An unmanned aerial vehicle recognition system and method based on artificial intelligence

Abstract

The application discloses an unmanned aerial vehicle recognition system and method based on artificial intelligence, target-background bidirectional separation processing is applied to a monitoring area image sequence, aircraft contour boundaries and environmental interference factors are extracted, and a visual detection benchmark corresponding rule is established; a multi-scale scanning analysis is used to generate a recognition response matrix, a rotor flicker feature sequence is captured, and an identification oscillation period is verified through a harmonic law; a flight trajectory is divided into motion domains, a flight steady state scale is established by monitoring a motion domain transition process, an optimal capture instant is selected based on the flight steady state scale, a stable recognition period is calibrated through a signal-to-noise ratio evolution curve; a dynamic determination threshold is subjected to deviation degree analysis with the rotor flicker feature sequence, non-target interference sources are detected, false alarms are screened out, and an unmanned aerial vehicle recognition confirmation result is output, so that unmanned aerial vehicle targets can be accurately recognized in a complex environment, and the false detection rate and the missed detection rate are effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle monitoring and identification, in particular to an unmanned aerial vehicle identification system and method based on artificial intelligence. BACKGROUND

[0002] With the rapid development and wide application of unmanned aerial vehicle technology, the popularity of unmanned aerial vehicles in various fields is continuously increasing, but at the same time it also brings serious challenges to airspace safety management. Illegal unmanned aerial vehicles frequently invade key protected areas such as airports, important facilities and core areas, posing a major threat to public safety. Therefore, it is an urgent need for airspace safety management to establish an efficient and accurate unmanned aerial vehicle identification and early warning mechanism.

[0003] Existing unmanned aerial vehicle identification technologies mainly include radar detection, radio spectrum monitoring and visual identification methods. However, these methods have obvious limitations in complex environments. Radar detection is difficult to identify small low-altitude unmanned aerial vehicles and is easily affected by terrain obstructions, radio monitoring relies on communication signals between unmanned aerial vehicles and ground stations but is ineffective for autonomous flying unmanned aerial vehicles, and traditional visual identification methods have insufficient recognition accuracy in conditions of strong background interference, large changes in light, and small target size. In particular, in the presence of environmental interference factors such as moving clouds, bird flight and tree swaying, existing methods have difficulty in accurately distinguishing unmanned aerial vehicle targets from interference objects, resulting in high false detection and missed detection rates. In addition, existing technologies lack deep mining of the unique motion characteristics and rotor flicker characteristics of unmanned aerial vehicles, and cannot fully utilize the dynamic information in the flight process of unmanned aerial vehicles, limiting further improvement of identification performance. Therefore, an intelligent unmanned aerial vehicle identification method is urgently needed. SUMMARY

[0004] The present application provides an unmanned aerial vehicle identification system and method based on artificial intelligence, aiming to accurately extract aircraft contour features through target-background bidirectional separation processing, capture the unique periodic optical features of unmanned aerial vehicles using rotor flicker feature sequences, establish a flight steady-state scale combined with motion domain analysis of flight trajectories, and calibrate the optimal identification period through signal-to-noise ratio evolution curves, ultimately achieving high-precision identification of unmanned aerial vehicle targets in complex environments and false alarm elimination, providing reliable technical support for airspace safety supervision.

[0005] The present application provides an unmanned aerial vehicle identification system and method based on artificial intelligence, aiming to accurately extract aircraft contour features through target-background bidirectional separation processing, capture the unique periodic optical features of unmanned aerial vehicles using rotor flicker feature sequences, establish a flight steady-state scale combined with motion domain analysis of flight trajectories, and calibrate the optimal identification period through signal-to-noise ratio evolution curves, ultimately achieving high-precision identification of unmanned aerial vehicle targets in complex environments and false alarm elimination, providing reliable technical support for airspace safety supervision.

[0006] An image sequence of a monitored area is obtained, target-background bidirectional separation processing is applied to the image sequence to generate aircraft contour boundaries and environmental interference factors, and a visual detection benchmark corresponding rule is created according to the aircraft contour boundaries and the environmental interference factors.

[0007] The visual detection benchmark corresponding rule is used to search a candidate target range, multi-scale scanning analysis is applied to the candidate target range to generate an identification response matrix, a rotor flicker feature sequence is captured along the identification response matrix, and the rotor flicker feature sequence is input into a deep convolution level to perform feature coding to establish a UAV discrimination criterion;

[0008] Trajectory monitoring analysis is performed on the UAV discrimination criterion to identify a posture conversion marker, the posture conversion marker is segmented and labeled according to a maneuver amplitude attribute to generate a rotary motion domain and a linear motion domain, a motion connection degree is collected by monitoring a transition process of the linear motion domain to the rotary motion domain, and the motion connection degree is used to determine a flight steady state scale;

[0009] Based on the flight steady state scale, an optimal capture instant is selected in the rotary motion domain, a target feature string is extracted according to the optimal capture instant, a stable identification period is calibrated by the target feature string and the environmental interference factor, and a dynamic determination threshold is configured in the stable identification period;

[0010] Non-target interference sources are detected by comparing the dynamic determination threshold with the rotor flicker feature sequence, background noise components are extracted from the non-target interference sources, and false alarm is removed from the UAV discrimination criterion by using the background noise components to output a UAV identification confirmation result.

[0011] The second aspect of the present application proposes an unmanned aerial vehicle identification system based on artificial intelligence, comprising:

[0012] An image acquisition module is configured to acquire an image sequence of a monitoring area, and a target-background bidirectional separation process is applied to the image sequence to generate an aircraft contour boundary and an environmental interference factor, and a visual detection benchmark corresponding rule is created according to the aircraft contour boundary and the environmental interference factor;

[0013] A target search module is configured to search a candidate target range through the visual detection benchmark corresponding rule, apply multi-scale scanning analysis to the candidate target range to generate an identification response matrix, capture a rotor flicker feature sequence along the identification response matrix, and input the rotor flicker feature sequence into a deep convolution level to perform feature coding to establish a UAV discrimination criterion;

[0014] A trajectory analysis module is configured to perform trajectory monitoring analysis on the UAV discrimination criterion to identify a posture conversion marker, segment and label the posture conversion marker according to a maneuver amplitude attribute to generate a rotary motion domain and a linear motion domain, collect a motion connection degree by monitoring a transition process of the linear motion domain to the rotary motion domain, and use the motion connection degree to determine a flight steady state scale;

[0015] The period calibration module is configured to select an optimal capture moment in the rotating motion domain based on the flight steady scale, extract a target feature string from a plurality of image features according to the optimal capture moment, calibrate a stable recognition period by comparing the target feature string with the environmental interference factor, and configure a dynamic determination threshold in the stable recognition period.

[0016] The result output module is configured to detect a non-target interference source by comparing the dynamic determination threshold with the rotor flicker feature sequence, extract a background noise component from the non-target interference source, and apply the background noise component to the false alarm screening of the UAV discrimination criterion to output a UAV recognition confirmation result.

[0017] The beneficial effects of the present application are embodied in the following aspects: first, the target-background bidirectional separation processing combined with the rotor flicker feature sequence capture technology simultaneously separates from two directions of forward recognition of the moving target and reverse exclusion of the background interference, labels the environmental interference factor while extracting the aircraft contour boundary. The time-domain pulsation tracking is applied to the recognition response matrix, the time-frequency coupling correlation detection is used to verify the recognition oscillation period, and the rotor flicker feature sequence is formed. The bidirectional separation mechanism avoids the confusion of target features and background features, and the rotor flicker feature sequence captures the periodic light reflection phenomenon caused by the rotation of the rotor of the rotor UAV, providing a unique feature dimension for the UAV type discrimination. Second, the motion domain division and stable recognition period calibration technology are used to segmentally label the motion amplitude attribute of the attitude conversion marker, divide the rotating motion domain and the linear motion domain, monitor the transition process of the motion domain to establish the flight steady scale. Based on the flight steady scale, the optimal capture moment is selected in the rotating motion domain, the target feature string is compared with the environmental interference factor, the anti-interference weight is generated by applying the inverse mapping to the environmental interference factor, the signal-to-noise ratio evolution curve is generated by bidirectional modulation, and the stable recognition period is calibrated. The motion domain division realizes the structural decomposition of the flight trajectory, the signal-to-noise ratio evolution curve accurately reflects the actual conditions of target recognition through bidirectional modulation of the anti-interference weight, and realizes the adaptive optimization of the recognition opportunity and the determination threshold. Finally, a perfect false alarm screening mechanism is established, the deviation degree sequence is obtained by point-by-point comparison of the dynamic determination threshold and the rotor flicker feature sequence, the high deviation aggregation area is identified by abnormal aggregation detection of the deviation degree sequence, the target attribute verification is performed to exclude the candidate targets that do not meet the UAV features, and the remaining candidate targets are determined as non-target interference sources. The background noise component is extracted from the non-target interference source, and the noise gating suppression is applied to the UAV discrimination criterion. The false alarm screening mechanism combines the deviation degree analysis with the target attribute verification, realizes the identification and suppression of the interference signal by integrating the noise features into the discrimination criterion, and reduces the false detection caused by common interference sources such as birds, clouds and light changes. BRIEF DESCRIPTION OF DRAWINGS

[0018] The drawings here show the specific examples of the technical solutions of the present application, and constitute a part of the specification together with the specific embodiments, for explaining the technical solutions, principles and effects of the present application.

[0019] Unless specifically stated, the same reference signs in different drawings represent the same or similar technical features, and different reference signs may also be used to represent the same or similar technical features.

[0020] Figure 1 is a flow diagram of a method for identifying a UAV based on artificial intelligence.

[0021] Figure 2 is a structural block diagram of a system for identifying a UAV based on artificial intelligence. DETAILED DESCRIPTION

[0022] In the following description, for the purpose of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0023] It should be understood that the term "comprise" as used in the specification and the appended claims indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0024] In the present specification, the reference "one embodiment" or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. Thus, the appearances of the phrases "in one embodiment" or "in some embodiments" in various places in the specification are not necessarily all referring to the same embodiment, but can refer to one or more but not all embodiments, unless otherwise indicated. The terms "including," "comprising," "having," and variations thereof are meant to encompass the presence of the stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0025] The technical solutions of the embodiments of the present application are introduced as follows.

[0026] As shown in Figure 1 The embodiments of the present application provide a method for identifying a UAV based on artificial intelligence, which comprises the following steps S110-S150:

[0027] Step S110, acquire the image sequence of the monitoring area, generate the aircraft contour boundary and the environmental interference factor by applying the target-background bidirectional separation processing to the image sequence, and create the visual detection base reference corresponding rule according to the aircraft contour boundary and the environmental interference factor.

[0028] Specifically, the image sequence of the monitoring area is acquired. A plurality of fixed cameras and rotating pan-tilt cameras are deployed in the monitoring area, the field of view angle of the cameras covers the entire airspace of the monitoring area, forming a visual monitoring network with no blind area. The fixed cameras are installed at the boundary position of the monitoring area, the field of view angle is set to 90-120 degrees, and the airspace condition of the fixed area is monitored. The rotating pan-tilt camera is installed at the center position of the monitoring area, has the ability of 360-degree horizontal rotation and -30 to +90-degree pitch adjustment, and is responsible for dynamic tracking of the key area. The cameras continuously collect video pictures of the monitoring area at a frame rate of 25-30 frames per second, arrange the continuously collected video pictures in chronological order to form an image sequence. Each frame of image in the image sequence contains complete visual information of the monitoring area at a specific time, including the sky background, the cloud layer, the building contour, the trees, the aircraft target and various environmental interference objects. Distortion correction and deblurring preprocessing are performed on the image sequence to eliminate the image edge deformation caused by the wide-angle lens and improve the clarity of the image sequence, providing stable input data for target recognition and separation.

[0029] The aircraft contour boundary and the environmental interference factor are generated by applying the target-background bidirectional separation processing to the image sequence. Bidirectional separation processing is applied to each frame of image in the image sequence, which is performed simultaneously from two directions. The first direction is forward separation from target to background, which identifies the moving target with aircraft characteristics in the image sequence and separates the moving target from the static background. The motion area is identified by the pixel value difference between adjacent frames, and the area with a difference value greater than a preset threshold is marked as a motion area. The edge of the target in the motion area that meets the size and shape characteristics of the aircraft is detected, the contour line is extracted and connected to form a closed boundary, which is the aircraft contour boundary. The aircraft contour boundary describes the spatial position and geometric shape of the aircraft in the image sequence. The second direction is reverse separation from background to target, which identifies the background area and interference in the image sequence. A reference model is established for the background area, and the real-time image is compared with the background model to identify the abnormal area deviating from the background. The remaining part of the abnormal area after excluding the real aircraft target is the environmental interference factor, which includes cloud movement, bird flight, tree shaking, light mutation and other interference factors.

[0030] A visual detection benchmark correspondence rule is created according to the aircraft contour boundary and the environmental interference factor. Geometric characteristics of the aircraft contour boundary are analyzed, and a length-width ratio and a circularity parameter of the contour boundary are extracted. The length-width ratio is obtained by dividing the maximum length of the contour boundary by the maximum width. The contour boundary of a fixed-wing unmanned aerial vehicle is long and narrow, and the length-width ratio is between 3:1 and 5:1. The contour boundary of a rotary-wing unmanned aerial vehicle is compact, and the length-width ratio is close to 1:1. The circularity parameter is quantified by using a formula C=4πA / P², where C is the circularity value, A is the area enclosed by the contour boundary, and P is the perimeter of the contour boundary. The circularity value C of the fixed-wing unmanned aerial vehicle is in the range of 0.3-0.5, and the circularity value C of the rotary-wing unmanned aerial vehicle is in the range of 0.7-0.9. Motion characteristics of the aircraft contour boundary are analyzed, and a displacement speed and a displacement direction of the centroid of the contour boundary are obtained. The fixed-wing unmanned aerial vehicle has a high flight speed and a stable straight-line motion, and the rotary-wing unmanned aerial vehicle has a low flight speed and flexible maneuvering characteristics. Interference modes of the environmental interference factor are analyzed, and the occurrence frequency and the influence range of each type of interference are counted. The environmental interference factor is quantified as an interference intensity index. Based on the geometric and motion characteristics of the aircraft contour boundary and in combination with the interference intensity of the environmental interference factor, a visual detection benchmark correspondence rule is established. The length-width ratio, the circularity value, and the motion characteristics are used as judgment thresholds. When the target characteristics meet the parameter range of the fixed-wing unmanned aerial vehicle, the target is determined as a fixed-wing target. When the target characteristics meet the parameter range of the rotary-wing unmanned aerial vehicle, the target is determined as a rotary-wing target. In a low interference intensity area, loose feature thresholds are used, and in a high interference intensity area, strict feature thresholds are used. The visual detection benchmark correspondence rule realizes a mapping relationship from an image sequence to aircraft target recognition.

[0031] In step S120, a candidate target range is searched through the visual detection benchmark correspondence rule, a recognition response matrix is generated by applying multi-scale scanning analysis to the candidate target range, a rotary-wing flicker feature sequence is captured along the recognition response matrix, and the rotary-wing flicker feature sequence is input into a deep convolution layer to perform feature encoding and establish a unmanned aerial vehicle discrimination criterion.

[0032] In some embodiments, searching the candidate target range through the visual detection benchmark correspondence rule includes: deriving a response potential projection from the visual detection benchmark correspondence rule.

[0033] Based on the response potential projection, a local extreme value clustering area is identified. A time period in which the local extreme value clustering area exceeds a pre-set signal threshold is divided into a high-suspected block. A candidate target range is determined according to an extended margin of the high-suspected block.

[0034] The response potential projection is derived from the visual detection reference correspondence rule. The visual detection reference correspondence rule contains the mapping relationship between different feature parameters and aircraft categories, as well as the distribution information of each feature parameter in the spatial domain and the feature domain. The feature weight parameters in the rule are extracted, including the aspect ratio weight, the roundness weight, and the motion feature weight. These weights reflect the contribution degree of different features to the target determination. For each spatial position in the image sequence, the matching degree is calculated according to the feature value of the position and the aircraft feature range defined in the rule. The position with high matching degree is assigned a high potential value, and the position with low matching degree is assigned a low potential value. The potential value P_potential is calculated by the formula P_potential = Σ (w_i·f_i), where w_i is the weight of the i-th feature, and f_i is the normalized matching degree of the i-th feature at the position. The potential values of all spatial positions are arranged according to the two-dimensional spatial coordinates to form a potential distribution map. The horizontal and vertical axes of the potential distribution map correspond to the spatial coordinates of the image, and the value of each point in the map represents the possibility strength of the existence of the aircraft target at the position. The potential distribution map is normalized and smoothed to eliminate isolated noise points and retain continuous high potential regions. The normalized potential distribution map is the response potential projection. The response potential projection maps the multi-dimensional feature mapping relationship in the visual detection reference correspondence rule to a two-dimensional space, and directly displays the spatial distribution of the potential target in the monitoring area.

[0035] Local extreme value clustering areas are identified based on the response potential projection. Local extreme value detection is performed on the response potential projection to find the local maximum value points in the projection. The judgment condition for the local maximum value point is that the potential value of the point is greater than the potential values of all points in its 8-neighborhood, and the potential value exceeds 1.5 times of the global average potential. The detected local maximum value points are marked as extreme value points, which correspond to the positions with the most prominent features in the image. Cluster analysis is performed on the extreme value points, and a spatial clustering method based on density is used to cluster the extreme value points that are adjacent in space and have similar potential values into the same cluster. The clustering distance threshold is set to 50 pixels, and the potential difference threshold is set to 0.3 times of the potential standard deviation. The extreme value points with a distance less than the threshold and a potential difference less than the threshold are clustered into one class. The number of extreme value points included in each cluster, the spatial range of the cluster area, and the average potential intensity are calculated. The cluster that contains more than 3 extreme value points and has an average potential intensity exceeding the threshold is defined as a local extreme value clustering area. The local extreme value clustering area represents a region with continuously high potential values in the response potential projection, which corresponds to the candidate target position in the monitoring image with high feature intensity and consistent with the aircraft feature mode. The center coordinates, boundary range, and average potential value of all identified local extreme value clustering areas are recorded.

[0036] The local extremum clustering region whose time period exceeds the preset confidence threshold is divided into a high-suspicious block. Each local extremum clustering region is tracked in consecutive time frames, and the potential energy intensity change of the clustering region at different time is extracted. The potential energy intensity of the clustering region forms a time sequence curve on the time axis, and the horizontal axis of the curve is the time frame number, and the vertical axis is the potential energy intensity value. Each clustering region is assigned a confidence score, which comprehensively considers the average potential energy intensity of the clustering region, the compactness of the pixels in the region, and the motion continuity of the region. The average potential energy intensity reflects the quality of feature matching, the compactness reflects the aggregation degree of the region, and the motion continuity reflects the trajectory stability of the target between consecutive frames. A preset confidence threshold is set, which is dynamically adjusted according to the complexity of the monitoring environment. A lower threshold is used in a low-interference environment, and a higher threshold is used in a high-interference environment. Analyze the confidence score time sequence change of the clustering region, and identify the continuous time period whose score continuously exceeds the preset confidence threshold. The local extremum clustering region corresponding to the continuous time period whose confidence score exceeds the threshold is divided into a high-suspicious block. The high-suspicious block represents a target region that stably exists in consecutive multiple frames of images and has high feature matching degree, and these regions are more likely to contain real aircraft targets. The time range, spatial position and average confidence of all high-suspicious blocks are recorded.

[0037] The candidate target range is determined according to the high-suspicious block and the extended margin. The spatial boundary coordinates of each high-suspicious block are extracted, including the top-left corner coordinates and the bottom-right corner coordinates of the block. Since the high-suspicious block is obtained based on the local extremum clustering, its boundary may be close to the core region of the target and may not completely cover the complete range of the target. An extended margin is set around the boundary of the high-suspicious block, and the width of the extended margin is dynamically determined according to the size of the block. For small-size blocks (edge length less than 64 pixels), the extended margin is set to 20 pixels; for medium-size blocks (edge length 64-128 pixels), the extended margin is set to 15 pixels; for large-size blocks (edge length greater than 128 pixels), the extended margin is set to 10 pixels. The boundary of the high-suspicious block is extended by the corresponding margin in all directions, and the extended rectangular region is the candidate target range corresponding to the block. The setting of the extended margin ensures that the candidate target range can completely cover the possible motion trajectory and attitude change of the target, avoiding missing the edge features of the target. The extended margin is set for all identified high-suspicious blocks, and a list of corresponding candidate target ranges is generated.

[0038] A multi-scale scanning analysis is applied to the candidate target range to generate an identification response matrix. The multi-scale scanning is applied to the image region in the candidate target range, and the scanning scales include three levels of small scale (32x32 pixels), medium scale (64x64 pixels), and large scale (128x128 pixels). The small scale scanning is used to capture the features of micro drones or long-distance aircraft, the medium scale scanning is used to capture the features of regular drones, and the large scale scanning is used to capture the features of close-range or large aircraft. At each scale level, a sliding window is used to scan the candidate target range pixel by pixel, and the window sliding step is set to 50% of the window size to ensure coverage integrity. For each scanning window, an aircraft feature vector is extracted, and the feature vector is matched with a pre-trained aircraft feature template. The window position with a higher matching degree is assigned a higher response value. The scanning response values of the three scale levels are fused to generate a comprehensive response intensity at each spatial position of the candidate target range. The response intensities of all spatial positions are arranged according to the spatial coordinates to form a two-dimensional identification response matrix. The rows and columns of the matrix correspond to the spatial positions, and the matrix element values represent the response intensity of the aircraft target at that position.

[0039] In some embodiments, the rotor flicker feature sequence is captured along the identification response matrix, including: applying time-domain pulsation tracking to the identification response matrix to extract a response intensity fluctuation curve; implementing a harmonic law verification to identify an oscillation period; extending the oscillation period forward and backward to collect a sustained oscillation event; and forming a rotor flicker feature sequence through amplitude and phase coding of the sustained oscillation event.

[0040] A time-domain pulsation tracking is applied to the identification response matrix to extract a response intensity fluctuation curve. A number of positions with high response intensity in the identification response matrix are selected as tracking points, which correspond to spatial positions where rotor drones may exist. For each tracking point, the response intensity values at the position in the identification response matrix of consecutive time frames are extracted, and the response intensity values in the consecutive time frames are arranged in time sequence to form a time-domain response sequence of the tracking point. The rotor drone has periodic light reflection changes due to rotor rotation, resulting in periodic fluctuation characteristics of the response intensity of the tracking point. Fixed-wing drones or non-aircraft interference objects do not have periodic light reflection characteristics, and their response intensities are relatively stable or randomly fluctuate. The time-domain response sequence is subjected to smoothing filter processing to eliminate high-frequency noise interference and retain the periodic fluctuation components in the sequence. The smoothed time-domain response sequence is the response intensity fluctuation curve, with the horizontal axis representing the time frame number and the vertical axis representing the response intensity value. The response intensity fluctuation curve clearly shows the change rule of the response intensity of the tracking point over time, providing a time-domain analysis basis for identifying the rotor flicker feature.

[0041] For example, the harmonic law verification on the response intensity fluctuation curve identifies the oscillation period, including: time-frequency coupling correlation detection on the response intensity fluctuation curve to obtain a correlation coefficient sequence; peak value retrieval based on the correlation coefficient sequence to extract a peak interval time; stability check on the peak interval time to generate a period locking strength; dynamic locking screening on the peak interval time by the period locking strength to confirm the oscillation period.

[0042] The time-frequency coupling correlation detection on the response intensity fluctuation curve obtains a correlation coefficient sequence. The response intensity fluctuation curve is analyzed in both time domain and frequency domain. In the time domain, the curve is analyzed by moving window to detect the distribution law of wave peaks and wave troughs, and the time interval between adjacent wave peaks is counted to extract a period candidate value. In the frequency domain, the curve is Fourier transformed to convert the time domain curve into a frequency spectrum representation, and the energy concentrated frequency component in the frequency spectrum is identified, and the reciprocal of the frequency is taken as the period candidate value. The candidate period values extracted in the time domain and the frequency domain are coupled and verified, and the autocorrelation function R(τ)=∫I(t)·I(t+τ)dt is used to perform matching degree test on each candidate period value, where I(t) is the intensity function of the response intensity fluctuation curve, and τ is the candidate period value. A series of candidate period values T1, T2,..., Tn are obtained, and the correlation coefficient of each candidate period value with the response intensity fluctuation curve is calculated, and the candidate value with a high correlation coefficient indicates that the time domain and frequency domain features are mutually verified. The correlation coefficients of each candidate period value are arranged in descending order of the period value to form a correlation coefficient sequence, and the horizontal axis of the sequence is the candidate period value, and the vertical axis is the correlation coefficient corresponding to the period value. The correlation coefficient sequence reflects the matching degree of different period values with the actual period characteristics of the response intensity fluctuation curve. m The time-frequency coupling correlation detection on the response intensity fluctuation curve obtains a correlation coefficient sequence. The response intensity fluctuation curve is analyzed in both time domain and frequency domain. In the time domain, the curve is analyzed by moving window to detect the distribution law of wave peaks and wave troughs, and the time interval between adjacent wave peaks is counted to extract a period candidate value. In the frequency domain, the curve is Fourier transformed to convert the time domain curve into a frequency spectrum representation, and the energy concentrated frequency component in the frequency spectrum is identified, and the reciprocal of the frequency is taken as the period candidate value. The candidate period values extracted in the time domain and the frequency domain are coupled and verified, and the autocorrelation function R(τ)=∫I(t)·I(t+τ)dt is used to perform matching degree test on each candidate period value, where I(t) is the intensity function of the response intensity fluctuation curve, and τ is the candidate period value. A series of candidate period values T1, T2,..., Tn are obtained, and the correlation coefficient of each candidate period value with the response intensity fluctuation curve is calculated, and the candidate value with a high correlation coefficient indicates that the time domain and frequency domain features are mutually verified. The correlation coefficients of each candidate period value are arranged in descending order of the period value to form a correlation coefficient sequence, and the horizontal axis of the sequence is the candidate period value, and the vertical axis is the correlation coefficient corresponding to the period value. The correlation coefficient sequence reflects the matching degree of different period values with the actual period characteristics of the response intensity fluctuation curve.

[0043] The peak value retrieval based on the correlation coefficient sequence extracts a peak interval time. In the correlation coefficient sequence, local maximum points of the correlation coefficient are found, and these local maximum points correspond to period candidate values with high matching degrees with the response intensity fluctuation curve. The local maximum points need to meet two conditions: first, the correlation coefficient of the point is greater than the correlation coefficients of its adjacent points; second, the correlation coefficient of the point exceeds a preset correlation threshold. The local maximum points that meet the conditions are marked as peak points, and the period candidate value corresponding to the peak points is the peak interval time. Usually, multiple peak points appear in the correlation coefficient sequence, and these peak points may correspond to the fundamental frequency period and its harmonic periods. The multiple peak interval times are screened, and the interval time corresponding to the peak with the highest correlation coefficient is preferentially selected. If the interval time corresponding to the highest peak is not within the reasonable range (12-60 milliseconds) of the rotor flicker, the interval time corresponding to the second highest peak is selected. The peak interval time reflects the most significant periodic characteristics in the response intensity fluctuation curve, and is an important intermediate result of identifying the oscillation period.

[0044] The peak interval time is applied to stability check to generate cycle lock strength. The response strength fluctuation curve is analyzed in segments, and the curve is equally divided into several time segments, each of which contains at least 3 complete candidate cycles. The cycle characteristics of each segment are extracted independently, and the peak interval time of the time segment is obtained. The peak interval time extracted from each time segment is compared with the candidate peak interval time, and the deviation Δ = |T_segment-T_candidate| / T_candidate, where T_segment is the cycle extracted from the time segment, and T_candidate is the candidate peak interval time. If the deviation Δ of all time segments is less than 10%, it is considered that the peak interval time remains highly stable in the entire curve, and a higher cycle lock strength is given. The cycle lock strength L has a value range of 0-1, and the higher the stability, the closer the L value to 1. If the deviation of part of the time segments is large, the value of the cycle lock strength is reduced. The cycle lock strength quantifies the reliability of the peak interval time as the true oscillation period, providing confidence evaluation for the final cycle confirmation.

[0045] The peak interval time is applied to stability check to generate cycle lock strength. The response strength fluctuation curve is analyzed in segments, and the curve is equally divided into several time segments, each of which contains at least 3 complete candidate cycles. The cycle characteristics of each segment are extracted independently, and the peak interval time of the time segment is obtained. The peak interval time extracted from each time segment is compared with the candidate peak interval time, and the deviation Δ = |T_segment-T_candidate| / T_candidate, where T_segment is the cycle extracted from the time segment, and T_candidate is the candidate peak interval time. If the deviation Δ of all time segments is less than 10%, it is considered that the peak interval time remains highly stable in the entire curve, and a higher cycle lock strength is given. The cycle lock strength L has a value range of 0-1, and the higher the stability, the closer the L value to 1. If the deviation of part of the time segments is large, the value of the cycle lock strength is reduced. The cycle lock strength quantifies the reliability of the peak interval time as the true oscillation period, providing confidence evaluation for the final cycle confirmation.

[0046] The system collects continuous oscillation events by extending the search forward and backward through frame segments based on the oscillation period. Using the identified oscillation period as the center, the search extends forward and backward through time frames to find continuously occurring periodic fluctuations in the response intensity fluctuation curve. The search extends forward by at least two oscillation periods and backward by at least two oscillation periods, ensuring at least five complete oscillation periods are collected. Within the extended time range, the system verifies whether the response intensity fluctuation curve maintains the periodic characteristics consistent with the identified oscillation period. Periodic consistency is assessed by the deviation between the time interval between adjacent peaks and the oscillation period; periods with a deviation less than 15% are considered to maintain periodicity. A continuous time period that meets the periodic consistency requirement is defined as a continuous oscillation event. A continuous oscillation event contains multiple complete oscillation periods along with their amplitude and phase information. The amplitude information of the continuous oscillation event reflects the range of variation in the rotor reflected light intensity, and the phase information reflects the relative position of the rotor rotation angle at different times. The longer the duration of the continuous oscillation event, the more stable the rotor flicker characteristics, and the higher the confidence that the target is a rotorcraft drone.

[0047] A rotor scintillation feature sequence is formed by amplitude and phase encoding of a continuous oscillation event. Amplitude and phase features are extracted for each complete oscillation cycle in the continuous oscillation event. Amplitude features are obtained by detecting the peak and trough values ​​of the response intensity fluctuation curve within that cycle. The oscillation amplitude is obtained by subtracting the trough value from the peak value, and the amplitude is normalized to the 0-1 range. Phase features are obtained by marking the time position of the peak within that cycle. The start time of the cycle is defined as phase 0 degrees, and the phase corresponding to the peak time is converted to an angle value of 0-360 degrees using a time scaling factor. The amplitude and phase values ​​of multiple consecutive cycles in the continuous oscillation event are arranged in chronological order to form amplitude and phase sequences. The amplitude sequence describes the variation law of rotor scintillation intensity, and the phase sequence describes the variation law of rotor rotation angle. The amplitude and phase sequences are interleaved and encoded using the encoding method [A1, φ1, A2, φ2, ..., A n , φ n ], where A i Let φ be the normalized amplitude value for the i-th period. i Let be the phase angle of the i-th cycle, and n be the number of complete acquisition cycles. The interleaved and encoded sequence is the rotor scintillation feature sequence, which fully describes the time and frequency domain characteristics of the rotor scintillation phenomenon.

[0048] The rotor flash feature sequence is input into a deep convolutional layer for feature coding to establish the UAV discrimination criterion. The rotor flash feature sequence is input into a deep convolutional neural network, which includes multiple convolutional layers and pooling layers. The first convolutional layer extracts local features from the feature sequence, and the convolution kernel size is 3x1. Each convolution kernel slides over the sequence to extract the correlation pattern between adjacent elements. The first convolution generates multiple feature maps, each of which emphasizes different local features in the sequence. The second convolutional layer extracts higher-level abstract features based on the first layer feature maps, and the convolution kernel size is 5x1, which captures the periodic pattern and amplitude-phase change pattern in a larger range of the sequence. The pooling layer reduces the dimension of the convolution features, and the maximum pooling retains the most significant features in each local region. After multiple convolution and pooling processing, the rotor flash feature sequence is encoded into a compact high-dimensional feature vector, which contains the core discrimination information of the rotor flash pattern. The encoded feature vector is input into a fully connected layer for classification and judgment. The output nodes of the fully connected layer correspond to different target categories, including rotor UAV, fixed-wing UAV and non-aircraft interference. The output of the fully connected layer is converted into a probability distribution through a softmax activation function, and the class with the highest probability value is the judgment result of the target. Based on the feature coding and classification judgment of the deep convolutional layer, the UAV discrimination criterion is established, which specifies the mapping relationship from the rotor flash feature sequence to the target category.

[0049] In step S130, the trajectory monitoring analysis is performed on the UAV discrimination criterion to identify the attitude transition marker. The attitude transition marker is segmented and labeled according to the maneuver amplitude attribute to generate a rotational motion domain and a linear motion domain. The transition process from the linear motion domain to the rotational motion domain is monitored to collect the motion continuity, and the flight steady state scale is determined based on the motion continuity.

[0050] Specifically, the trajectory monitoring analysis discriminates the attitude transition markers for the UAV discriminant criterion. Based on the UAV target identified by the UAV discriminant criterion, the spatial position of the target in consecutive time frames is tracked and recorded to form a motion trajectory sequence of the target. The motion trajectory sequence contains the center coordinates, bounding box size and orientation angle information of the target in each frame of image. The motion trajectory sequence is subjected to time series analysis to extract the motion speed, motion direction and motion acceleration parameters of the trajectory. The motion speed is obtained by dividing the displacement distance of the target center coordinates between adjacent frames by the time interval, the motion direction is determined by the angle of the displacement vector, and the motion acceleration is obtained by dividing the change in speed by the time interval. The time series variation characteristics of the motion parameters are analyzed to identify the time points at which the motion state in the trajectory changes significantly. When the motion speed, motion direction or motion acceleration of the target changes abruptly, the time point is marked as an attitude transition marker. The attitude transition marker represents the moment when the UAV switches from one flight state to another flight state, such as from straight-line flight to rotating maneuver, or from rotating maneuver to straight-line flight. The time stamp, position coordinates and motion parameter difference values before and after each attitude transition marker are recorded to form the attribute description of the attitude transition marker.

[0051] In some embodiments, the segmenting and labeling of the motion amplitude attribute of the attitude transition marker generates a rotating motion domain and a straight-line motion domain, including: applying the pre-frame and post-frame velocity vector decoupling to the attitude transition marker to obtain a dynamic jump variable; implementing intensity-increasing grading for the dynamic jump variable to generate a maneuver intensity level; determining a low-intensity period and a high-intensity period based on the maneuver intensity level; labeling the low-intensity period as a straight-line motion domain and the high-intensity period as a rotating motion domain.

[0052] The dynamic jump variable is obtained by decoupling the front and back frame speed vector applied to the pose transition marker. At the time corresponding to each pose transition marker, the speed vector of the UAV in the previous frame and the next frame is extracted. The speed vectors of the front and back frames are decoupled and analyzed, and the speed change in the horizontal and vertical directions is calculated respectively, and then vector synthesis is performed to obtain the total amplitude of the speed change. The total amplitude of the speed change is the dynamic jump variable at the pose transition marker. The larger the jump variable value, the more intense the speed adjustment of the UAV. The smaller the jump variable value, the more gentle the speed adjustment. The dynamic jump variable corresponding to all the pose transition markers in the motion trajectory is extracted to form a sequence of dynamic jump variables. The dynamic jump variable sequence reflects the spatiotemporal distribution characteristics of the speed adjustment of the UAV in the entire flight process. In actual monitoring, the dynamic jump variable of the fixed-wing UAV is usually small and changes gently, because its flight mode is mainly straight line at constant speed. The dynamic jump variable of the rotor UAV fluctuates greatly and frequently appears peak value, because it frequently performs hovering, turning and maneuvering adjustment. By analyzing the numerical distribution and variation law of the dynamic jump variable, the flight type and control characteristics of the UAV can be preliminarily judged. The collection frequency of the dynamic jump variable is consistent with the frame rate of the image sequence, which is usually 25-30 times per second, ensuring that the details of the speed change in the rapid maneuvering process of the UAV can be captured.

[0053] A four-level grading scheme is adopted to divide the dynamic jump variable into four levels: micro jump, light jump, moderate jump and severe jump. Micro jump corresponds to the case of extremely small speed adjustment amplitude, which usually occurs during the constant speed cruising stage of the UAV. Light jump corresponds to small amplitude speed fine tuning, which usually occurs when the UAV maintains the heading or height. Moderate jump corresponds to obvious speed adjustment, which usually occurs when the UAV turns or changes the flight height. Severe jump corresponds to rapid speed change, which usually occurs when the UAV avoids obstacles, turns quickly or suddenly accelerates or decelerates. Each dynamic jump variable is assigned to the corresponding level according to its numerical value, and the level is the maneuvering intensity level at the pose transition marker. The maneuvering intensity level quantifies the maneuvering intensity of the UAV during the pose transition, providing classification basis for the division of the motion domain. The grading threshold is dynamically adjusted according to the type of UAV and the monitoring environment. The grading threshold of the micro UAV is usually lower than that of the large UAV, because their maneuvering ability and speed range are different.

[0054] The low-intensity period and the high-intensity period are determined based on the maneuver intensity level. The sequence of motion trajectories is traversed in time sequence to all attitude transition markers and their corresponding maneuver intensity levels. Markers with adjacent same or similar maneuver intensity levels are clustered to form continuous maneuver intensity intervals. The markers of slight jump and slight change are clustered to form the low-intensity period, and the markers of moderate jump and severe jump are clustered to form the high-intensity period. The starting time, ending time and average maneuver intensity of each period are marked. The speed of the unmanned aerial vehicle in the low-intensity period changes gently, and the motion state is relatively stable, corresponding to low-speed adjustment actions such as uniform cruise and stable straight flight. The speed of the unmanned aerial vehicle in the high-intensity period changes dramatically, and the motion state is quickly adjusted, corresponding to high-intensity maneuver actions such as emergency turning, rapid circling and emergency obstacle avoidance. All periods on the time axis are marked to determine whether each time point belongs to the low-intensity period or the high-intensity period. The division of the low-intensity period and the high-intensity period provides a time range basis for the final marking of the motion domain.

[0055] The low-intensity period is marked as a straight-line motion domain, and the high-intensity period is marked as a rotating motion domain. Each low-intensity period is assigned a label of a straight-line motion domain, which indicates that the unmanned aerial vehicle mainly performs uniform flight actions in this period. The characteristic of uniform flight is that the speed changes relatively gently, the motion state is stable, and the maneuver intensity is low. The typical flight scenarios corresponding to the straight-line motion domain include: the unmanned aerial vehicle patrols along the predetermined route, performs constant-speed reconnaissance tasks, performs straight-line transfer between two target points, or maintains stable flight with fixed heading. In the straight-line motion domain, the control instructions of the unmanned aerial vehicle are relatively simple, mainly maintaining the heading and speed, and the flight state is highly predictable. Each high-intensity period is assigned a label of a rotating motion domain, which indicates that the unmanned aerial vehicle mainly performs maneuver adjustment actions in this period. Maneuver adjustment includes rotation, turning, rapid change of heading and other actions, which are characterized by dramatic speed changes and high maneuver intensity. The typical flight scenarios corresponding to the rotating motion domain include: the unmanned aerial vehicle rapidly circles in the target area for observation, performs emergency avoidance maneuver, performs dynamic search scanning, or rapidly changes the heading in response to sudden instructions. In the rotating motion domain, the control instructions of the unmanned aerial vehicle change frequently, and the uncertainty of the flight state is high. The sequence of motion trajectories is divided into multiple continuous motion domain segments according to the time axis, and each segment is marked as a straight-line motion domain or a rotating motion domain. After the motion domain marking is completed, the motion mode switching of the unmanned aerial vehicle in the entire flight process can be clearly identified, providing a motion domain classification basis for flight steady-state analysis.

[0056] The motion connection degree is collected in the transition process from the linear motion domain to the rotary motion domain. The transition interval from the linear motion domain to the rotary motion domain in the motion trajectory sequence is identified, and the transition interval is located at the junction position of the two motion domains. The motion trajectory in the transition interval is analyzed in detail, and the velocity variation, direction variation and acceleration variation curves of the interval are extracted. The velocity variation curve reflects the speed adjustment of the UAV in the transition process, the direction variation curve reflects the angle change rate of the UAV turning, and the acceleration variation curve reflects the maneuvering force of the UAV. The smoothness of the three variation curves is evaluated, and the absolute value integral of the second derivative of the curve is used to quantify the smoothness S = ∫|d²v / dt²|dt, wherein v represents the speed, direction or acceleration parameter. The smaller the smoothness value is, the more gentle the transition process is, and the larger the value is, the more intense the transition process is. The velocity smoothness, direction smoothness and acceleration smoothness of the transition interval are weighted and fused to obtain the motion connection degree index of the transition process. The high motion connection degree indicates that the UAV smoothly transitions when switching between motion domains, and the low motion connection degree indicates that the UAV has intense adjustment when switching.

[0057] In some embodiments, the motion connection degree is used to establish a flight steady state scale, including: applying a time sequence steady state test to the motion connection degree to extract a stability fluctuation feature; identifying a jump protruding point based on the stability fluctuation feature; determining a steady state maintenance interval according to the extension of the jump protruding point to both sides; and establishing a flight steady state scale through the time span of the steady state maintenance interval.

[0058] The time sequence steady state test is applied to the motion connection degree to extract a stability fluctuation feature. The collected motion connection degree time sequence is taken as an analysis object, and the sequence records the smoothness of the UAV when transitioning from the linear motion domain to the rotary motion domain at different times. The motion connection degree sequence is subjected to a time sequence steady state test to evaluate the stability of the sequence values on the time axis. The steady state test is performed in a sliding window manner, and the window length is set to contain at least 3 motion connection degree sampling points. The mean μ and the standard deviation σ of the motion connection degree are calculated in each window, and the mean reflects the average level of the connection degree in the window, and the standard deviation reflects the fluctuation degree of the connection degree. The standard deviation σ is taken as a quantitative index of the stability fluctuation feature, and the smaller the standard deviation is, the more stable the motion connection degree in the window is, and the larger the standard deviation is, the more intense the motion connection degree fluctuates. The standard deviation values of all windows in the entire time sequence are extracted to form a stability fluctuation feature sequence. The horizontal axis of the stability fluctuation feature sequence is time, and the vertical axis is the fluctuation degree, and the sequence clearly shows the time sequence variation law of the motion smoothness of the UAV.

[0059] The jump protruding point is identified based on the stability fluctuation feature. The stability fluctuation feature sequence is analyzed to find the position where the fluctuation degree suddenly increases. The sudden increase in the fluctuation degree indicates that the smoothness of the motion of the unmanned aerial vehicle significantly decreases at this moment, and the transition process becomes unstable. A threshold detection method is used to identify the jump protruding point. A fluctuation threshold σ_threshold is set. When the stability fluctuation feature value at a certain moment exceeds the threshold, the moment is marked as a candidate jump protruding point. The candidate jump protruding point is further screened. The fluctuation value of the point is required to be obviously higher than the fluctuation values of adjacent points before and after it, that is, it satisfies the local maximum value condition. The point that satisfies the threshold condition and the local maximum value condition is confirmed as the jump protruding point. The jump protruding point represents the moment when the flight stability of the unmanned aerial vehicle is destroyed. At this moment, the motion control of the unmanned aerial vehicle has a large fluctuation or performs a sharp maneuver adjustment. The time position and the fluctuation intensity of all the identified jump protruding points are recorded to provide boundary reference for the division of the steady state interval.

[0060] The steady state maintenance interval is determined according to the extension of the jump protruding point to both sides. Each jump protruding point is taken as a boundary. The time period before and after the point belongs to different steady state statuses. The jump protruding point is traced back to find the time period where the stability fluctuation feature value is continuously lower than the fluctuation threshold. The time period indicates that the unmanned aerial vehicle is in a stable flight state before the jump protruding point. The jump protruding point is extended to find the time point where the stability fluctuation feature value is reduced to below the fluctuation threshold again. The time period from the time point to the next jump protruding point indicates that the unmanned aerial vehicle returns to a new stable flight state. The time interval between two adjacent jump protruding points and the stability fluctuation feature value continuously below the threshold is defined as the steady state maintenance interval. The motion continuity of the unmanned aerial vehicle in the steady state maintenance interval remains stable, the transition process is smooth, and the flight control stability is high. All the steady state maintenance intervals identified in the entire flight process are labeled, and the start time and the end time of each interval are recorded. The distribution of the steady state maintenance interval reflects the time and space distribution characteristics of the flight stability of the unmanned aerial vehicle.

[0061] The flight steady state scale is established by the time span of the steady state keeping interval. The time span of all identified steady state keeping intervals is counted, and the time span T_i is equal to the end time minus the start time of the i-th interval. The time spans of different steady state keeping intervals are different, and the longer time span indicates that the UAV can maintain stable flight state for a long time, and the shorter time span indicates that the steady state keeping ability of the UAV is weaker. The flight steady state scale is quantified by using a weighted average method, and the formula is T_steady = Σ(w_i·T_i) / Σw_i, where T_steady is the flight steady state scale, T_i is the time span of the i-th steady state keeping interval, w_i is the weight coefficient of the i-th interval, and the weight coefficient is inversely proportional to the average value of the stability fluctuation characteristics in the interval, and the smaller the fluctuation, the greater the weight. For a flight process with n steady state keeping intervals, the flight steady state scale comprehensively considers the time length and stability degree of all steady state intervals. The larger the flight steady state scale, the better the flight steady state performance of the UAV, and the higher the control accuracy. The flight steady state scale of a professional UAV is usually 30-60 seconds, and the flight steady state scale of a consumer UAV is usually 10-30 seconds. The UAV affected by interference or abnormal control may be less than 10 seconds.

[0062] In step S140, the best capture moment is selected in the rotational motion domain based on the flight steady state scale, the target feature string is generated by extracting multiple image features according to the best capture moment, and the stable recognition period is calibrated by the calibration of the target feature string and the environmental interference factor, and the dynamic judgment threshold is configured in the stable recognition period.

[0063] Specifically, the best capture moment is selected in the rotational motion domain based on the flight steady state scale. The flight steady state scale value is analyzed, and the larger the value, the more stable the UAV flight, and the higher the quality of image capture. In the rotational motion domain, the attitude and orientation of the UAV continuously change when it performs a rotational maneuver, which provides favorable conditions for multi-angle feature capture. When all the rotational motion domain periods are traversed, the flight stability in each period is evaluated. The flight steady state scale is used as the stability evaluation benchmark, and the motion continuity in the period is compared with the flight steady state scale. The time when the motion continuity is close to or exceeds the average level of the flight steady state scale is marked as a high stability time. At the high stability time, the motion of the UAV is smooth and predictable, and the image clarity and feature extraction success rate are high. The high stability times in each rotational motion domain are sorted, and the time with the highest motion continuity is preferentially selected as the candidate best capture moment. The image quality corresponding to the candidate moment is further verified, including image clarity, target visibility and background interference degree. The best capture moment is selected by comprehensive scoring S_capture = α·Q_clear + β·Q_visible - γ·D_interference, where Q_clear is the image clarity score, Q_visible is the target visibility score, D_interference is the background interference degree, and α, β and γ are weight coefficients.

[0064] A target feature string is generated by extracting multiple frame image features at the best capture moment. A time window is formed by extending several frames of images before and after the time point corresponding to the best capture moment. The length of the time window is dynamically adjusted according to the rotation speed of the UAV, and the window is shorter when the rotation speed is faster and longer when the rotation speed is slower, ensuring that the window contains multiple images of the UAV at different angles. Usually, the time window contains 5-9 frames of images, and the frame interval is set to 0.1-0.2 seconds. Feature extraction is performed on each frame of image in the time window, and the extracted features include the edge features, texture features and color features of the target. The edge features are obtained by edge detection algorithm to obtain the geometric information of the target contour, the texture features describe the detailed pattern of the target surface, and the color features statistics the color distribution of the target. The feature vectors extracted from each frame of image are standardized to eliminate the differences in the dimensions of different features. The feature vectors of all frames in the time window are connected in time sequence to form a long feature vector, which is the target feature string. The target feature string contains the feature information of the UAV at multiple angles near the best capture moment, and fully describes the visual characteristics of the target.

[0065] In some embodiments, the stable recognition period is calibrated by comparing the target feature string with the environmental interference factors, including: applying time axis alignment to the target feature string to extract feature saliency at each time; performing reverse contrast analysis on the feature saliency and the environmental interference factors to obtain a signal-to-noise ratio evolution curve; applying peak interval retrieval to the signal-to-noise ratio evolution curve to identify a high signal-to-noise duration segment; and calibrating the stable recognition period according to the start and end boundaries of the high signal-to-noise duration segment.

[0066] The feature saliency at each time is extracted by applying time axis alignment to the target feature string. The feature vectors in the target feature string are unfolded in time sequence, and each feature vector corresponds to the image features at a specific time. The time axis is standardized by setting the best capture moment as the time origin, and the time points of the previous and subsequent frames are labeled according to the relative time offset. Time axis alignment ensures that the target feature strings extracted under different flight scenarios have a unified time reference basis. The saliency of each time feature vector is evaluated, and the saliency reflects the prominence of the target feature at that time relative to the background feature. The feature saliency D_saliency is quantified by the ratio of the norm of the feature vector to the mean of the background feature, and the larger the ratio, the more significant the feature. The saliency values of all times in the target feature string are extracted to form a time sequence of feature saliency. The horizontal axis of the feature saliency sequence is the standardized time, and the vertical axis is the saliency value. This sequence clearly shows the change in the prominence of the target feature over time. The time with high feature saliency indicates that the target feature is clear and identifiable, and the time with low feature saliency indicates that the target feature is blurred or submerged in the background.

[0067] The signal-to-noise ratio evolution curve is obtained by performing reverse contrast analysis on the feature saliency and the environmental interference factor. The feature saliency is taken as a signal intensity reference to develop a signal reference curve along a time axis. An anti-interference weight is generated by applying an inverse mapping to the environmental interference factor. A modulated response pair is generated by performing bidirectional modulation on the signal reference curve and the environmental interference factor through the anti-interference weight. A signal-to-noise ratio evolution curve is obtained by taking a ratio of a signal component to a noise component of the modulated response pair.

[0068] The feature saliency is taken as a signal intensity reference to develop a signal reference curve along a time axis. The feature saliency time series is taken as a signal source, and each value in the series represents the saliency of the target feature at a specific time. The feature saliency values are standardized to a uniform intensity range, typically normalized to the 0-1 interval. The normalized values are taken as signal intensity reference values. The signal intensity reference values are developed along the time axis in chronological order, with the horizontal axis representing standardized time and the vertical axis representing signal intensity reference values. The continuous curve formed after development is the signal reference curve, and the shape of the curve reflects the variation of the target feature intensity over time. The areas with higher values in the signal reference curve correspond to periods when the target feature is salient, and the areas with lower values correspond to periods when the target feature is weak. When the UAV is flying steadily in clear skies, the signal reference curve maintains a high and stable value. When the UAV passes through clouds or enters a shadow area, the curve value decreases significantly. The signal reference curve is subjected to smoothing filter processing to eliminate short-term fluctuation noise and retain the overall trend of the curve. The smoothed signal reference curve is taken as the signal reference benchmark for reverse contrast analysis.

[0069] An anti-interference weight is generated by applying an inverse mapping to the environmental interference factor. The interference intensity values corresponding to the time periods of the signal reference curve are extracted from the time series of the environmental interference factor. The greater the interference intensity value, the stronger the environmental interference and the greater the impact on target recognition. An inverse mapping transformation is applied to the interference intensity values, with the mapping formula being w_anti-interference = 1 / (N_interference + ε), where w_anti-interference is the anti-interference weight, N_interference is the interference intensity value, and ε is a small positive number to avoid division by zero. The characteristic of the inverse mapping is that the greater the interference intensity, the smaller the anti-interference weight, and the smaller the interference intensity, the greater the anti-interference weight. This inverse relationship ensures that the contribution weight of the signal is reduced in high-interference environments and increased in low-interference environments. Under clear sky conditions, the environmental interference factor value is low, and the corresponding anti-interference weight is close to the maximum value. In cloudy weather, the interference is enhanced, causing the weight to decrease moderately. In rainy and foggy weather, strong interference causes the weight to decrease significantly. All the anti-interference weights at different times are arranged in chronological order to form an anti-interference weight time series. The anti-interference weight sequence is taken as a modulation parameter for adjusting the relative contributions of the signal reference curve and the environmental interference factor.

[0070] The signal reference curve and the environmental interference factor are bidirectionally modulated by the anti-interference weight to generate a modulation response pair. The bidirectional modulation includes two directions of signal enhancement modulation and interference suppression modulation. In the signal enhancement modulation, the value of the signal reference curve at each time is multiplied by the corresponding anti-interference weight to obtain a weighted signal intensity S_weighted(t) = S_reference(t) w_anti-interference(t), where S_reference(t) is the value of the signal reference curve at the time. The weighted signal intensity is amplified in the low interference period and suppressed in the high interference period. In the interference suppression modulation, the interference intensity value of the environmental interference factor is multiplied by (1-anti-interference weight) to obtain a modulation noise intensity N_modulation(t) = N_interference(t) (1-w_anti-interference(t)). The modulated noise intensity is further reduced in the low interference period and remains at the original level in the high interference period. The weighted signal intensity S_weighted(t) and the modulation noise intensity N_modulation(t) at each time are combined into a two-tuple [S_weighted(t), N_modulation(t)], which is the modulation response pair at the time. The modulation response pairs at all times are arranged in chronological order to form a modulation response pair sequence, which records the signal and noise components after bidirectional modulation.

[0071] The ratio of the signal component to the noise component of the modulation response pair is taken as the signal-to-noise ratio evolution curve. For each time in the modulation response pair sequence, the signal component S_weighted(t) and the noise component N_modulation(t) are extracted. At each time, the value of the signal-to-noise ratio is the ratio of the two components SNR_modulation(t) = S_weighted(t) / N_modulation(t). This ratio reflects the degree of advantage of the signal over the noise after bidirectional modulation. The signal-to-noise ratio values at all times are arranged in chronological order and plotted into a curve, which is the signal-to-noise ratio evolution curve after modulation. Compared with the simple signal-to-noise ratio obtained by directly dividing the feature saliency by the interference intensity, the signal-to-noise ratio evolution curve after bidirectional modulation by the anti-interference weight more accurately reflects the actual conditions of target recognition. When the unmanned aerial vehicle performs a cruise task in a clear sky background, the signal-to-noise ratio evolution curve maintains a high and stable value. When the monitoring area is blocked by clouds or the light changes suddenly, the curve fluctuates and decreases, but still maintains within the identifiable range. The bidirectional modulation mechanism amplifies the signal-to-noise ratio difference in periods of weak interference and suppresses false high signal-to-noise ratios in periods of strong interference, improving the robustness and discriminant accuracy of the signal-to-noise ratio evolution curve.

[0072] A peak interval retrieval is imposed on the SNR evolution curve to identify a high SNR duration section. The morphological features of the SNR evolution curve are analyzed to find the interval in which the SNR remains at a high level. A SNR threshold SNR_threshold is set, which is determined according to the accuracy requirement of target recognition. The time points at which the SNR is higher than the threshold are marked as high SNR points by traversing the SNR evolution curve. The continuity of the marked high SNR points is checked. If multiple consecutive time points are high SNR points, the time interval formed by these consecutive points is defined as a candidate high SNR interval. A peak retrieval is imposed on the candidate interval to find the maximum value point of the SNR in each interval, which corresponds to the time at which the recognition condition is optimal. The accurate boundaries of the interval are determined by extending forward and backward from the maximum value point to the positions at which the SNR drops to the threshold. The continuous high SNR interval obtained after the peak interval retrieval is the high SNR duration section. The high SNR duration section represents the time period in which the SNR is continuously higher than the threshold, and in this period, the target features are stable and prominent, and the environmental interference is weak. The start time and end time of all the identified high SNR duration sections are recorded to provide candidate intervals for the calibration of stable recognition period.

[0073] A stable recognition period is calibrated according to the start and end boundaries of the high SNR duration section. The quality of all the identified high SNR duration sections is evaluated, and the evaluation indicators include the period length, average SNR, and SNR stability. The period length reflects the duration of the section, and the longer the period, the larger the time window for stable recognition. The average SNR reflects the overall recognition condition level in the section, and the higher the average value, the more superior the recognition condition. The SNR stability is measured by the standard deviation of the SNR in the section, and the smaller the standard deviation, the more stable the SNR. Each high SNR duration section is scored comprehensively, and the score Q_section = w1 · T_length + w2 · SNR_mean - w3 · σ_SNR, where T_length is the period length, SNR_mean is the average SNR, σ_SNR is the SNR standard deviation, and w1, w2, w3 are weight coefficients. The high SNR duration section with the highest score is selected, and its start and end boundaries are used as the boundaries of the stable recognition period. The start boundary of the stable recognition period corresponds to the start time of the section, and the end boundary corresponds to the end time of the section. A buffer zone is set at the start and end boundaries, and the buffer zone width is 5-10% of the period length to avoid the influence of the sudden change of the SNR at the boundaries on the recognition stability. The calibrated stable recognition period is used as the time window for subsequent target recognition and determination, and the recognition operation performed in this period has higher accuracy and reliability.

[0074] The dynamic determination threshold is configured in the stable recognition period. Statistical analysis is performed on the target feature in the stable recognition period, and the mean, standard deviation and variation range of the feature are extracted. The feature mean reflects the typical feature level of the target, and the standard deviation reflects the fluctuation degree of the feature. The dynamic determination threshold is set based on the statistical characteristics of the feature, and the threshold value T_threshold = μ_feature-k·σ_feature, wherein μ_feature is the feature mean, σ_feature is the feature standard deviation, and k is the adjustment coefficient. The adjustment coefficient k is dynamically adjusted according to the recognition accuracy requirement. When high-precision recognition is required, increase the value of k to increase the threshold. When high-recall rate recognition is required, reduce the value of k to reduce the threshold. The dynamic determination threshold is used as the reference threshold for target determination. If the feature value of the input image is higher than the threshold, it is determined as a target, otherwise it is determined as a non-target. Since the target feature is relatively stable in the stable recognition period, the dynamic determination threshold based on the period has high adaptability and accuracy. Different dynamic determination thresholds are configured for different sub-periods in the stable recognition period to adapt to the slight changes of the target feature. The dynamic determination threshold is self-adaptive to the time and environmental conditions, and the robustness of target recognition is improved.

[0075] In step S150, the non-target interference source is detected by comparing the dynamic determination threshold with the rotor flicker feature sequence, the background noise component is extracted from the non-target interference source, and the background noise component is used to apply false alarm screening to the unmanned aerial vehicle discrimination criterion to output the unmanned aerial vehicle recognition confirmation result.

[0076] In some embodiments, the detection of the non-target interference source by comparing the dynamic determination threshold with the rotor flicker feature sequence includes: using the dynamic determination threshold as a screening reference to perform point-by-point comparison on the rotor flicker feature sequence to obtain a deviation sequence; performing abnormal aggregation detection on the deviation sequence to identify a high deviation aggregation area; performing target attribute verification on the high deviation aggregation area to exclude candidate targets that do not meet the characteristics of the unmanned aerial vehicle; and determining the identified candidate target as the non-target interference source.

[0077] The dynamic determination threshold is taken as a screening criterion to impose point-by-point comparison on the rotor flicker feature sequence to obtain a deviation degree sequence. The numerical values of feature points are extracted from the rotor flicker feature sequence one by one, and the feature points include oscillation amplitude values and phase angle values. For each feature point, its numerical value is compared with the dynamic determination threshold at the corresponding time. The comparison mode is to calculate the difference value between the numerical value of the feature point and the threshold value, the difference value Δ = F_feature - T_threshold, where F_feature is the numerical value of the feature point, and T_threshold is the numerical value of the dynamic determination threshold. A positive difference value indicates that the numerical value of the feature point is higher than the threshold, and a negative difference value indicates that the numerical value of the feature point is lower than the threshold. The difference value is normalized to obtain the deviation degree d_deviation = Δ / T_threshold. The normalized deviation degree reflects the deviation degree of the feature point relative to the threshold. The larger the absolute value of the deviation degree, the more serious the deviation degree, and the smaller the absolute value, the closer the feature point to the threshold level. The deviation degrees of all feature points in the rotor flicker feature sequence are arranged in time sequence to form a deviation degree sequence. The horizontal axis of the deviation degree sequence is the feature point number or time, and the vertical axis is the deviation degree value. Positive deviation degree indicates that the feature is stronger than the threshold, and negative deviation degree indicates that the feature is weaker than the threshold. The deviation degree sequence clearly shows the compliance of the rotor flicker feature sequence and the dynamic determination threshold.

[0078] Anomaly clustering detection is performed on the deviation degree sequence to identify high deviation clustering areas. The points with large absolute values of deviation degree in the deviation degree sequence are analyzed, and these points correspond to the time when the feature deviates seriously from the threshold. A deviation threshold d_threshold is set, and the points with absolute values of deviation degree exceeding the threshold are marked as high deviation points. The high deviation points indicate that the feature at that time is obviously not within the typical feature range of the UAV. The deviation degree sequence is analyzed for clustering to detect whether the high deviation points are distributed in clusters on the time axis. If multiple high deviation points appear continuously within a short period of time, these points form a clustering area. The determination condition of the clustering area is that the density of high deviation points in the interval exceeds a preset density threshold. The sliding window method is used for clustering detection, and the window length is set to contain 5-10 feature points. The number of high deviation points in the window is counted. If the number of high deviation points in the window exceeds 60% of the window length, the time period corresponding to the window is marked as a high deviation clustering area. The high deviation clustering area indicates that the rotor flicker feature sequence deviates from the dynamic determination threshold continuously within the time period, and the target feature has a significant difference from the typical feature of the UAV. The starting position, length, and average deviation degree of all identified high deviation clustering areas are recorded.

[0079] According to the high deviation aggregation area, the target attribute verification excludes the candidate targets that do not meet the characteristics of the unmanned aerial vehicle. In-depth analysis is conducted on each high deviation aggregation area to verify whether the deviation of the area is caused by the special flight state of the real unmanned aerial vehicle. The special flight state includes the unmanned aerial vehicle performing maneuvering actions such as sharp turns, rapid climbs or descents, etc. In these states, the rotor flicker feature may temporarily deviate from the normal range. The unmanned aerial vehicle motion trajectory data of the time period corresponding to the high deviation aggregation area is extracted, and the speed, acceleration and direction change of the time period are analyzed. If the unmanned aerial vehicle is in a high maneuvering state during the time period, the deviation is considered to be a normal phenomenon, and the candidate target should not be excluded. If the unmanned aerial vehicle is in a stable flight state but still has high deviation, it indicates that the candidate target may not be a real unmanned aerial vehicle. Further verify the duration and frequency of the high deviation aggregation area. The high deviation of the real unmanned aerial vehicle usually has a short duration and a low frequency of occurrence. The high deviation of the interference source has a long duration or frequently occurs. For candidate targets with high deviation duration exceeding 5 seconds or frequency of occurrence exceeding 30% in the entire observation period, it is determined that they do not meet the characteristics of the unmanned aerial vehicle. The candidate targets determined not to meet the characteristics of the unmanned aerial vehicle are excluded from the unmanned aerial vehicle candidate list. These targets have certain rotor flicker characteristics, but their overall feature pattern is significantly different from that of a real unmanned aerial vehicle.

[0080] The identified candidate targets are determined to be non-target interference sources. After target attribute verification and exclusion, the identified candidate targets are objects that do not pass the unmanned aerial vehicle feature verification. These objects were marked as candidate targets in the preliminary screening because they had certain characteristics similar to rotor flicker, but after detailed comparison and verification, it was found that their feature sequence deviated from the dynamic determination threshold by a large margin, and did not meet the feature pattern of a real unmanned aerial vehicle. These identified candidate targets are collectively labeled as non-target interference sources, which include birds, insect groups, reflective objects, cloud edges, and other objects that may produce similar flicker phenomena under certain conditions. The feature sequence, occurrence position, duration, and deviation characteristics of each non-target interference source are recorded. The feature data of the non-target interference source will be used to update the background noise model, helping to improve the discrimination criteria for unmanned aerial vehicle identification. By removing explicit non-target interference sources from the candidate list, it is ensured that the final output of the unmanned aerial vehicle identification result only contains real unmanned aerial vehicle targets with high confidence. The identification and recording of non-target interference sources also provide important data for feature analysis of the monitored environment, which helps to understand the typical interference source types and interference patterns under different environmental conditions.

[0081] The background noise component is extracted from non-target interference sources. Feature analysis is performed on all identified non-target interference sources to extract common features of these interference sources. The common features include typical spectral features, time-domain fluctuation patterns, and spatial distribution rules of the interference sources. The frequency spectrum of the rotor flicker feature sequence of each non-target interference source is analyzed to identify the main frequency components in the sequence. The frequency components of real drones are concentrated in the frequency band corresponding to the rotor speed, while the frequency components of non-target interference sources are scattered or concentrated in other frequency bands. The frequency components unique to the interference sources are marked as interference frequency bands, and the energy of these frequency bands constitutes the frequency domain noise. Statistical modeling is performed on the time-domain fluctuation curve of the non-target interference sources to establish a probability distribution model of the interference fluctuation. The interference fluctuation usually exhibits randomness or low-frequency drift characteristics, which are significantly different from the periodic characteristics of the drone rotor flicker. The statistical parameters of the interference fluctuation are used as time-domain noise features. The frequency domain noise and time domain noise features are integrated to form a description model of the background noise component. The background noise component quantifies the influence mode of typical interference sources in the monitoring environment on target recognition, providing a noise benchmark for false alarm elimination.

[0082] The background noise component is used to impose false alarm elimination on the drone discrimination criterion to output the drone recognition confirmation result. The feature model of the background noise component is fused with the drone discrimination criterion, and a noise suppression mechanism is added to the discrimination criterion. The noise suppression mechanism is realized through a noise gating method. First, the noise of the input feature sequence is detected. If the feature sequence contains components similar to the background noise component, it is considered to be contaminated by noise, and the confidence of determining it as a drone is reduced. Noise detection uses similarity measurement to measure the feature distance between the input sequence and the background noise component model. If the distance is less than a threshold, it is determined to be contaminated by noise. The feature sequence that passes the noise gating is subjected to normal drone category determination, and the target category and confidence score are output. The sequence intercepted by the noise gating is marked as a false alarm and is not output as a drone target. The false alarm elimination mechanism effectively reduces the false detection caused by common interference sources such as birds, clouds, and light changes. The remaining targets after false alarm elimination are finally confirmed to verify the spatial and temporal continuity and feature stability of the targets. The spatial and temporal continuity requires the target to maintain a coherent trajectory in consecutive frames, and the feature stability requires the target features to fluctuate within a reasonable range in the time window. Through false alarm elimination and final confirmation, the drone recognition confirmation result is output, including the target's category label, position coordinates, velocity direction, and confidence score.

[0083] In order to perform the above-mentioned method embodiment corresponding to the artificial intelligence-based drone recognition method, to realize the corresponding functions and technical effects. Referring to Figure 2 , Figure 2A structural block diagram of an unmanned aerial vehicle recognition system 200 based on artificial intelligence is shown. For ease of illustration, only parts related to the present embodiment are shown. The unmanned aerial vehicle recognition system 200 based on artificial intelligence provided by the present embodiment comprises:

[0084] An image acquisition module 201 is configured to acquire an image sequence of a monitoring area, apply target-background bidirectional separation processing to the image sequence to generate an aircraft contour boundary and an environmental interference factor, and create a visual detection benchmark corresponding rule according to the aircraft contour boundary and the environmental interference factor.

[0085] A target search module 202 is configured to search a candidate target range through the visual detection benchmark corresponding rule, apply multi-scale scanning analysis to the candidate target range to generate an identification response matrix, capture a rotor flicker feature sequence along the identification response matrix, input the rotor flicker feature sequence into a deep convolution level for feature encoding to establish an unmanned aerial vehicle discrimination criterion.

[0086] A trajectory analysis module 203 is configured to perform trajectory monitoring analysis on the unmanned aerial vehicle discrimination criterion to identify a posture conversion marker, segmentally label a maneuvering amplitude attribute of the posture conversion marker to generate a rotary motion domain and a linear motion domain, monitor a transition process of the linear motion domain to the rotary motion domain to collect a motion connection degree, and use the motion connection degree to determine a flight steady state scale.

[0087] A time period calibration module 204 is configured to select an optimal capture instant in the rotary motion domain based on the flight steady state scale, extract a plurality of image features according to the optimal capture instant to generate a target feature string, calibrate a stable recognition time period through the target feature string and the environmental interference factor, and configure a dynamic determination threshold in the stable recognition time period.

[0088] A result output module 205 is configured to detect a non-target interference source through comparison between the dynamic determination threshold and the rotor flicker feature sequence, extract a background noise component from the non-target interference source, and apply false alarm screening to the unmanned aerial vehicle discrimination criterion using the background noise component to output an unmanned aerial vehicle recognition confirmation result.

[0089] The unmanned aerial vehicle recognition system 200 based on artificial intelligence described above can implement the unmanned aerial vehicle recognition method based on artificial intelligence of the method embodiment described above. The optional items in the method embodiment described above are also applicable to the present embodiment, and will not be described in detail here. The remaining content of the present embodiment can refer to the content of the method embodiment described above, and will not be described in detail in the present embodiment.

[0090] The above embodiments are also not exhaustive enumeration based on the present application, in addition to which, there can be a plurality of other embodiments not listed. Any substitution and improvement made without violating the concept of the present application is within the scope of protection of the present application.

Claims

1. An artificial intelligence-based unmanned aerial vehicle identification method, characterized in that, The method comprises the following steps: acquiring an image sequence of a monitoring area, generating an aircraft contour boundary and an environmental interference factor by applying target-background bidirectional separation processing to the image sequence, and creating a visual detection benchmark corresponding rule according to the aircraft contour boundary and the environmental interference factor; searching for a candidate target range through the visual detection benchmark corresponding rule, applying multi-scale scanning analysis to the candidate target range to generate an identification response matrix, capturing a rotor flicker feature sequence along the identification response matrix, inputting the rotor flicker feature sequence into a deep convolution level for feature coding to establish a UAV discrimination criterion; the step of capturing a rotor flicker feature sequence along the identification response matrix comprises the following steps: applying time-domain pulsation tracking to the identification response matrix to extract a response intensity fluctuation curve; implementing harmonic law verification to identify an oscillation period of the response intensity fluctuation curve; extending the oscillation period to adjacent frames to collect a sustained oscillation event; 2.The method of claim 1, wherein, forming a rotor flicker feature sequence through amplitude and phase coding of the sustained oscillation event; applying trajectory monitoring analysis to the UAV discrimination criterion to identify a posture conversion marker, segmenting the posture conversion marker according to maneuvering amplitude attributes to generate a rotary motion domain and a linear motion domain, monitoring a transition process of the linear motion domain to the rotary motion domain to collect a motion continuity degree, and using the motion continuity degree to determine a flight steady state scale, including the following steps: applying time sequence steady state inspection to the motion continuity degree to extract a stability fluctuation feature; identifying a jump protrusion point based on the stability fluctuation feature; extending the jump protrusion point to both sides to determine a steady state maintenance interval; 3.The method of claim 1, wherein, determining a flight steady state scale through a time span of the steady state maintenance interval; selecting an optimal capture instant in the rotary motion domain based on the flight steady state scale, including the following steps: comparing the motion continuity degree in the rotary motion domain with the flight steady state scale to determine a high stability time, selecting a time with the highest motion continuity degree from the high stability time as the optimal capture instant, extracting a plurality of image features from the optimal capture instant to generate a target feature string, and calibrating a stable identification period through alignment of the target feature string and the environmental interference factor; detecting a non-target interference source through comparison of the dynamic determination threshold and the rotor flicker feature sequence, extracting a background noise component from the non-target interference source, and applying false alarm screening to the UAV discrimination criterion using the background noise component to output a UAV recognition confirmation result. the step of searching for a candidate target range through the visual detection benchmark corresponding rule comprises the following steps: deriving a response potential projection from the visual detection benchmark corresponding rule; identifying a local extreme value clustering area based on the response potential projection; dividing a time period in which the local extreme value clustering area exceeds a pre-set signal threshold into a high suspicious block; determining a candidate target range according to an extended margin of the high suspicious block. the step of segmenting the posture conversion marker according to maneuvering amplitude attributes to generate a rotary motion domain and a linear motion domain comprises the following steps: applying forward and backward frame speed vector decoupling to the posture conversion marker to obtain a dynamic jump variable; A strength-increment hierarchical generation maneuver intensity level is implemented for the dynamic jump variable; A low-intensity period and a high-intensity period are determined based on the maneuver intensity level; The low-intensity period is marked as a straight-line motion domain, and the high-intensity period is marked as a rotational motion domain. 4.The method of claim 1, wherein, The stable identification period is calibrated through the calibration of the target feature string and the environmental interference factor, including: A time-axis alignment is applied to the target feature string to extract feature saliency at each time; A reverse comparison analysis is performed on the feature saliency and the environmental interference factor to obtain a signal-to-noise ratio evolution curve; A peak value interval retrieval is applied to the signal-to-noise ratio evolution curve to identify a high signal-to-noise duration; The stable identification period is calibrated according to the start and end boundaries of the high signal-to-noise duration.

5. The method of claim 1, wherein the method further comprises: The non-target interference source is detected by comparing the dynamic determination threshold with the rotor flicker feature sequence, including: The dynamic determination threshold is used as a screening reference to apply a point-by-point comparison to the rotor flicker feature sequence to obtain a deviation degree sequence; An abnormal aggregation detection is performed on the deviation degree sequence to identify a high deviation aggregation area; Target attribute verification is performed on the high deviation aggregation area to exclude candidate targets that do not meet the characteristics of the unmanned aerial vehicle; The identified candidate target is determined as a non-target interference source.

6. The method of claim 1, wherein the method further comprises: The oscillation period is identified by performing a harmonic law verification on the response intensity fluctuation curve, including: A time-frequency coupling correlation detection is performed on the response intensity fluctuation curve to obtain a correlation coefficient sequence; A peak value retrieval is performed based on the correlation coefficient sequence to extract a peak value interval time; A cycle locking strength is generated by applying stability checking to the peak value interval time; The peak value interval time is dynamically locked by the cycle locking strength to confirm it as an oscillation period.

7. The method of claim 4, wherein the method further comprises: The signal-to-noise ratio evolution curve is obtained by performing a reverse comparison analysis on the feature saliency and the environmental interference factor, including: The feature saliency is used as a signal strength reference to develop a signal reference curve along the time axis; An anti-interference weight is generated by applying an inverse mapping to the environmental interference factor; A modulation response pair is generated by performing bidirectional modulation on the signal reference curve and the environmental interference factor through the anti-interference weight; The ratio of signal components to noise components of the modulation response pair is used as the signal-to-noise ratio evolution curve.

8. An artificial intelligence-based unmanned aerial vehicle identification system, characterized by, It includes: An image acquisition module is used to obtain an image sequence of a monitoring area, and a target-background bidirectional separation processing is applied to the image sequence to generate an aircraft contour boundary and an environmental interference factor, and a visual detection reference corresponding rule is created according to the aircraft contour boundary and the environmental interference factor; A target search module is used to search for a candidate target range through the visual detection reference corresponding rule, apply a multi-scale scanning analysis to the candidate target range to generate an identification response matrix, capture a rotor flicker feature sequence along the identification response matrix, and input the rotor flicker feature sequence into a deep convolution level to perform feature encoding and establish an unmanned aerial vehicle discrimination criterion; The capturing the rotor flicker feature sequence along the identification response matrix comprises: applying time-domain pulsation tracking to the identification response matrix to extract a response intensity fluctuation curve; implementing harmonic law verification to identify an oscillation period of the response intensity fluctuation curve; extending the oscillation period to collect a sustained oscillation event; forming a rotor flicker feature sequence through amplitude-phase coding of the sustained oscillation event; The trajectory analysis module is configured to perform trajectory monitoring analysis to identify a posture conversion marker according to the UAV identification criterion, segmentally label the posture conversion marker according to a maneuvering amplitude attribute to generate a rotary motion domain and a linear motion domain, monitor a transition process of the linear motion domain to the rotary motion domain to collect a motion connection degree, and use the motion connection degree to determine a flight steady state scale, including: applying time sequence steady state inspection to the motion connection degree to extract a stability fluctuation feature; identifying a jump protruding point based on the stability fluctuation feature; determining a steady state maintenance interval according to the jump protruding point; and determining a flight steady state scale through a time span of the steady state maintenance interval. The time period calibration module is configured to select an optimal capture moment in the rotary motion domain based on the flight steady state scale, including: comparing the motion connection degree in the rotary motion domain with the flight steady state scale to determine a high stability time, selecting a time with the highest motion connection degree from the high stability time as the optimal capture moment, extracting a target feature string from a plurality of image features according to the optimal capture moment, and calibrating a stable identification period through the target feature string and the environmental interference factor, and configuring a dynamic determination threshold in the stable identification period. The result output module is configured to detect a non-target interference source through comparison of the dynamic determination threshold and the rotor flicker feature sequence, extract a background noise component from the non-target interference source, and use the background noise component to apply false alarm screening to the UAV identification criterion to output a UAV identification confirmation result.

Images