to a preset numerical range. The larger the value, the higher the value. The first frame The more easily a candidate aerial target area is identified in the current low-altitude context; The smaller the value, the more likely the candidate target may be affected by factors such as long distance, small scale, low contrast, motion blur, noise interference, or background disturbance, requiring stronger targeted enhancement or more careful identification and judgment.

[0030] The image enhancement module determines the target region based on the current recognition period. In addition to the enhancement adjustment parameters output from the previous recognition cycle, global enhancement is performed on consecutive video image frames, and local enhancement is performed on candidate aerial target regions.

[0031] The image enhancement module determines the target based on multiple candidate targets within the current recognition period. Calculate global identifiability statistics, such as the average. Minimum value and the proportion of low-recognition candidate targets .in, for Below the first preset threshold The proportion of the number of candidate targets to the total number of candidate targets can be expressed as: ; in, This is an indicator function; it takes a value of 1 when the condition within the parentheses is true, and a value of 0 otherwise. The image enhancement module determines global image enhancement parameters, including global denoising intensity, based on the above statistics. Low light intensity enhancement Strong light suppression intensity Defogging or contrast tensile strength and global image stabilization compensation intensity At least one of them.

[0032] For example, when higher and Increase global denoising strength when the level is low. When the overall frame brightness is low and When the intensity is low, increase the weak light enhancement intensity. When the sky background is overexposed or there is significant strong light reflection, increase the intensity of strong light suppression. ;when When the intensity is high, increase the global image stabilization compensation intensity. Alternatively, reduce the weight of the background disturbance region on the extraction of candidate targets.

[0033] For each candidate aerial target area The image enhancement module further... Determine enhancement parameters for candidate regions, including local contrast enhancement intensity. Motion blur suppression intensity Local sharpening intensity Local noise reduction intensity Scale normalization ratio and over-enhanced inhibition coefficient At least one of them.

[0034] The enhancement adjustment parameters output from the previous recognition cycle are used to offset the aforementioned enhancement parameters, for example, to... , , , and Applying an increment or attenuation coefficient allows the image enhancement process to smoothly change across multiple recognition cycles as the low-altitude environment changes, rather than jumping drastically within a single frame.

[0035] S3: Input the enhanced image frame into the deep learning object detection model to obtain the detection confidence of candidate aerial targets. And target visual features.

[0036] The enhanced continuous video image frames are input into the deep learning detection module. This module can employ a single-stage object detection network, a two-stage object detection network, a Transformer-based object detection network, a lightweight edge detection network, or a multi-scale feature fusion detection network. The deep learning detection module can be pre-trained using a dataset containing samples of drones, birds, kites, clouds, floating objects, and background interference, or it can be incrementally updated based on feedback samples after deployment on the low-altitude security platform.

[0037] In one implementation, the deep learning detection module extracts features from both the enhanced full-frame image and local image patches of the candidate region. For the full-frame image, the deep learning detection module extracts multi-scale feature maps; for local image patches of the candidate region, the deep learning detection module performs fine-grained feature extraction using the scale-normalized target region. The deep learning detection module outputs candidate bounding boxes, initial probabilities of target categories, and detection confidence scores. and target visual feature vector .

[0038] Detection confidence Used to indicate that the deep learning detection model considers the first... The first frame The probability or score of a candidate aerial target belonging to a drone target or a suspected drone target. Target visual feature vector. It can include at least one of the following: features from intermediate layers of deep networks, candidate box region features, target contour features, target texture features, target aspect ratio features, local edge distribution features, and multi-scale semantic features. Target visual feature vector It is used not only for target determination in the current frame, but also for subsequent target association and interference evidence generation.

[0039] S4: Based on the target association parameters output from the previous recognition cycle, perform target association on the same candidate aerial target in consecutive frames, and calculate the rotor-rigid body cooperative features based on the temporal changes of edge gradients, body attitude changes, and center of mass trajectory changes of the candidate aerial target region. .

[0040] The rotor-rigid body co-analysis module uses the current identification period The candidate target detection results in the frame image are used as input to associate candidate regions of the same candidate aerial target in consecutive frames. The target association parameters are output by the reinforcement learning policy module in the previous recognition cycle, and the target association parameters may include the association time window length. Candidate matching distance threshold Intersection over Union (IOU) threshold Appearance feature similarity weight Motion prediction weights Missed detection maintains frame count and trajectory smoothing coefficient At least one of them.

[0041] In one implementation, the rotor-rigid body co-analysis module uses Kalman filtering, uniform motion model, uniform acceleration motion model, or optical flow-assisted prediction model to predict the position of candidate targets in the next frame, and combines this with candidate box intersection-over-union ratio, centroid distance, and detection confidence. Target visual feature vector Similarity and motion direction consistency are used to construct an association cost matrix, and then Hungarian matching or greedy matching is used to obtain the trajectory of the same candidate target in consecutive frames. For candidate targets that are missed in a short time, when the number of missed frames does not exceed At this time, the rotor-rigid body cooperative analysis module retains the trajectory state of the candidate target to reduce tracking interruptions of the UAV when there is motion ambiguity, short-term occlusion, or insufficient local enhancement.

[0042] For the associated first For each candidate target trajectory, the rotor-rigid body co-analysis module calculates the edge gradient map within the candidate target region in each frame and obtains the edge gradient time series. Within the candidate target region, fuselage contour edges, suspected rotor edges, target circumscribed region boundaries, or sets of significant gradient points can be selected to calculate the gradient magnitude statistics for each frame. ,in This indicates the frame number within the current identification period. Perform Fourier transform, short-time Fourier transform, or wavelet transform to obtain the edge gradient time series at frequency. Energy at the location .

[0043] Micro-periodic response frequency band of rotor or propeller region Based on the scale of candidate target regions The candidate frequency bands are adaptively determined based on the video frame rate, the dominant frequency distribution of the historical edge gradient time series, the micro-periodic response frequency band adjustment parameters output from the previous recognition cycle, and the preset UAV rotor sample frequency band range. The candidate frequency band set is denoted as... And satisfy: Since ordinary video frame rates may not be able to directly capture the high-frequency period corresponding to the actual rotor speed, the micro-period response in this embodiment includes not only the actual period response but also the equivalent period response formed by motion blur, rolling shutter effect, video sampling aliasing, rotor occlusion of fuselage edges, and local brightness variations. For example, in a 30 fps video, the micro-period response frequency band... Can fall into to Within the observable frequency range; in 60 frames per second video, the micro-period response band can fall within to Within the observable frequency range. This frequency band does not limit the actual rotor speed, but is used to detect the observable edge gradient periodic components formed by rotor or propeller motion under video sampling conditions.

[0044] Simultaneously, the rotor-rigid body co-analysis module calculates the main body orientation angle sequence of the candidate target. The main orientation angle can be determined by the principal axis direction of the candidate target region, the direction of the minimum bounding rectangle, the direction of the second-order central moment, or the principal direction of the target profile. For rigid-body UAVs, the main orientation of the fuselage is usually relatively stable within a short time window, and the main attitude changes little. For interfering targets such as birds and kites, their shape may undergo non-rigid changes or sway with the wind, and the main orientation angle changes more significantly. Therefore, the variance of the main orientation angle can be used. Characterize the short-time rigid body stability of candidate targets.

[0045] In addition, the rotor-rigid body co-analysis module calculates the centroid trajectory jitter based on the deviation between the candidate target centroid trajectory and the predicted trajectory. The predicted centroid position of the candidate target in each frame can be obtained using a Kalman filter, a uniform velocity model, or a uniform acceleration model. The normalized mean or variance of the residuals between the actual associated centroid position and the predicted centroid position can then be calculated as... For drone targets, short-term trajectories are usually highly predictable; for bird flapping, kite swaying, or cloud drift, the center-of-mass trajectory may exhibit more pronounced non-smooth changes or low-texture drift.

[0046] In obtaining , , , and Then, the rotor-rigid body cooperative characteristics are calculated according to the following formula. : ; in, For the candidate frequency band set, For the first The micro-periodic response frequency band of the rotor or propeller region corresponding to each candidate aerial target To prevent constants with a denominator of zero. The larger the value, the more obvious the equivalent periodic response of the rotor or propeller, the more stable rigid body attitude, and the smoother center of mass trajectory of the candidate target, thus better matching the short-term visual motion characteristics of UAV targets. The smaller the value, the more likely the candidate target may lack rotor micro-periodic response, exhibit significant non-rigid deformation, or experience trajectory jitter. Therefore, it is not advisable to rely solely on detection confidence levels. It was determined to be a drone.

[0047] S5: Generate evidence of bird interference based on target visual features and continuous frame motion features. Evidence of kite interference Evidence of cloud interference .

[0048] The interference evidence generation module uses the target visual feature vector output by the deep learning detection module. The target-related trajectory, main body attitude change, edge gradient time series, candidate target region scale change and centroid trajectory change output by the rotor-rigid body co-analysis module are used as inputs to calculate the interference evidence corresponding to different interference sources.

[0049] Evidence of bird interference The interference evidence generation module extracts wing periodic oscillation features, non-rigid morphological change features, and trajectory non-smoothness features. The wing periodic oscillation features can be determined based on the periodic area changes, contour width changes, and local edge opening and closing changes of the left and right or upper and lower edges of the candidate target region; the non-rigid morphological change features can be determined based on the Hu moment changes of the candidate target mask, the contour principal axis ratio changes, the aspect ratio changes of the candidate box, or the edge point distribution changes; the trajectory non-smoothness features can be determined based on the sudden change in the centroid velocity direction, short-term acceleration changes, or trajectory curvature changes.

[0050] In one implementation, the above features can be normalized and then input into a bird interference classifier, or bird interference evidence can be obtained by weighted summation. : ; in, This is a characteristic of periodic wing oscillation. This is a characteristic of non-rigid morphological change. This is a feature of non-smooth trajectory. , and These are the weighting coefficients. It is a normalization function or a sigmoid function.

[0051] Evidence of kite interference The interference evidence generation module extracts contour flexibility deformation features, wind-driven swaying features, and traction-type trajectory constraint features. Contour flexibility deformation features can be determined based on the low-frequency swaying of the target contour edge, slow changes in local shape, and the degree of edge curvature; wind-driven swaying features can be determined based on the periodic lateral swaying of the target's center of mass on both sides of the main motion direction; traction-type trajectory constraint features can be determined based on whether the candidate target trajectory sways around a certain approximately fixed traction point or an arc-shaped trajectory range.

[0052] In one implementation, evidence of kite interference can be generated according to the following formula. : ; in, It is a feature of flexible contour deformation. It is characterized by swaying with the wind. This is a traction-type trajectory constraint feature. , and These are the weighting coefficients.

[0053] Evidence of cloud interference The interference evidence generation module extracts low-texture drift features, edge diffusion features, and continuous expansion features of regional morphology. Low-texture drift features can be determined based on low gradient energy, low texture entropy, and slow overall movement with the background within the candidate region; edge diffusion features can be determined based on low gradient peaks at the target edges, large edge widths, or high degree of boundary blur; continuous expansion features of regional morphology can be determined based on the candidate region area slowly increasing across multiple consecutive frames, unstable morphological boundaries, or increased connectivity with surrounding cloud regions.

[0054] In one implementation, cloud interference evidence can be generated according to the following formula. : ; in, It is a low texture drift feature. It is a feature of edge diffusion. This is a characteristic of continuous expansion of regional morphology. , and These are the weighting coefficients.

[0055] Evidence of bird interference Evidence of kite interference Evidence of cloud interference All values ​​are normalized to a preset range, such as [0,1]. A higher interference evidence value indicates a higher probability of the corresponding interference type. In this embodiment, interference evidence does not independently determine whether a candidate target is a drone, but rather serves as a counter-evidence item in the subsequent interference counter-evidence margin. The calculations are used to avoid directly misclassifying targets with "high detection confidence but obvious characteristics of flying birds, kites, or clouds" as drones.

[0056] S6: Improve the identifiability of candidate target-level low-altitude scenes Detection confidence Rotor-rigid body cooperative characteristics Evidence of interference from birds Evidence of kite interference Evidence of cloud interference Input the reinforcement learning policy model to obtain the decision adjustment parameters, interference evidence weights, and decision margin threshold for the current recognition cycle. , as well as the enhanced adjustment parameters, target correlation parameters, and micro-period response frequency band adjustment parameters for the next identification cycle.

[0057] The reinforcement learning strategy module constructs state variables for each recognition cycle. State variables This can include statistical features and target-level features of multiple candidate targets within the current identification period. For example, state variables. It can include the current candidate target. , , , , , During the current identification period Mean, minimum, and variance; test confidence level. Mean and maximum values; number of consecutive frames associated with the target; number of frames held up in the current missed detection; intensity of background motion disturbance. Image noise intensity ; as well as the judgment results and alarm delay information of the previous identification cycle.

[0058] In a preferred implementation, the reinforcement learning policy model employs a deep Q-network model. To ensure consistency between the deep Q-network model and the parameter output method, this embodiment sets the action quantity as a discrete set of actions, where each discrete action... Update vector corresponding to a set of parameters Input state variables of a deep Q-network model Then, the Q-value of each discrete action is output, and the action with the largest Q-value is selected as the adjustment action for the current recognition cycle. The calculation formula is as follows: ; in, For a discrete set of actions, For deep Q-network models in state variables Next action The value of the output action.

[0059] action The vector can be updated by mapping it to the following parameters: ; in, , , , and Used to adjust image enhancement parameters for the next recognition cycle; , and Used to adjust the target association parameters for the next identification cycle; The micro-periodic response band used to adjust the next identification cycle; , and Used to adjust the weight of interfering evidence; Used to adjust the decision margin threshold.

[0060] Discrete actions can include, but are not limited to, the following types: The first type of action is an enhancement adjustment action, used to increase, maintain, or decrease the global denoising intensity in the next recognition cycle. Local contrast enhancement intensity Motion blur suppression intensity Local sharpening intensity and scale normalization ratio ; The second type of action is the target association window adjustment action, which is used to increase, maintain, or decrease the length of the association time window in the next recognition cycle. Matching distance threshold Missed detection maintains frame count And appearance similarity weight ; The third type of action is the micro-periodic response frequency band adjustment action, used to adjust the frequency band according to the current target scale. The frame rate and historical main frequency distribution, and the micro-periodic response frequency band of the rotor or propeller region in the next identification cycle, shifting upwards, downwards, expanding, or narrowing. ; The fourth type of action is the interference evidence weight adjustment action, which is used to increase or decrease the interference evidence weight of birds. Kite interference evidence weighting Weight of evidence related to cloud interference ; The fifth type of action is the decision margin threshold adjustment action, which is used to increase, maintain, or decrease the decision margin threshold in the current recognition cycle or the next recognition cycle. .

[0061] For example, when the status indicator is displayed Lower medium, When the threshold is high and various types of interference evidence are low, the reinforcement learning policy model can output an increase in the enhancement strength of the candidate region and an appropriate decrease in the decision margin threshold. And maintain a relatively wide target association window to reduce missed detections under low image quality conditions; when the status variable is displayed Higher but or When the value is high, the reinforcement learning policy model can output an increase in the weight of the corresponding perplexing evidence. or Increase the judgment margin threshold It also adds actions requiring continuous frame confirmation to suppress false alarms from birds or kites; when the status indicator displays... When the target association is high and the continuity of the association decreases, the reinforcement learning policy model can output an increase in the target association window, increase the motion prediction weights, and adjust the micro-periodic response band. The action is to reduce tracking interruptions caused by camera shake or background disturbances.

[0062] The reward for a reinforcement learning strategy model can be determined based on at least two of the following: recognition confirmation results, false positives, false negatives, target tracking continuity, and recognition delay. For example, a positive reward is given when the system's output drone recognition result is confirmed as correct by subsequent continuous tracking results, manual confirmation results, or feedback results from a low-altitude security platform; a negative reward is given when the system misidentifies birds, kites, clouds, or other interfering targets as drones and generates false alarms; a negative reward is given when the system fails to recognize a real drone target within a preset time; a positive reward is given for tracking continuity when the same drone target maintains a stable association and continuously outputs consistent recognition results across multiple recognition cycles; and a delay penalty is given when the recognition delay is too long.

[0063] S7: Calculate the interference rebuttal judgment margin for candidate aerial targets based on the judgment adjustment parameters and interference evidence weights. And determine the margin based on the evidence of interference. Output drone identification results and early warning information.

[0064] The fusion-based judgment and early warning module receives the detection confidence score output by the deep learning detection module. The scene recognizability calculation module output The output of the rotor-rigid body co-analysis module The output of the interference evidence generation module , and The decision adjustment parameters, interference evidence weights, and decision margin thresholds output by the reinforcement learning strategy module. The fusion-based judgment and early warning module calculates the interference counter-evidence margin of candidate aerial targets according to the following formula. : ; in, , , , , and the judgment margin threshold It is obtained from the action mapping output by the reinforcement learning policy model or selected by the reinforcement learning policy model; The discernibility adjustment index is used to adjust... The strength of the impact on the decision margin; The interference evidence smoothing parameter is used to control the degree of influence of the stronger interference evidence among multiple interference evidence on the counter-evidence. , and These are the weights of evidence interference from birds, kites, and clouds, respectively. To prevent constants with a denominator of zero, in a preferred implementation, the parameters satisfy: The numerator of the above formula represents positive evidence that the candidate target is a drone, where... Provides deep learning detection confidence. Provides rotor-rigid body cooperative positive discrimination features. Provides adjustment for the identifiability and reliability of candidate target-level scenes. The denominator of the above formula represents the interference counter-evidence, where the interference evidence is birds. Evidence of kite interference Evidence of cloud interference The fusion is performed using a weighted log-sum-exp method. This method can approximately reflect the stronger interference evidence among multiple interference types, while maintaining a continuously adjustable and smooth fusion effect. When a candidate target has a high detection confidence level... However, when there is strong evidence of interference from birds, kites, or clouds, the denominator of the counter-evidence increases. This reduces false alarms; when a candidate target has a high detection confidence level... More obvious rotor-rigid body cooperative characteristics And when the evidence of interference is weak, This increases the likelihood that the target will be identified as a drone target.

[0065] when Greater than the decision margin threshold At that time, the fusion judgment and early warning module will... The first frame The candidate aerial targets were identified as drone targets: ; when Not greater than the judgment margin threshold At that time, the fusion judgment and early warning module will... The first frame The candidate aerial targets are marked as interference targets, non-UAV targets, or targets awaiting confirmation: ; When a target is identified as a drone, the fusion judgment and early warning module outputs the drone target's location, target number, and detection confidence level. Information such as target trajectory, warning level, and warning time is collected. Warning information can be sent to low-altitude security platforms, perimeter protection platforms, airport airspace monitoring platforms, audible and visual alarm devices, video surveillance clients, or command and dispatch systems. When a candidate target is marked as an interference target, a non-UAV target, or a target awaiting confirmation, the system continues to track and confirm it in subsequent identification cycles based on the target's association status.

[0066] In a preferred implementation, to reduce false alarms caused by sporadic noise, single-frame false detections, or instantaneous background disturbances, the fusion judgment and early warning module can also... Perform short-term smoothing. For example, in continuous... A frame or a series In each recognition cycle, when the same candidate target All are greater than the judgment margin threshold Or the smoothed mean of the decision margin Greater than the decision margin threshold When this occurs, a higher-level warning is output. The smoothed decision margin mean... It can be represented as: ; The corresponding continuous confirmation condition can be expressed as: ; When only a single frame Greater than the decision margin threshold At this time, a low-level prompt or a result awaiting confirmation can be output. This short-term smoothing process does not change the margin of the interference-based proof-of-contrast judgment in this embodiment. The basic judgment logic for drone identification.

[0067] Furthermore, this embodiment also includes an identification feedback update step. The low-altitude security platform can return subsequent tracking results, manual confirmation results, radar / radio detection collaborative confirmation results, or platform processing feedback results as identification feedback information to the reinforcement learning strategy module. The reinforcement learning strategy module updates the reward amount based on the identification feedback information and further optimizes the enhancement adjustment parameters, target association parameters, micro-periodic response band adjustment parameters, interference evidence weights, and decision margin thresholds in subsequent identification cycles. Therefore, this embodiment forms a closed-loop recognition link consisting of low-altitude video image input, candidate target-level discernibility evaluation, global and local image enhancement, deep learning detection, continuous frame target association, rotor-rigid body collaborative analysis, interference evidence rebuttal, reinforcement learning cross-cycle parameter tuning, and fusion judgment and early warning.

[0068] The technical advantage of this embodiment is that, firstly, by calculating the low-altitude scene discernibility of each candidate aerial target region, the discernibility of the candidate target level is achieved. It can distinguish the identifiable states of different targets in the same frame image, enabling more targeted local enhancement of distant, small-scale, low-contrast, or motion-blurred UAV targets; secondly, by... Simultaneously used for image enhancement and decision margin adjustment, the image quality assessment results no longer remain in the preprocessing stage but participate in the final recognition reliability judgment; furthermore, through rotor-rigid body cooperative features... Introducing a collaborative discrimination criterion of edge gradient micro-periodic response, subject posture stability, and centroid trajectory smoothness can distinguish it from relying solely on detection confidence. Or it could rely solely on centroid trajectory identification; further, it could utilize bird interference evidence. Evidence of kite interference Evidence of cloud interference The fusion of evidence by contradiction can suppress false alarms when the detection model gives a high confidence level to the interfering target; finally, the reinforcement learning policy model is used to adjust the enhancement parameters, target association parameters, micro-periodic response band, and decision margin threshold. By adjusting across recognition cycles, it can adapt to complex situations in low-altitude security scenarios, such as sudden changes in lighting, camera shake, changes in target scale, enhanced background disturbances, and changes in interference types, thereby improving the stability, robustness, and real-time early warning capabilities of UAV recognition.

[0069] Example 2: This example provides an implementation method for UAV recognition processing based on low-altitude surveillance simulation video data. This example mainly illustrates the construction method of the simulation video data, the generation method of candidate aerial target regions, the calculation method of the candidate target-level low-altitude scene discernibility, and the specific operation process of performing global image enhancement and local enhancement of candidate regions based on the candidate target-level low-altitude scene discernibility. Subsequent examples can further provide recognition performance, false alarm suppression effect, and ablation comparison results based on the simulation video data generated in this example.

[0070] like Figure 4 As shown, this embodiment first constructs low-altitude monitoring simulation video data. The generation process of the simulation video data includes low-altitude monitoring simulation scene construction, target and interference object injection, image degradation processing, continuous video sequence output, and candidate target real label generation. Through the above process, continuous video image frames containing UAV targets, bird interference, kite interference, cloud edge interference, and floating object interference can be obtained.

[0071] In one specific implementation, the resolution of the simulated video is set to 1920×1080, the frame rate is set to 25fps, and each 16 consecutive frames constitute a recognition cycle, with a step size of 8 frames between adjacent recognition cycles. This configuration ensures that one recognition cycle corresponds to a video duration of approximately 0.64 seconds, which not only covers the short-term motion changes of small low-altitude targets but also provides a continuous frame basis for subsequent edge gradient temporal analysis, subject pose change analysis, and centroid trajectory change analysis.

[0072] The low-altitude monitoring simulation scene library includes six scene categories: clear sky scene, foggy low-contrast scene, backlight scene, complex background scene with distant mountains / buildings, slight camera shake scene, and cloud drift scene. Specifically, the clear sky scene simulates typical monitoring conditions where background texture is weak and target edges are relatively clear; the foggy low-contrast scene simulates situations where air scattering, increased humidity, or haze reduces the grayscale difference between the target and the sky background; the backlight scene simulates situations where the target is locally too dark and the background locally overexposed when the sun's incident direction is close to the camera's line of sight; the complex background scene with distant mountains / buildings simulates complex background interference when the target crosses distant mountain outlines, building edges, tower structures, or city skylines; the slight camera shake scene simulates background motion disturbances caused by gimbal wind disturbance, support vibration, or long-distance zoom; and the cloud drift scene simulates false target interference caused by slow cloud edge movement, local diffusion, and changes in low-texture areas.

[0073] When constructing the simulation scene, the sky background can be set as a brightness gradient background or a layered sky background, and clouds, distant mountain outlines, building outlines, thin linear tower structures, or drifting areas at the edge of clouds can be superimposed in different regions. For low-contrast fog scenes, an imaging degradation model can be used to process the original background image and the target image to reduce the grayscale difference between the target and the sky background. After the low-altitude monitoring simulation scene is constructed, UAV targets and non-UAV interference objects are injected into the scene. UAV targets include multi-rotor UAVs and small fixed-wing UAVs. Multi-rotor UAVs can be represented by a cross-shaped or rectangular fuselage outline, rotor arm outline, and weakly textured local edges; small fixed-wing UAVs can be represented by a combination of fuselage and wing outlines. Interference objects include birds, kites, drifting targets at the edge of clouds, and floating objects. Among them, birds can be represented by non-rigid contours with periodic wing oscillations; kites can be represented by flexible polygonal contours with wind-driven oscillation and traction-type trajectory constraints; drifting targets at the edge of clouds can be represented by grayscale blocks with low texture, edge diffusion, and slow expansion of regional morphology; and floating objects can be represented by irregular spots with small scale, weak texture, and low movement speed.

[0074] In one specific implementation, the equivalent scale of the target and the interfering object in the image is set to 9×9 pixels to 48×48 pixels. For small targets at a distance, the equivalent side length of the candidate target region can be set to 9 pixels to 18 pixels; for medium-distance targets, the equivalent side length of the candidate target region can be set to 19 pixels to 35 pixels; and for relatively close targets, the equivalent side length of the candidate target region can be set to 36 pixels to 48 pixels. By controlling the equivalent scale of the target, the small-scale imaging characteristics of "low, small, and slow" targets in low-altitude security scenarios in long-distance monitoring images can be simulated.

[0075] For the motion trajectory of a UAV target, trajectory modes such as uniform straight line, low-speed turn, short-term hover, slow climb, or lateral crossing can be adopted. The position change of the target's centroid in consecutive frames can be represented as: ; Where c is the target's centroid position, v is the target's velocity, a is the target's acceleration, Δt is the inter-frame time interval, and r is a small random term used to simulate wind disturbance or platform perturbation. For multi-rotor UAVs, short-term periodic changes in edge intensity can be superimposed on the rotor or propeller region, based on a relatively stable fuselage profile, to simulate the real rotor response, motion blur response, or equivalent periodic response formed by video sampling aliasing. For small fixed-wing UAVs, this allows for smoother changes in the main body's attitude angle and makes the rotor's equivalent periodic response weaker than that of multi-rotor UAVs.

[0076] For birds causing interference, periodic wing oscillations can be set, causing their local contours to expand, contract, and deform irregularly in consecutive frames. The bird's center of mass trajectory can be set to exhibit short-term vertical fluctuations or curved turns, with the main body's azimuth angle change amplitude being greater than that of UAV targets. For kites causing interference, a flexible contour can be set to deform with the wind, causing the region boundary to oscillate slowly in consecutive frames. By setting traction constraint points, the center of mass trajectory can exhibit oscillating characteristics around the constraint direction. For targets drifting at the cloud edge, low-texture grayscale areas, edge diffusion, and slow changes in region area can be set, potentially causing them to be mistaken for slowly moving aerial targets during the candidate target extraction stage.

[0077] After the target and interference objects are injected, the simulated video undergoes image degradation processing. This processing includes noise superposition, motion blur, local overexposure, local underexposure, low contrast degradation, and background motion disturbance. Noise superposition can use Gaussian or Poisson noise, and motion blur can use a directional blur kernel, with the blur direction related to the target's motion direction or the camera's disturbance direction. For low-contrast scenes, the gradient magnitude at the target edges can be reduced; for backlit scenes, background brightness can be enhanced and target dark texture reduced; for cloud drift scenes, slow grayscale changes and shape expansion can be introduced at the cloud edge regions. Through these processes, the simulated video includes not only ideal targets but also common issues in actual low-altitude security deployments, such as weak texture, low contrast, small scale, background disturbance, and similarity problems with interfering targets.

[0078] While generating the simulated video sequence, a real-world label is generated for each frame. The real-world label includes at least the target category, target number, target centroid location, target bounding box, target scale, target trajectory number, and a label indicating whether the target belongs to a drone target. For drone targets, the real-world label can further record its target type as a multi-rotor drone or a small fixed-wing drone; for interfering targets, the real-world label can record its target type as a bird, kite, cloud-edge drifting target, or floating object. These real-world labels do not participate in the identification and judgment process; they are only used for subsequent simulation verification, false detection statistics, missed detection statistics, and identification feedback generation.

[0079] After obtaining the simulated video sequence, the continuous video image frames are processed according to the method of this invention. First, candidate aerial target regions are extracted from the continuous video image frames of the current recognition period. In one specific embodiment, background motion compensation can be performed on the continuous frames first, and then inter-frame difference images or optical flow residual images can be calculated; subsequently, threshold segmentation, connected component analysis, and candidate box generation are performed on the difference images or optical flow residual images to obtain candidate aerial target regions. For each candidate region, A can be used to represent the candidate aerial target region.

[0080] Background motion compensation can be achieved by obtaining the inter-frame global transformation matrix through background region feature point matching or global optical flow estimation, and then registering the previous frame image to the coordinate system of the current frame image. After background compensation, the gray-level residual or gradient residual between the current frame and the compensated frame is calculated to obtain the motion-salient region. Morphological opening and closing operations are performed on the motion-salient region to remove isolated noise points and connect the target local region. Then, candidate regions are selected based on the area of ​​the connected components, the aspect ratio of the bounding rectangle, the target scale range, and the region motion consistency. Isolated noise regions with too small an area, cloud regions with too large an area, and regions located at the image edge and lacking continuous motion relationships can be eliminated or marked as low-priority candidate regions during the candidate extraction stage.

[0081] like Figure 5 As shown, after obtaining candidate aerial target regions, the candidate target-level low-altitude scene discernibility is calculated for each candidate region. This candidate target-level low-altitude scene discernibility is not an evaluation of the overall image quality of a frame, but rather a measure of target discernibility calculated for each candidate target region and its surrounding background. For the i-th candidate region in the t-th frame, the local contrast C, high-frequency structure preservation G, image noise intensity N, background motion disturbance intensity B, and candidate region scale S are extracted, and the candidate target-level low-altitude scene discernibility Q is calculated.

[0082] The local contrast ratio C can be calculated based on the grayscale difference between the candidate region and its neighboring background. In one specific embodiment, a preset width is extended outward from the candidate region to obtain the candidate region's neighborhood, and the portion of the candidate region's neighborhood excluding the candidate region is taken as the local background region. The local contrast ratio can be expressed as: ; The average gray value of the candidate region, the average gray value of the neighboring background, and the standard deviation of the gray values ​​of the candidate region's neighbors are used to characterize the gray value difference between the target and the background and the degree of local background fluctuation, respectively. ε is a constant to prevent the denominator from being zero. The calculated local contrast can be normalized to a preset numerical range.

[0083] The high-frequency structure retention G is used to indicate whether the edges and details of the candidate region are clear. In one specific implementation, it can be calculated using Sobel gradient energy, Tenengrad gradient energy, or Laplacian variance. Taking the Sobel gradient as an example, the horizontal and vertical gradients of the candidate region can be calculated first, and the gradient magnitude can be obtained. The calculation formula is as follows: ; The high-frequency structure preservation amount G is then obtained based on the mean, high quantile, or proportion of pixels exceeding a preset gradient threshold within the candidate region. When motion blur, out-of-focus blur, or low-contrast degradation exist in the candidate region, the edge gradient energy decreases, and the high-frequency structure preservation amount G decreases accordingly.

[0084] The image noise intensity N can be estimated based on the flat sky-background region of the current frame. The high-pass filter residual can be calculated in the background region after excluding candidate target regions, and the noise intensity can be estimated based on the median absolute deviation, as shown in the following formula: ; Where R is the residual image obtained after high-pass filtering or small-window mean filtering of the background region of the current frame, and norm represents the normalization process. The higher the noise intensity, the more difficult it is to distinguish the candidate target edge from the background noise.

[0085] The background motion perturbation intensity B can be calculated based on the optical flow residual of the background region. In one specific implementation, the global motion of the background region in the current frame is first estimated, then the residual between the actual optical flow at each background sampling point and the predicted global motion optical flow is calculated, and the variance or mean of the residual is taken as the background motion perturbation intensity. If there is slight camera shake, cloud drift, or background texture movement, the background motion perturbation intensity increases.

[0086] The candidate region scale S can be represented by the area of ​​the bounding rectangle of the candidate region, the pixel area of ​​the candidate region, or the equivalent side length of the candidate region. Since low-altitude UAVs are small in scale during long-range imaging, excessively small scales can reduce recognition reliability. Therefore, a candidate region scale term is introduced when calculating the discriminability of candidate target-level low-altitude scenes. The discriminability of candidate target-level low-altitude scenes can be calculated using the following formula: ; Where S0 is the preset reference scale, and the local contrast C, high-frequency structure preservation G, image noise intensity N, background motion disturbance intensity B, and the ratio of candidate region scale to reference scale are all normalized. This formula shows that when the candidate target has high local contrast, clear edge structure, and a large scale, the candidate target-level low-altitude scene has high discernibility; when the current frame has strong noise, background motion disturbance, motion blur, or the target scale is small, the discernibility of the candidate target-level low-altitude scene decreases.

[0087] After obtaining the identifiability Q of the candidate target in the low-altitude scene, the image enhancement parameters for the current recognition cycle are determined based on Q and the enhancement adjustment parameters output from the previous recognition cycle. The image enhancement parameters include global image enhancement parameters and candidate region local enhancement parameters. The global image enhancement parameters improve the brightness, noise, and overall contrast of the entire frame; the candidate region local enhancement parameters improve the local edges, target contrast, and motion blur of the candidate target region.

[0088] In one specific implementation, periodic-level identifiability statistics, such as average identifiability, minimum identifiability, and the proportion of low-identifiability candidates, can be calculated first based on the identifiability of candidate targets in all candidate regions within the current identification period. The proportion of low-identifiability candidates can be expressed as: ; Where M is the total number of candidate regions in the current identification period, and T Q The indicator function is used to count the number of candidate regions below the low identifiability threshold. If the proportion of low identifiability candidates is high, it indicates that most candidate targets in the current recognition cycle are in a low-contrast, strongly disturbed, or small-scale state, and the enhancement intensity needs to be increased. If the average identifiability is high and the proportion of low identifiability candidates is low, it indicates that the current scene is identifiable well, and the enhancement intensity should be reduced to avoid over-enhancement leading to increased edge artifacts and false detections.

[0089] Global image enhancement parameters can include denoising intensity, brightness compensation intensity, strong light suppression intensity, and global contrast adjustment intensity. The denoising intensity can be determined based on the image noise intensity and the proportion of periodically low-discrimination candidates; the brightness compensation intensity can be determined based on the image brightness average and the proportion of low-brightness areas; the strong light suppression intensity can be determined based on the proportion of high-brightness saturated areas; and the global contrast adjustment intensity can be determined based on the overall frame grayscale distribution and the average contrast of the candidate regions.

[0090] The local enhancement parameters for the candidate region may include local contrast enhancement intensity, motion blur suppression intensity, candidate region scale normalization ratio, and edge artifact suppression intensity. In one specific implementation, the local enhancement parameters for the candidate region can be determined according to the following rules: when the discernibility of the candidate target-level low-altitude scene is lower than a first threshold, the candidate region is determined to be in a low discernibility state, and the local contrast enhancement intensity, denoising intensity, and motion blur suppression intensity are increased; when the discernibility of the candidate target-level low-altitude scene is between the first and second thresholds, the candidate region is determined to be in a medium discernibility state, and a medium enhancement intensity is adopted; when the discernibility of the candidate target-level low-altitude scene is higher than the second threshold, the candidate region is determined to be in a high discernibility state, the local enhancement intensity is reduced, and the edge artifact suppression intensity is increased. In one example, the first threshold can be set to 0.45, and the second threshold can be set to 0.65. The specific thresholds can be adjusted according to device parameters and application scenarios.

[0091] To avoid drastic changes in enhancement parameters between adjacent recognition cycles, this embodiment can also employ a smooth update method to determine the enhancement parameters for the current recognition cycle. For example, the local contrast enhancement intensity of the candidate region can be updated according to the following formula: ; In this process, the local contrast enhancement intensity actually used in the current recognition cycle is jointly determined by the local contrast enhancement intensity of the previous recognition cycle and the candidate enhancement intensity calculated in the current cycle, where ρ is a smoothing coefficient. This method can reduce abrupt changes in enhancement parameters caused by illumination variations or transient noise.

[0092] Subsequently, global enhancement processing is performed on consecutive video image frames within the current recognition period. Global enhancement processing may include the following steps: first, noise suppression is applied to the entire frame to reduce random noise in the background area; then, weak light enhancement or strong light suppression is performed based on the image brightness distribution; finally, appropriate contrast adjustment is applied to the entire frame. The purpose of global enhancement processing is not to forcibly highlight all edges, but rather to improve the overall image quality first, providing a more stable image foundation for subsequent local enhancement of candidate regions.

[0093] After global enhancement, local enhancement processing is performed on each candidate aerial target region. Local enhancement processing may include candidate region cropping, neighborhood expansion, local contrast enhancement, motion blur suppression, scale normalization, and edge artifact suppression. Specifically, a local enhancement region is obtained by expanding outwards from the candidate region by a preset ratio. The expansion ratio can be set to 0.5 to 1.5 times the width and height of the candidate box to ensure that background information from the target's neighborhood is utilized during local enhancement. Subsequently, local contrast enhancement and edge enhancement are performed on the locally enhanced region to make the target boundaries and local structures clearer. For candidate regions with significant motion blur, the blur direction can be estimated based on the motion direction of the candidate target in consecutive frames, and directional deblurring or deconvolution is used to suppress target edge trailing. For candidate regions with excessively small scales, scale normalization is performed to scale the candidate region to an input scale that is easily recognized by the detection model. For candidate regions with high discernibility in low-altitude scenes at the candidate target level, the intensity of local contrast enhancement is reduced, and edge artifact suppression is used to avoid generating false edges.

[0094] exist Figure 5 In one candidate target example shown, the pre-enhancement simulated image frame suffers from low contrast, motion blur, noise, and small target scale. The candidate target-level low-altitude scene discernibility example value Q is 0.36, and the detection confidence example value p is 0.61. Based on the low discernibility state of this candidate region, the system increases the local contrast enhancement intensity, denoising intensity, and motion blur suppression intensity. After enhancement, the edge structure of the same candidate target region is clearer, background noise is suppressed, and the candidate target-level low-altitude scene discernibility example value increases to Q of 0.49, and the detection confidence example value increases to p of 0.72. This example is used to illustrate the correspondence between the candidate target-level low-altitude scene discernibility and image enhancement processing, and does not limit the invention to the above values.

[0095] The enhanced consecutive video image frames are input into a deep learning object detection model. The deep learning object detection model can output detection boxes, detection confidence scores, and visual features of candidate aerial objects. Because the local enhancement of the candidate regions has performed contrast, edge, and blur compensation for low-discernibility targets, the detection model can obtain more stable candidate object detection results in scenes with distant small targets, low-contrast targets, and targets with background disturbances.

[0096] It should be noted that the simulation data construction process, imaging degradation parameters, target scale range, and enhancement parameter thresholds in this embodiment can all be adjusted according to different low-altitude security application scenarios. For example, in airport airspace protection scenarios, the proportion of long-distance small-scale UAV targets and bird interference targets can be increased; in urban low-altitude monitoring scenarios, complex building backgrounds, reflective points, and cloud drift interference can be added; in coastal or border patrol scenarios, large-scale sky backgrounds, kites, and floating object interference can be added. This invention does not limit the specific number and background form of the simulation scenarios. As long as the simulated video data can form a low-altitude monitoring sequence containing low-altitude targets, continuous frame motion, image degradation, and interference targets, it can be used to execute the identification and processing flow of this invention.

[0097] Through the processing method of this embodiment, the simulated video data not only includes the UAV target itself, but also various non-UAV interference objects, different image quality degradation conditions, and continuous frame motion relationships. Simultaneously, the discernibility of candidate target-level low-altitude scenes is used to guide global image enhancement and local enhancement of candidate regions, establishing a correspondence between image enhancement processing and subsequent detection and recognition processes. Therefore, this embodiment provides a data and processing foundation for the detection confidence improvement, false alarm suppression, tracking continuity improvement, and reinforcement learning closed-loop adjustment effect verification in subsequent embodiments.

[0098] Example 3: Based on the low-altitude monitoring simulation video data constructed in Example 2, this example quantitatively verifies the image enhancement effect, detection confidence improvement effect, rotor-rigid body collaborative feature differentiation effect, interference rebuttal evidence suppression effect, and complete recognition performance of the method of the present invention. This example illustrates that the present invention does not only complete UAV recognition through a deep learning detection model, but also improves the stability of UAV recognition and the reliability of early warning in complex low-altitude scenarios through candidate target-level low-altitude scene identifiability, adaptive image enhancement, rotor-rigid body collaborative analysis, interference evidence rebuttal evidence fusion, and cross-recognition cycle parameter adjustment.

[0099] This embodiment uses the low-altitude monitoring simulation video data generated in Embodiment 2. The simulation video resolution is 1920×1080, the frame rate is 25 fps, and each 16 consecutive frames constitute a recognition cycle, with a step size of 8 frames between adjacent recognition cycles. The simulation scenarios include six types: A) clear sky with light clouds, B) thin fog with low contrast, C) evening backlight, D) distant mountains / buildings with complex background, E) slight camera shake, and F) cloud drift interference. Each scenario is injected with UAV targets and non-UAV interference targets. The UAV targets include multi-rotor UAVs and small fixed-wing UAVs, and the interference targets include birds, kites, cloud edge drifting targets, and floating objects.

[0100] To ensure fairness in the experimental comparison, all comparison methods in this embodiment use the same simulated video sequence, the same candidate target extraction method, the same deep learning target detection model, and the same detection model weights. The differences between the comparison methods lie only in whether to use enhancement processing driven by the low-altitude scene discernibility of candidate targets, whether to use rotor-rigid body collaborative features, whether to use bird / kite / cloud interference to provide counter-evidence, and whether to use a cross-recognition cycle parameter adjustment mechanism based on the reinforcement learning strategy model output.

[0101] I. Composition of the Simulation Dataset The simulation dataset used in this embodiment is shown in Table 1. This dataset covers common imaging conditions in low-altitude security, such as clear skies, low contrast, backlighting, complex backgrounds, camera shake, and cloud interference, and is used to verify the adaptability of the present invention under different image quality and interference conditions.

[0102] Table 1: Composition of Simulated Low-Altitude Monitoring Video Dataset

[0103] As shown in Table 1, this embodiment constructed a total of 6 simulated video sequences with a total duration of 540 seconds and a total of 13,500 frames, including 30 UAV trajectories and 46 non-UAV interference trajectories. Among them, the average Q before enhancement for scenes B, C, D, E, and F is significantly lower than that for scene A, indicating that these scenes have adverse imaging factors such as low contrast, backlighting, complex backgrounds, camera shake, or cloud drift, which can be used to verify the effectiveness of candidate target-level low-altitude scene discernibility-driven enhancement. The target scale distribution is shown in Table 2.

[0104] Table 2: Scale Distribution of Candidate Targets

[0105] As shown in Table 2, the candidate targets in this embodiment are mainly small-scale targets, with targets ranging from 9 to 25 pixels accounting for 69%. This scale distribution is consistent with the imaging characteristics of "small, weak, and slow" long-distance UAV targets in low-altitude security videos, and also reflects the technical value of this invention in enhancing and identifying small-scale targets.

[0106] II. Comparison Experiment of Detection Confidence Before and After Adaptive Enhancement This embodiment first verifies the image enhancement effect driven by the discernibility of candidate target-level low-altitude scenes. Specifically, for the same set of simulated video sequences, without using the adaptive enhancement of this invention, the original image frames or image frames that have only undergone basic preprocessing are input into a deep learning target detection model, and the average detection confidence of real UAV targets is calculated. Then, based on the candidate target-level low-altitude scene discernibility calculation results, global enhancement is performed on consecutive video image frames, and local enhancement is performed on candidate aerial target regions. These are then input into the same deep learning target detection model, and the average detection confidence after enhancement is calculated.

[0107] During statistical analysis, only detection bounding boxes that successfully match real drone targets are included in the calculation. A detection bounding box is considered to have successfully matched a real drone target if the intersection-union ratio (IU) between the detection bounding box and the ground truth bounding box is greater than a preset threshold. Detection bounding boxes that do not match are not included in the average detection confidence score to avoid background false positives interfering with the mean detection confidence score.

[0108] The comparison results of detection confidence before and after adaptive enhancement are shown in Table 3. The data in Table 3 are... Figure 6 The bar charts shown correspond perfectly.

[0109] Table 3: Comparison of Detection Confidence Before and After Adaptive Enhancement in Simulation Scenarios

[0110] like Figure 6As shown in Table 3, the average detection confidence of real UAV targets was improved in all six simulation scenarios after applying the adaptive enhancement processing of this invention. The improvement value for the clear sky with light clouds scenario was 0.03, a relatively small increase. This is because the original image quality is good in this scenario, and the low-altitude scene at the candidate target level is highly identifiable, so the system does not excessively increase the local enhancement intensity. The improvement values ​​for the foggy low-contrast scenario, the evening backlight scenario, and the complex background scenario with distant mountains / buildings were all 0.13, indicating that even with weak target edges, insufficient local contrast, or complex background textures, the identifiability of the low-altitude scene at the candidate target level can effectively trigger stronger local contrast enhancement, denoising, and motion blur suppression. The improvement value for the slightly shaken camera scenario was 0.11, indicating that even with strong background motion disturbances, the global enhancement and local enhancement of candidate regions of this invention can still improve the detection model's response to real UAV targets. The improvement value for the cloud drift interference scenario was 0.09, indicating that even with the presence of cloud edges and floating object interference, this invention can still improve the detectability of real UAV targets.

[0111] To further illustrate the relationship between the discernibility of candidate targets in low-altitude scenes and the enhancement effect, this embodiment also statistically analyzed the average Q-value before and after enhancement, the mean detection confidence, and the side effects of enhancement. The results are shown in Table 4.

[0112] Table 4: Statistics on the Distinguishingness and Side Effects of Candidate Targets Before and After Adaptive Enhancement

[0113] Table 4 shows that in low-discriminability scenes such as B, C, D, E, and F, the average Q is significantly improved after enhancement, and the average detection confidence also increases simultaneously. In scene A, the average Q only increases from 0.66 to 0.70 after enhancement, and the detection confidence only increases from 0.81 to 0.84. This indicates that when the scene itself has high discriminability, this invention does not perform excessive enhancement, thus avoiding over-enhancement of clear targets. The edge artifact rate statistics also show that although this invention improves the enhancement intensity in low-discriminability scenes, the edge artifact rate remains at a low level, and no large number of false edges are generated due to enhancement processing.

[0114] III. Experiment on the ability to distinguish rotor-rigid body cooperative characteristics To verify the ability of rotor-rigid-body cooperative features to distinguish between UAV targets and interference targets, this embodiment statistically analyzes the performance of multi-rotor UAVs, small fixed-wing UAVs, birds, kites, and cloud edge targets in terms of edge micro-periodic response, main body attitude angle change, center of mass trajectory jitter, and integrated rotor-rigid-body cooperative features. The statistical results are shown in Table 5.

[0115] Table 5: Statistical Table of Rotor-Rigid Body Cooperative Characteristics for Different Target Types

[0116] Table 5 shows that multi-rotor UAVs have a higher proportion of micro-periodic frequency band energy, and relatively lower main body attitude angle variance and center-of-mass trajectory jitter, resulting in a higher average R. Small fixed-wing UAVs have weaker micro-periodic rotor or propeller responses than multi-rotor UAVs, but their main body attitude and center-of-mass trajectory are more stable, so their average R is still higher than that of birds, kites, and cloud edge targets. Although birds may exhibit periodic wing oscillations, their main body attitude angle variance and center-of-mass trajectory jitter are large, resulting in a significantly lower average R than UAV targets. Kites and cloud edge targets have a lower proportion of micro-periodic frequency band energy, and their boundary morphology changes or trajectory drifts do not conform to the rigid body motion characteristics of UAVs, resulting in an even lower average R. These results indicate that rotor-rigid body cooperative characteristics can provide a positive discrimination criterion for UAV targets, distinguishing them from ordinary motion trajectory characteristics.

[0117] IV. Comparison Experiment of False Alarm Numbers of Different Identification Methods After completing the detection confidence level and rotor-rigid body cooperative feature statistics, this embodiment further conducts a comparative experiment on the number of false alarms. This experiment corresponds to... Figure 7 To demonstrate the contribution of different technical features to false alarm suppression, this embodiment sets up six sets of comparison methods.

[0118] The first group uses the basic YOLO detection method. This method only uses a deep learning target detection model for UAV detection, without using candidate target-level low-altitude scene discernibility-driven enhancement, rotor-rigid body cooperative features, interference-based counter-evidence, or cross-recognition cycle parameter adjustment.

[0119] The second group uses YOLO with global enhancement. This method performs uniform global enhancement on the entire frame before inputting it into the deep learning object detection model, but does not perform local enhancement on candidate regions, nor does it employ rotor-rigid body collaborative features, interference-based counter-evidence, or cross-recognition cycle parameter adjustment.

[0120] The third group is the YOLO plus candidate region enhancement method. This method enhances the candidate aerial target region locally on the basis of global enhancement, but does not use rotor-rigid body cooperative features, interference to prove evidence, or cross-recognition cycle parameter adjustment.

[0121] The fourth group involves methods that incorporate rotor-rigid body cooperative features R. This method, based on candidate region enhancement, further calculates rotor-rigid body cooperative features and uses them as a positive criterion for UAV target discrimination.

[0122] The fifth group involves adding interference evidence E. Building upon the fourth group, this method further generates interference evidence from birds, kites, and clouds, and uses this interference evidence to suppress non-drone targets through counter-evidence.

[0123] The sixth group represents the complete solution of this invention. This method simultaneously employs candidate target-level low-altitude scene discernibility-driven enhancement, rotor-rigid body collaborative features, interference-based counter-evidence, and a cross-recognition cycle parameter adjustment mechanism based on the output of the reinforcement learning strategy model.

[0124] In this embodiment, the number of false alarms is defined as the number of events in which a non-UAV target or background interference area is identified as a UAV target by the system and triggers a warning output. To avoid statistical bias caused by accidental false detections in a single frame, a false alarm event is only counted when a candidate target is identified as a UAV target in two or more consecutive recognition cycles and actually triggers a warning message. For alarm outputs triggered continuously by the same interference target within a short period of time, they are combined and counted as a single continuous false alarm event. Finally, the number of false alarm events for each method under the same simulated video duration is converted into the number of false alarms per 10 minutes.

[0125] Table 6 shows a complete performance comparison of different recognition methods, where the data in the false alarm count column is compared with... Figure 7 The bar charts shown correspond perfectly.

[0126] Table 6: Comparison of Overall Performance of Different Recognition Methods

[0127] From Table 6 and Figure 7 The results show that the basic YOLO detection method generates 6.8 false alarms per 10 minutes, indicating that in complex low-altitude scenes, relying solely on detection confidence can easily misclassify birds, kites, cloud edges, or floating objects as drone targets. The YOLO plus global augmentation method generates 6.5 false alarms per 10 minutes, only slightly lower than the basic YOLO detection method. This suggests that while frame-wide augmentation can improve some low-quality images, it may also simultaneously enhance background edges and interfering targets, thus limiting its effectiveness in controlling false alarms. The YOLO plus candidate region augmentation method generates 6.1 false alarms per 10 minutes, indicating that candidate region augmentation can improve the visibility of real targets, but augmentation alone is insufficient to effectively distinguish drones from low-altitude interfering targets.

[0128] After adding the rotor-rigid body cooperative feature R, the number of false alarms decreased to 4.2 per 10 minutes, indicating that this feature can distinguish between interfering targets such as birds and kites by utilizing the characteristics of the UAV target in terms of the equivalent micro-periodic response of the rotor or propeller, the stability of the main body attitude, and the smoothness of the center of mass trajectory. Further adding interference counter-evidence E reduced the number of false alarms to 2.5 per 10 minutes, indicating that bird interference evidence, kite interference evidence, and cloud interference evidence can suppress false warnings when the detection confidence is high but the target exhibits non-UAV characteristics. The complete scheme of this invention further reduces the number of false alarms to 1.9 per 10 minutes, indicating that the enhancement parameters, target association parameters, micro-periodic response frequency band adjustment parameters, interference evidence weights, and decision margin thresholds output by the reinforcement learning strategy model can play a dynamic compensation role in continuous recognition cycles, thereby further reducing false alarms.

[0129] Table 6 also shows that, compared to the basic YOLO detection method, the complete solution of this invention improves the recall rate from 78.6% to 88.9%, the precision from 75.4% to 92.1%, the F1 score from 76.9% to 90.5%, the average alarm latency from 0.78s to 0.49s, and the number of tracking interruptions from 2.4 per target to 1.0 per target. These results demonstrate that this invention not only reduces the number of false alarms but also simultaneously improves the detection capability and continuous identification stability of UAV targets.

[0130] It is worth noting that after adding the interfering evidence E, the recall rate slightly decreased from 85.5% to 84.8%, but the precision increased from 84.7% to 90.6%, and the number of false alarms decreased from 4.2 per 10 minutes to 2.5. This result is consistent with the working mechanism of the interfering evidence mechanism: in some boundary samples, the interfering evidence may have a certain inhibitory effect on low-confidence UAV targets, but overall it can significantly reduce false alarms caused by non-UAV interference targets. The complete solution of this invention further restores and improves the recall rate by adjusting parameters across the recognition cycle, raising the recall rate to 88.9%, indicating that the reinforcement learning strategy model can achieve a more reasonable dynamic balance between false alarm suppression and missed detection control.

[0131] V. Comparison Experiment of False Alarm Rates for Typical Interference Types To further illustrate the suppressive effect of interference-based counter-evidence on different types of interference targets, this embodiment statistically analyzes the false alarm rates of three typical interference targets: birds, kites, and cloud edges. The statistical results are shown in Table 7.

[0132] Table 7: Comparison of False Alarm Rates for Typical Interference Types

[0133] Table 7 shows that when relying solely on detection confidence, birds, kites, and cloud edges are all prone to false alarms. Kites have the highest false alarm rate at 21.2%, mainly because kites, when imaged at long distances, have small-scale outlines similar to drones and may exhibit relatively stable hovering characteristics when swaying in the wind. Birds have a false alarm rate of 18.5%, primarily because the local appearance and short-term movement trajectory of birds may resemble small drones. Cloud edges have a false alarm rate of 13.8%, mainly because drifting areas at cloud edges may appear as slowly moving candidate targets against a low-texture background.

[0134] After adding the rotor-rigid body cooperative feature R, the false alarm rates for birds, kites, and cloud edges all decreased, indicating that the rotor-rigid body cooperative feature can suppress some interference targets with significant non-rigid body morphological changes, unstable body attitude, or large trajectory jitter. Further adding interference counter-evidence E, the false alarm rates for the three types of interference targets continued to decrease, indicating that constructing interference evidence for birds, kites, and clouds respectively has a significant counter-evidence effect. The complete scheme of this invention further reduces the false alarm rate for birds to 3.7%, the false alarm rate for kites to 4.8%, and the false alarm rate for cloud edges to 4.1%, indicating that the adjustment of evidence weight and decision margin threshold across recognition periods can further adapt to changes in the frequency and intensity of different interference types.

[0135] VI. Experiment on the Moderating Effect of Reinforcement Learning Across Recognition Cycles To verify the moderating effect of the reinforcement learning strategy model in continuous monitoring scenarios, this embodiment statistically analyzed the changes in average Q, average interference evidence strength, enhancement actions, target association window, micro-periodic response band, and decision margin threshold at different recognition cycle stages. The statistical results are shown in Table 8.

[0136] Table 8: Statistics on the Adjustment Process Across Recognition Cycles in Reinforcement Learning

[0137] As shown in Table 8, during periods 1-20, the average Q was 0.43, indicating a moderately low level of scene discernibility. The system employed a moderate enhancement strategy, but short-term false negatives still occurred. During periods 21-40, the average Q decreased to 0.36, indicating a further reduction in image quality and target discernibility. The reinforcement learning strategy model output improved denoising and deblurring actions, while the target association window was adjusted from 5 frames to 7 frames, and the micro-periodic response band was expanded. This enabled short-term blurred or short-term false negative targets to be more stably associated in consecutive frames, thus improving recall. During periods 41-60, the average Q increased to 0.55, but the average interference evidence strength rose to 0.48, indicating a stronger influence from interfering targets at this stage. The system correspondingly reduced the enhancement strength, increased the weight of counter-evidence, and raised the decision margin threshold, resulting in a decrease in false alarms. During periods 61-80, the average Q and interference evidence strength tended to stabilize. The system employed moderate enhancement, a moderate target association window, and a moderate decision margin threshold, maintaining stable recognition performance.

[0138] The results demonstrate that the reinforcement learning strategy model does not simply replace the detection model for classification. Instead, it dynamically adjusts the enhancement parameters, target association parameters, micro-periodic response band, and decision threshold based on changes in candidate target-level discriminability, detection confidence, rotor-rigid body cooperative features, and interference evidence throughout continuous recognition cycles. This cross-cycle adjustment mechanism improves recall in low-discriminability scenarios, reduces false alarms in scenarios with strong interference evidence, and maintains stable recognition output when the scene state tends to stabilize.

[0139] Combine Tables 3 to 8 and Figure 6 and Figure 7 It can be seen that the present invention has at least the following technical effects in low-altitude monitoring simulation scenarios.

[0140] Improve the confidence of drone detection in low-discernibility scenarios. (See Table 3 and...) Figure 6 It can be seen that in scenes with low contrast due to light fog, backlighting at dusk, complex backgrounds such as distant mountains / buildings, and slight camera shake, the average detection confidence after enhancement is significantly improved compared to before enhancement. This indicates that the discernibility of candidate targets at low altitudes can effectively reflect the imaging quality of the target area and guide the image enhancement module to perform more targeted enhancement on low-quality candidate areas.

[0141] To avoid false alarms caused by simple global enhancement, Table 6 shows that the number of false alarms with the YOLO plus global enhancement method only decreased from 6.8 to 6.5 per 10 minutes, a limited reduction. The YOLO plus candidate region enhancement method further reduced the number to 6.1, indicating that local enhancement of candidate regions is more suitable for low-altitude, small-target scenes than uniform enhancement of the entire frame. Furthermore, Table 4 shows that this invention maintains a low edge artifact rate after enhancement, indicating that candidate target-level discernibility is used not only to improve enhancement intensity but also to suppress over-enhancement when discernibility is high.

[0142] The ability to distinguish UAVs from interfering targets is improved by utilizing rotor-rigid-body cooperative features. Table 5 shows that the average R-value of multi-rotor UAVs and small fixed-wing UAVs is significantly higher than that of birds, kites, and cloud edge targets. Table 6 shows that after adding the rotor-rigid-body cooperative feature R-value, the number of false alarms decreased from 6.1 times per 10 minutes to 4.2 times, indicating that this feature can effectively distinguish targets with similar appearances or similar short-term movements, such as birds and kites.

[0143] By using interference as a counter-evidence, false alarms caused by typical low-altitude interference can be reduced. As shown in Table 7, the complete solution of this invention can reduce the false alarm rate for birds to 3.7%, for kites to 4.8%, and for cloud edges to 4.1%. This demonstrates that bird interference evidence, kite interference evidence, and cloud interference evidence can effectively suppress different types of interference targets, rather than relying solely on a single detection confidence level or a single motion trajectory feature.

[0144] The stability of continuous monitoring is improved by adjusting parameters across recognition cycles. As shown in Table 8, the reinforcement learning strategy model can increase enhancement and association strength when the average Q decreases, increase the rebuttal weight and decision margin threshold when interference evidence is enhanced, and maintain moderate adjustment parameters when the scene is stable. This adjustment method enables the system to maintain relatively stable recognition performance under changes in illumination, camera shake, target scale, and interference type.

[0145] In summary, this embodiment verifies the complete technical chain of the present invention through simulation data. Table 3 and Figure 6 This verifies the effect of candidate target-level low-altitude scene identifiability-driven enhancement on improving detection confidence; Table 6 and Figure 7The results verified the differences in the number of false alarms among different recognition methods; Table 5 verified the target discrimination capability of the rotor-rigid body collaborative features; Table 7 verified the ability of interference-based counter-evidence to suppress interference from birds, kites, and cloud edges; Table 8 verified the dynamic adjustment effect of the reinforcement learning strategy model across recognition cycles. These experimental results collectively demonstrate that the present invention can improve the stability of UAV target detection in complex low-altitude security scenarios, reduce false alarms from non-UAV targets, and decrease short-term missed detections and tracking interruptions, thus exhibiting better low-altitude scene adaptability and recognition robustness compared to conventional video UAV detection methods.

[0146] The above embodiments merely illustrate implementation methods of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for identifying unmanned aerial vehicles (UAVs) for low-altitude security, characterized in that: Includes the following steps: S1: Acquire continuous video image frames of the low-altitude monitoring area, take a preset number of continuous video image frames as a recognition period, and extract candidate aerial target areas from the continuous video image frames of the current recognition period. S2: According to the first The first frame The image sharpness, local contrast, noise intensity, background motion disturbance, and target scale of each candidate aerial target region are used to calculate the low-altitude scene discernibility of the candidate targets. And based on the candidate target level low-altitude scene identifiability corresponding to each candidate aerial target area. In addition to the enhancement adjustment parameters output from the previous recognition cycle, the global image enhancement parameters and candidate region enhancement parameters for the current recognition cycle are determined, and global enhancement is performed on continuous video image frames, while local enhancement is performed on candidate aerial target regions. S3: Input the enhanced image frame into the deep learning object detection model to obtain the detection confidence of candidate aerial targets. and target visual features; S4: Based on the target association parameters output from the previous recognition cycle, perform target association on the same candidate aerial target in consecutive frames, and calculate the rotor-rigid body cooperative features based on the temporal changes of the edge gradient of the candidate aerial target region, the changes of the main body attitude, and the changes of the center of mass trajectory. ; S5: Generate evidence of bird interference based on target visual features and continuous frame motion features. Evidence of kite interference Evidence of cloud interference ; S6: Improve the identifiability of candidate target-level low-altitude scenes Detection confidence Rotor-rigid body cooperative characteristics Evidence of interference from birds Evidence of kite interference Evidence of cloud interference Input the reinforcement learning policy model to obtain the decision adjustment parameters, interference evidence weights and decision margin thresholds for the current recognition cycle, as well as the enhancement adjustment parameters, target association parameters and micro-cycle response frequency band adjustment parameters for the next recognition cycle; S7: Calculate the interference rebuttal judgment margin for candidate aerial targets based on the judgment adjustment parameters and interference evidence weights. and in the margin of interference rebuttal judgment When the threshold for judgment is exceeded, the corresponding candidate aerial target is judged as a drone target, and the drone identification result and warning information are output.

2. The UAV identification method for low-altitude security according to claim 1, characterized in that: The formula for calculating the identifiability of candidate target-level low-altitude scenes in step S2 is as follows: ； in, For the first The first frame Local contrast of candidate aerial target regions To maintain the high-frequency structure of candidate airborne target regions. For the first Image noise intensity of the frame For the first The intensity of background motion perturbation in the frame. The scale is the area of ​​the candidate aerial target. For the preset reference scale; among which, , , , and All are dimensionless quantities normalized to a preset numerical range.

3. The UAV identification method for low-altitude security according to claim 1, characterized in that: In step S2, when the candidate target level low-altitude scene is identifiable... When the noise level is reduced, increase the denoising intensity, motion blur suppression intensity, or local contrast enhancement intensity of the candidate region; when the candidate target-level low-altitude scene discernibility is reduced... When increasing the intensity, the image enhancement intensity of the candidate region is reduced to decrease edge artifacts caused by over-enhancement.

4. The UAV identification method for low-altitude security according to claim 1, characterized in that: Rotor-rigid body cooperative features in step S4 The calculation formula is: ； in, The edge gradient time series of candidate aerial target regions at frequency The energy at the location, For the first The micro-periodic response frequency band of the rotor or propeller region corresponding to each candidate aerial target For the candidate frequency band set, This is the main orientation angle sequence of the candidate aerial targets. The variance of the main orientation angle, This represents the jitter in the candidate aerial target's centroid trajectory relative to the predicted trajectory. To prevent constants with a denominator of zero; The micro-periodic response of the rotor or propeller region includes the true periodic response and the equivalent periodic response formed by motion blur or video sampling aliasing.

5. The UAV identification method for low-altitude security according to claim 4, characterized in that: Micro-periodic response frequency band of rotor or propeller region The frequency band adjustment parameters are adaptively determined based on the candidate aerial target region scale, video frame rate, main frequency distribution of historical edge gradient time series, micro-period response frequency band adjustment parameters output in the previous recognition cycle, and preset UAV rotor sample frequency band range.

6. The UAV identification method for low-altitude security according to claim 1, characterized in that: Bird interference evidence in step S5 Generated based on wing periodic oscillation characteristics, non-rigid morphological changes, and trajectory non-smoothness characteristics; evidence of kite interference. Generated based on contour flexibility deformation characteristics, wind-driven swaying characteristics, and traction trajectory constraint characteristics; cloud interference evidence. It is generated based on low-texture drift features, edge diffusion features, and continuous expansion features of regional morphology.

7. The UAV identification method for low-altitude security according to claim 1, characterized in that: In step S6, the reinforcement learning strategy model is a deep Q-network model. The state variables of the deep Q-network model include the identifiability of the candidate target-level low-altitude scene, detection confidence, rotor-rigid body cooperative features, and various interference evidence; the action variables include enhancement adjustment action, target association window adjustment action, rotor or propeller region micro-periodic response frequency band adjustment action, interference evidence weight adjustment action, and decision margin threshold adjustment action. The reward amount is determined based on at least two of the following: identification confirmation result, false detection result, missed detection result, target tracking continuity, and identification delay.

8. The UAV identification method for low-altitude security according to claim 1, characterized in that: Error proof margin in step S7 The calculation formula is: ； in, For the first The first frame Detection confidence of candidate aerial targets, For the first The first frame Rotor-rigid body cooperative characteristics of candidate aerial targets For the first The first frame The identifiability of candidate aerial targets at the low-altitude scene level; The index is the recognizability adjustment index. To smooth out interference evidence parameters, , , These are the evidence weights for interference from birds, kites, and clouds, respectively. , , , , The decision margin threshold is obtained by the output of the reinforcement learning policy model or by the action mapping selected by the reinforcement learning policy model; To prevent constants with a denominator of zero; , and All are dimensionless quantities normalized to a preset numerical range; when When the threshold is greater than the judgment margin threshold, the first... The first frame The candidate aerial targets were identified as drone targets.

9. A drone identification method for low-altitude security according to claim 1, characterized in that: The method further includes: comparing the output UAV identification result with subsequent tracking results, manual confirmation results, or feedback results from a low-altitude security platform to generate identification feedback information, and using the identification feedback information to update the reward amount of the reinforcement learning strategy model.

10. A drone identification system for low-altitude security, characterized in that: Includes the following modules: T1: Image sensing module, used to acquire continuous video image frames of the low-altitude monitoring area; T2: Candidate target extraction module, used to extract candidate aerial target regions from consecutive video image frames; T3: Scene identifiability calculation module, used to calculate the identifiability of candidate target-level low-altitude scenes; T4: Image enhancement module, used to perform global enhancement on continuous video image frames and local enhancement on candidate aerial target regions based on the identifiability of candidate target-level low-altitude scenes and the enhancement adjustment parameters output from the previous recognition cycle. T5: Deep learning detection module, used to output the detection confidence and visual features of candidate aerial targets; T6: Rotor-rigid body cooperative analysis module, used to calculate the rotor-rigid body cooperative characteristics of candidate aerial targets; T7: Interference Evidence Generation Module, used to generate interference evidence from birds, kites, and clouds; T8: Reinforcement learning strategy module, used to output the decision adjustment parameters, interference evidence weights and decision margin thresholds for the current recognition cycle, as well as the enhancement adjustment parameters, target association parameters and micro-cycle response frequency band adjustment parameters for the next recognition cycle; T9: Fusion judgment and early warning module, used to calculate the interference counter-evidence judgment margin, and output UAV identification results and early warning information; The image enhancement module is deployed in a programmable logic processing unit or an image signal processor, while the deep learning detection module, rotor-rigid body collaborative analysis module, and reinforcement learning strategy module are deployed in an intelligent computing unit. The programmable logic processing unit includes an FPGA or CPLD, and the intelligent computing unit includes an AI chip, a SOC chip, a GPU, an NPU, or an ASIC.

Images