Accurate identification method and system based on air dynamic target

Abstract

The invention relates to the technical field of aerial target identification, and discloses an aerial dynamic target accurate identification method and system. The method comprises the following steps: acquiring multi-dimensional spectral data of a target through a multi-spectral imaging device, extracting spatial distribution characteristics, establishing a three-dimensional motion track model according to the spatial distribution characteristics, and calculating an instantaneous velocity vector; a motion deviation index set is generated in combination with preset reference motion parameters, abnormal behavior categories are divided through clustering analysis, and a key frame sequence is marked; reconstructing a local motion characteristic spectrum based on the key frame sequence, extracting morphological change parameters, and inputting the morphological change parameters into a pre-trained target recognition network to obtain a type recognition result and a confidence score; when the confidence coefficient is lower than a threshold value, activating a supplementary recognition process, acquiring high-resolution texture data, and fusing the high-resolution texture data with the initial recognition result to generate a final recognition tag; and updating the feature database and adjusting the network weight according to the final identification tag.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aerial target recognition, in particular to an aerial dynamic target accurate recognition method and system. BACKGROUND

[0002] Aerial dynamic target recognition has wide applications in military defense, aviation traffic management, low-altitude safety monitoring and other fields. With the increasing number of various aerial targets and the increasingly complex motion patterns, higher requirements are put forward for the accuracy and adaptability of the recognition technology. Currently, aerial dynamic target recognition mostly relies on single imaging devices to collect data. Such devices can only obtain limited dimensional information of the target, and it is difficult to fully reflect the characteristic attributes of the target, which leads to the inability to accurately capture the spatial distribution characteristics of the target in the subsequent feature extraction link, and further affects the judgment of the target motion state.

[0003] In establishing the target motion trajectory model, the prior art often directly uses the original coordinate data without effective smoothing processing of the data, so that the generated trajectory curve has much noise interference and cannot accurately present the actual motion path of the target. On this basis, the trajectory point instantaneous velocity vector calculated also has large errors and cannot truly reflect the motion speed change of the target. Meanwhile, in analyzing the target motion deviation, the prior art only judges whether the target motion is abnormal by simply comparing the speed parameters without comprehensively considering other key motion parameters such as acceleration, so that the generated motion deviation index is not comprehensive enough and cannot accurately distinguish between normal motion and abnormal motion of the target, which is prone to misjudgment or omission.

[0004] In the process of identifying the target type, the prior art mostly uses a single recognition network to process the extracted feature parameters. When the confidence score output by the recognition network is low, there is no effective supplementary recognition mechanism, the key feature information of the target cannot be further obtained, and the recognition result can only be output by relying on the existing information, which leads to insufficient reliability of the final recognition result. In addition, after completing a target recognition, the prior art does not establish an associated updating mechanism between the recognition result and the feature database and the recognition network parameters, the feature database cannot timely incorporate new target feature information, and the weight parameters of the recognition network cannot be adjusted according to the actual recognition situation, so that the recognition method has poor adaptability when facing new types or new motion patterns of aerial targets, and it is difficult to maintain stable recognition performance. SUMMARY

[0005] The present application aims to provide an aerial dynamic target accurate recognition method and system to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides an aerial dynamic target accurate recognition method, which comprises: The multi-dimensional spectral data of the aerial dynamic target is collected by a multispectral imaging device, and a spatial distribution feature in the multi-dimensional spectral data is extracted; A three-dimensional motion trajectory model of the dynamic target is established according to the spatial distribution feature, and an instantaneous speed vector of each trajectory point in the three-dimensional motion trajectory model is calculated; In combination with the instantaneous speed vector and a preset reference motion parameter, a motion deviation index set of the dynamic target is generated; The motion deviation index set is subjected to cluster analysis, an abnormal behavior category of the dynamic target is divided, and a key frame sequence corresponding to the abnormal behavior category is marked; A local motion feature graph of the dynamic target is reconstructed based on the key frame sequence, and a morphological change parameter in the local motion feature graph is extracted; The morphological change parameter is input into a pre-trained target recognition network, and a type recognition result and a confidence score of the dynamic target are output; When the confidence score is lower than a preset threshold, a supplementary recognition process is activated, and high-resolution texture data of the dynamic target is acquired; The high-resolution texture data and the type recognition result are fused, and a final recognition label of the dynamic target is generated; According to the final recognition label, a feature database of the dynamic target is updated, and a weight parameter of the target recognition network is adjusted.

[0007] Preferably, the spatial distribution feature in the multi-dimensional spectral data is extracted, including: The multi-dimensional spectral data is subjected to time domain alignment processing, and a standardized spectral sequence is generated; An energy mutation region in the standardized spectral sequence is detected, and a spectral gradient distribution of the energy mutation region is calculated; Based on the spectral gradient distribution, a contour boundary of the dynamic target is divided, and a geometric topological structure corresponding to the spatial distribution feature is generated.

[0008] Preferably, the three-dimensional motion trajectory model of the dynamic target is established according to the spatial distribution feature, including: The geometric topological structure is used to determine a center of mass coordinate sequence of the dynamic target; The center of mass coordinate sequence is subjected to Kalman filter smoothing processing, and a continuous trajectory curve is generated; Discrete trajectory points are sampled on the continuous trajectory curve, and a tangent direction angle of each discrete trajectory point is calculated.

[0009] Preferably, the instantaneous speed vector of each trajectory point in the three-dimensional motion trajectory model is calculated, including: A time stamp difference and a spatial displacement amount of adjacent discrete trajectory points are acquired; Determine the instantaneous speed of the dynamic target according to the ratio of the timestamp difference and the spatial displacement; Generate the polar coordinate expression corresponding to the instantaneous speed vector in combination with the tangent direction angle.

[0010] Preferably, the generation of the motion deviation indicator set of the dynamic target comprises: Extract the typical speed range and acceleration threshold from the preset reference motion parameters; Compare the deviation degree of the instantaneous speed vector and the typical speed range to generate a first deviation coefficient; Detect the acceleration mutation event of the dynamic target and count the frequency of the acceleration mutation event exceeding the acceleration threshold to generate a second deviation coefficient; Combine the first deviation coefficient and the second deviation coefficient into the motion deviation indicator set.

[0011] Preferably, the division of the abnormal behavior category of the dynamic target comprises: Perform spatial density analysis on the motion deviation indicator set using a density clustering algorithm; Identify the core data points of the high-density area and map the core data points to the key frame sequence; Divide the level labels corresponding to the abnormal behavior categories according to the distribution density of the core data points.

[0012] Preferably, the reconstruction of the local motion feature map of the dynamic target based on the key frame sequence comprises: Extract the region of interest of the dynamic target from the key frame sequence; Perform optical flow field calculation on the region of interest to generate a motion vector distribution map; Extract the frequency domain feature components corresponding to the morphological change parameters based on the motion vector distribution map.

[0013] Preferably, the activation of the supplementary identification process comprises: Control the high-resolution imaging device to zoom in and track the dynamic target for shooting; Perform super-resolution reconstruction on the image sequence obtained by shooting to generate the high-resolution texture data; Extract the microstructure features in the high-resolution texture data through a convolutional neural network.

[0014] Preferably, the updating of the feature database of the dynamic target according to the final identification label comprises: Perform similarity matching of the final identification label and historical identification records; When the matching degree is lower than the preset standard, create a new target feature entry; Optimize the convolution kernel size of the target recognition network based on the new target feature item.

[0015] Preferably, the present application also includes an air dynamic target precision identification system based on the air, comprising a memory, a processor and a computer program stored in the memory and running on the processor, the processor implements the steps of the above-mentioned air dynamic target precision identification method when executing the computer program.

[0016] Compared with the prior art, the present application has the following advantages: The present method can obtain more dimensional feature information of the target by using a multi-spectral imaging device, compared with a traditional single imaging device, which provides sufficient data basis for subsequent precise extraction of spatial distribution features, so that the extracted spatial distribution features are more consistent with the actual feature attributes of the target, which helps to more accurately grasp the spatial form and position information of the target, and provides comprehensive feature support for subsequent establishment of a three-dimensional motion trajectory model.

[0017] In the step of establishing a three-dimensional motion trajectory model, the present method is based on the extracted spatial distribution features, rather than directly using the original data, which can reduce the interference of irrelevant information on model establishment, and calculate the instantaneous velocity vector of each trajectory point in the model, so as to comprehensively grasp the motion speed of the target at different times, clearly present the motion speed change trend of the target, and provide accurate motion parameter basis for subsequent analysis of target motion deviation, which helps to more deeply understand the motion state of the target.

[0018] The generation of the motion deviation index set combining the instantaneous velocity vector and the preset reference motion parameter can comprehensively judge the difference between the target motion and the reference motion from the speed dimension, and subsequent processing of the motion deviation index set through clustering analysis can effectively distinguish different types of motion deviation, and then accurately classify the abnormal behavior categories of the dynamic target, while marking the corresponding key frame sequence, which facilitates subsequent focusing on key information to carry out further feature extraction and analysis work, and reduces the interference of invalid information on the identification process.

[0019] The reconstruction of the local motion feature spectrum of the dynamic target based on the key frame sequence can focus on the key stage in the motion process of the target, extract the morphological change parameters therein, so that the obtained feature parameters are more targeted, input these parameters into the pre-trained target recognition network, fully exert the feature processing capability of the network, output the recognition result and confidence score which are more consistent with the actual type of the target, and provide intuitive basis for judging the reliability of the recognition result.

[0020] When the confidence score is lower than the preset threshold, a supplementary identification process is activated to obtain high-resolution texture data, which can further obtain detailed feature information of the target in the case of insufficient original identification information, and the high-resolution texture data and the type identification result are fused to generate a final identification label, which can make up for the defects of the original identification information, improve the accuracy and reliability of the final identification label, and avoid deviation of the identification result caused by insufficient information.

[0021] The feature database of the dynamic target is updated according to the final identification label, which can continuously enrich the target feature information in the database, so that the database always maintains comprehensive coverage of various target features, and the weight parameters of the target identification network are adjusted, so that the network can better adapt to different types and different motion modes of the air dynamic target in the subsequent identification process, enhance the adaptability of the identification method in complex application scenarios, ensure long-term stable identification performance, and meet the actual needs of different fields for air dynamic target identification. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A working principle diagram of the air dynamic target accurate identification method is described. Figure 2 A flowchart of multi-dimensional spectral data space distribution feature extraction is described. Figure 3 A flowchart of dynamic target three-dimensional motion trajectory model establishment is described. Figure 4 A Kalman filter state estimation and mean square error analysis diagram of the dynamic target three-dimensional motion trajectory is described. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0024] Please refer to Figure 1The application provides a method and system for accurate identification of aerial dynamic targets. The method includes multi-dimensional spectral data acquisition, motion trajectory modeling, abnormal behavior analysis, feature map reconstruction, and multi-source data fusion processing, achieving comprehensive and accurate identification of aerial dynamic targets. Multi-dimensional spectral data is synchronously acquired by multi-spectral imaging equipment on an observation platform. Spatial distribution feature extraction includes noise elimination and geometric contouring. A three-dimensional motion trajectory model is constructed based on the position changes of the target in consecutive frames. Instantaneous velocity vectors are obtained by differential operation of trajectory points. A set of motion deviation indicators is generated by comparing real-time data with historical parameters. Cluster analysis uses an unsupervised learning algorithm to identify abnormal patterns. Local motion feature maps are extracted from the optical flow field in the key frame region of interest. Morphological change parameters are processed through frequency domain transformation and feature dimension reduction. The target recognition network uses a pre-trained convolutional neural network. A high-resolution acquisition process is started when the confidence level is insufficient. The final identification label is used to optimize system performance. This method is suitable for aerial target monitoring in complex environments and ensures identification accuracy through a multi-stage processing flow. The output of the previous step provides input data for the subsequent processing, forming a complete target identification chain. Multi-dimensional spectral data provides rich spectral feature information. Motion trajectory analysis captures the behavior characteristics of the target. Feature map reconstruction enhances the expression of morphological features. Multi-source data fusion improves the reliability of identification. The system has self-learning ability and can continuously optimize identification performance through database updates and network parameter adjustments.

[0025] Example 1: see Figure 2 The spatial distribution features in the multi-dimensional spectral data are extracted, which involves a series of ordered processing and transformation of the original multi-dimensional spectral data. The original data stream captured by the multi-dimensional spectral imaging device usually contains timestamp information and intensity information of multiple spectral bands (such as visible light, near-infrared, thermal infrared, etc.). These data may have slight desynchronization in the time domain, which is caused by the response time difference of each channel of the device or the delay in the data transmission process. Time domain alignment processing requires parsing the timestamp attached to each data packet and selecting a reference time reference (usually the main clock signal or the acquisition time of a certain specified band). Through interpolation algorithms, the data points of other bands are resampled on the time axis, so that all bands have corresponding intensity values at the same time. Linear interpolation or cubic spline interpolation is commonly used in this process to smoothly reconstruct the data. Generating standardized spectral sequences is the basis for subsequent processing. Standardization aims to eliminate intensity value fluctuations caused by changes in lighting, atmospheric conditions, or differences in sensor sensitivity. A common method is to normalize the intensity sequence of each band, such as mapping it to the [0, 1] interval or converting it to a distribution with a mean of 0 and a standard deviation of 1. This step makes spectral data from different sources and collected at different times comparable.

[0026] Detecting energy abrupt change regions in a standardized spectral sequence requires a sliding window analysis technique. A fixed or adaptive-sized window moves pixel-by-pixel or region-by-region across the image sequence, calculating the variance or standard deviation of the spectral intensity within each window. When this statistic exceeds a preset threshold, the region is marked as an energy abrupt change region. The threshold is not fixed; it can be dynamically adjusted based on the overall noise level of the image. For example, a global statistic (such as the average variance of all pixels plus a certain multiple of the standard deviation) can be used as the threshold benchmark. Energy abrupt change regions typically correspond to the edges of moving targets, areas of abrupt changes in surface material, or parts with significantly different radiation characteristics from the background. Calculating the spectral gradient distribution of these energy abrupt change regions further reveals the contour characteristics of the target. Spectral gradient calculation not only focuses on the spatial gradient within a single band but also involves cross-band gradient analysis, comparing the rate of intensity change at the same spatial location in different bands. Ultimately, this yields the gradient magnitude and direction information of each pixel in different spectral dimensions, which together constitute a multidimensional gradient field. Specifically, after completing temporal alignment and standardization, spatial gradient extraction is performed frame-by-frame on the multispectral images within each labeled energy abrupt change region. Specifically, a 3x3 Sobel kernel (the Sobel kernel is a discrete convolution template pair used to approximate the first-order spatial derivative of an image) is used to convolve in the horizontal and vertical directions to obtain the gradient components of each band. Then, based on the channel weight table generated in the radiometric calibration stage, the gradient components of each band are calibrated and linearly fused to obtain a pair of fused gradient component maps. Subsequently, the magnitude and direction of the fused gradient are calculated at the pixel level, and a small-scale Gaussian smoothing is performed within the region to suppress isolated noise points, thereby forming a spectral gradient distribution map (composed of an amplitude map and a direction map). This distribution map is directly used for subsequent contour boundary division and geometric topology generation, ensuring that the contour information simultaneously reflects the material response differences across bands and the spatial variation characteristics of a single band.

[0027] Delineating the contour boundaries of dynamic targets based on spectral gradient distribution is a process of pixel classification and region aggregation. Pixels with high gradient magnitudes are initially considered as potential edge points, but these may contain a large number of false edges caused by noise. Region growing algorithms can start from these seed points and, based on the continuity of gradient directions and the spatial proximity between pixels, aggregate pixels with similar gradient characteristics to form connected boundary segments. Morphological operations (such as opening and closing operations) are often used in subsequent processing to connect broken boundary segments and fill small holes, thereby generating a complete and closed contour. Specifically, non-maximum suppression is performed along the gradient direction on the fused gradient magnitude map to extract local maxima as seed points. Region growth is then performed using an eight-neighborhood connection criterion, oriented towards directional consistency. The growth rule is: when the angle between a candidate pixel and the average direction of the current boundary segment is less than a set angle threshold and the gradient magnitude is not lower than a local adaptive threshold, the pixel is incorporated into that segment. During the growth process, the average direction within the segment is continuously updated. After obtaining several connected boundary segments, an opening operation (structuring element is 3x3 cross) is applied to the binary boundary map to remove burrs and isolated short segments. Then, a closing operation (structuring element is 5x5 square) is applied to bridge small gaps and smooth notches. Subsequently, contour tracking based on Moore's neighborhood is used to generate closed contours. Hole filling based on seed filling and topological consistency verification output a set of closed and self-intersecting contours, which is used to generate the geometric topology. This contour is ultimately expressed as a geometric topology, which can be a polygonal chain code describing the contour boundary or a more complex mesh representation (such as a triangular mesh). It encodes the shape, size, and spatial connection relationships of the target's components, providing precise spatial constraints for subsequent 3D motion trajectory modeling.

[0028] The specific implementation details of temporal alignment processing also include the parsing of hardware synchronization signals. Multispectral imaging devices are typically equipped with synchronization signal input / output interfaces to ensure that multiple sensors can be driven by the same external trigger signal. When the external trigger is unavailable, post-synchronization processing relies on internal clocks and software timestamps. The generation of normalized spectral sequences sometimes requires consideration of radiometric calibration, converting the digital quantization values ​​output by the sensors into physically meaningful radiance values, followed by normalization. This improves the physical consistency and accuracy of data processing. Algorithms for detecting energy abrupt changes in regions need to be optimized for computational efficiency because they handle high-frame-rate, multi-band data streams. Integral image or pyramid multi-resolution analysis techniques can accelerate variance calculation for large-area images. The calculation of spectral gradient distribution has various variations in practice. Some methods calculate histograms of gradient magnitudes (histogram of gradient directions, HOG) to form a description of the local shape; however, in this application, the focus is more on using gradient information for precise boundary localization. Specifically, local descriptive units are constructed using fixed-size sliding windows within energy mutation regions. After dividing the window into equal-sized grid cells, the magnitude and direction of the aforementioned fusion gradient are accumulated in each cell according to directional intervals to form an orientation histogram. Then, block-level normalization is performed by grouping adjacent 2x2 cells. After completion, multi-spectral band HOG descriptors are constructed using a "band-by-band calculation and fixed-order splicing" method and registered on the dense descriptor map at the center of the window. This dense map is used in conjunction with the boundary map: boundary responses caused solely by noise and lacking a stable orientation histogram pattern are suppressed, while boundary segments consistent with the target's local shape pattern are preserved and extended. The generated HOG features serve as both a criterion for contour integrity and an auxiliary channel for subsequent local morphological feature extraction and recognition network input, thereby stably expressing the target's local shape changes in the keyframe sequence.

[0029] During the contour boundary segmentation process, the accuracy of the generated geometric topology directly affects the accuracy of subsequent centroid calculation. Therefore, sub-pixel edge detection technology is sometimes used to improve the boundary positioning accuracy. The generated topology data is usually stored in vector format, which facilitates geometric transformation and feature calculation.

[0030] Multidimensional spectral data inherently has high information redundancy. Temporal alignment and normalization effectively compress the data volume and improve data quality, creating conditions for subsequent feature extraction. The detection of energy abrupt change regions is essentially a saliency detection method, helping the system concentrate computational resources on areas where the target may appear, rather than uniformly processing the entire image. Spectral gradient distribution provides richer boundary information than single-band gradients because different materials may have different edge responses at different bands. Integrating gradient information from all bands can generate more robust contours. The resulting geometric topology is an abstract yet precise shape representation that transforms the visual appearance of the target into a set of geometric parameters that can be used for quantitative analysis and model calculations. This structure is a crucial bridge connecting the underlying spectral data with high-level motion analysis.

[0031] Example 2: See Figure 3 A three-dimensional motion trajectory model of a dynamic target is established based on its spatial distribution characteristics. This begins by determining the target's centroid coordinate sequence from the generated geometric topology. The geometric topology clearly expresses the target's spatial shape and position in each frame of the image in the form of a point cloud or mesh. The calculation of the centroid coordinates is not simply the arithmetic mean of the pixels, but comprehensively considers the density distribution of the target's shape. For non-uniform or complex dynamic targets, a weighted centroid algorithm can more accurately reflect the spatial displacement of its center of mass. After each frame of the image is processed, the system records the specific position of the centroid in the three-dimensional coordinate system at that moment, thus forming a centroid coordinate sequence arranged in chronological order. This sequence is the most direct digital representation of the target's motion path, but it usually contains measurement errors introduced by sensor noise, image jitter, or background interference.

[0032] Smoothing the centroid coordinate sequence using Kalman filtering is a core step in improving the accuracy of the trajectory model. As an optimal estimation algorithm, Kalman filtering effectively suppresses the influence of random noise through two alternating stages: state prediction and measurement update. Its state equation is based on the physical laws of the target's motion, such as assuming the target moves at a constant or uniformly accelerated speed over a short period. The observation equation is directly related to the centroid coordinate measurements extracted from the image. The filtering process continuously fuses the noisy measured centroid positions with the model predictions, generating a smoother, more continuous trajectory curve. This curve is presented in the form of parametric equations or spline functions, not only filtering out abnormal jumps but also maintaining the smoothness and dynamic rationality of the trajectory.

[0033] On the generated continuous trajectory curve, a set of discrete trajectory points needs to be sampled for subsequent instantaneous motion analysis. The sampling strategy can be selected based on actual needs, choosing between equal time intervals or equal spatial intervals. Equal time interval sampling can uniformly capture the temporal characteristics of the target motion, while equal spatial interval sampling provides denser data points in areas with high trajectory curvature to better describe the geometric details of the path. For each sampled discrete trajectory point, its tangent direction angle needs to be calculated, which defines the instantaneous forward direction of the trajectory curve at that point. The tangent direction angle is calculated by solving the first derivative of the trajectory curve function. In the discrete case, it can be approximated by the coordinate difference between adjacent points, or a more accurate result can be obtained by curve fitting and differentiation. To calculate the instantaneous velocity vector of each trajectory point in the 3D motion trajectory model, it is first necessary to accurately obtain the timestamp difference between adjacent discrete trajectory points and their displacement in 3D space. The timestamp difference is obtained directly from the metadata of the image acquisition system, with an accuracy typically reaching the millisecond level, while the spatial displacement is obtained by calculating the Euclidean distance between two adjacent points, which is a scalar value. Dividing the spatial displacement by the timestamp difference yields the average velocity of the dynamic target within that interval, serving as an approximate estimate of the instantaneous velocity. To obtain a more accurate instantaneous velocity, the central difference method can be used, which utilizes information from the previous and next trajectory points to calculate the velocity, thereby reducing errors.

[0034] After determining the instantaneous velocity magnitude, the instantaneous velocity vector needs to be fully expressed by combining it with the previously calculated tangent direction angle. Velocity is a vector, possessing both magnitude and direction. In three-dimensional space, the velocity vector can be represented by three components (Vx, Vy, Vz) in Cartesian coordinates, but converting it to polar coordinates often more intuitively reflects the directional characteristics of the motion. Polar coordinates describe the vector using velocity magnitude and direction angles (e.g., pitch and azimuth). This form facilitates intuitive comparison and deviation analysis with preset benchmark motion parameters, such as directly determining whether the velocity direction deviates from the expected flight path. The entire implementation process relies on stable and accurate centroid information provided by spatial distribution characteristics. The parameters of the Kalman filter, such as process noise covariance and measurement noise covariance, need to be pre-calibrated and adjusted according to the target's motion characteristics and sensor performance in the specific application scenario. The specific calibration process can adopt the following steps: First, acquire sensor data in a static environment, and determine the initial value of the measurement noise covariance R by statistically analyzing the variance of the measured values. A typical value is 0.1-0.5 pixels². The initial value of the process noise covariance Q is set based on the target motion characteristics. For a uniform motion model, the diagonal element of Q is typically set to 0.01-0.1 pixels² / second². In practical applications, adaptive filtering techniques are used to dynamically adjust Q and R based on the statistical characteristics of the innovation sequence (the difference between observed and predicted values): when the variance of the innovation sequence is consistently greater than the theoretical value, Q is appropriately increased; when the innovation sequence shows a significant correlation, R is adjusted. The convergence criterion for filtering is that the mean of the innovation sequence is less than 3 times the standard deviation over 10 consecutive periods. The sampling density of discrete trajectory points is a factor that needs to be balanced; excessively dense sampling increases computation and may introduce noise, while excessively sparse sampling will lose motion details. The accuracy of instantaneous velocity vector calculation directly affects the reliability of subsequent motion deviation analysis; therefore, error propagation must be strictly controlled at every stage of the data processing chain.

[0035] See Figure 4This paper demonstrates the modeling effect and performance analysis of the Kalman filter algorithm in the 3D trajectory estimation of dynamic targets, corresponding to the quantitative evaluation of the target's motion trajectory along the Y-axis (horizontal direction) and Z-axis (vertical direction). The comparison chart of the Kalman filter effect in the Y-direction reveals three types of data comparisons regarding the evolution of the Y-direction coordinate over time: the true trajectory is represented by a solid line, presenting the ideal motion path; the original measurement data with Gaussian white noise is shown by a dashed line, highlighting the random fluctuations introduced by the sensor; the estimated trajectory after Kalman filtering is marked with dotted lines, effectively suppressing noise through state prediction and measurement update mechanisms, significantly improving trajectory smoothness, and keeping the root mean square error with the true path at a low level. The comparison chart of the Kalman filter effect in the Z-direction focuses on the motion characteristics in the Z-direction. The periodic changes of the true trajectory (peak points t=2.5s, t=15s and valley points t=7.5s, t=12.5s) are highly consistent with the filtered estimation results, with the maximum residual range limited to ±0.5 meters, verifying the robustness of the filter in nonlinear motion scenarios. Both subplots employ a high-contrast visualization design, with the time axis scaled in 0.1-second increments and the vertical axis dynamically adapting to the motion amplitude, conforming to engineering analysis standards. The core advantage of Kalman filtering lies in its state-space model's synergistic optimization of dynamic priors and observation noise. By adjusting the process noise covariance (typically 0.01–0.1 unit² / s²) and the measurement noise covariance (typically 0.1–0.5 unit²), it achieves asymptotic unbiasedness in trajectory estimation, providing a high signal-to-noise ratio data foundation for multidimensional motion analysis.

[0036] Example 3: Generating a set of motion deviation indicators for dynamic targets. The first step is to systematically extract typical speed ranges and acceleration thresholds from preset benchmark motion parameters. These benchmark parameters constitute a reference system for judging whether the target's motion is abnormal. Typical speed ranges are usually determined based on historical data statistics or prior knowledge of the target. They define the reasonable upper and lower limits of speed that such targets may exhibit in specific scenarios, such as the cruising speed range of large aircraft or the typical flight speed band of small drones. Acceleration thresholds set the limit value of acceleration changes during normal maneuvers, used to capture violent movements that exceed normal maneuverability. This threshold may need to distinguish between positive acceleration (acceleration) and negative acceleration (deceleration) and set them separately.

[0037] To generate a first deviation coefficient, the deviation of the instantaneous velocity vector from a typical velocity range is compared. This process requires separating the magnitude of the instantaneous velocity (i.e., speed) from the velocity vector for scalar comparison. A common approach is to calculate the relative degree of deviation of the instantaneous speed from the midpoint of the typical velocity range and then normalize it. Assume the typical velocity range is... ,in The median value of the range represents the instantaneous rate. Then the first deviation coefficient It can be calculated using the following formula: ; in: This represents the first deviation coefficient obtained from the calculation. This represents the instantaneous velocity of the dynamic target calculated from the three-dimensional motion trajectory model. and These are the upper and lower limits of the typical speed range extracted from the reference motion parameters, respectively. It is the median of this typical speed range.

[0038] The system detects abrupt acceleration events of dynamic targets and counts the frequency of these events exceeding an acceleration threshold to generate a second deviation coefficient. Detecting these abrupt acceleration events requires first calculating the rate of change between consecutive instantaneous velocity vectors. By calculating the ratio of the velocity vector difference at adjacent sampling time points to the time interval, the average acceleration vector of the target during that time period can be obtained. When the magnitude of the acceleration vector (i.e., the magnitude of acceleration) exceeds a preset acceleration threshold, an abrupt acceleration event is recorded. The generation of the second deviation coefficient not only counts the number of such events occurring within a set observation time window but may also consider the magnitude of acceleration exceeding the threshold in each event, thus forming a comprehensive quantitative indicator reflecting the intensity or stability of the target's motion.

[0039] The first deviation coefficient and the second deviation coefficient are combined into a set of motion deviation indices, which can be represented as a multi-dimensional vector, for example... This includes deviation information quantified from different kinematic dimensions (velocity, acceleration). A set of indicators representing the motion deviation of dynamic targets. This represents the first deviation coefficient. This represents the second deviation coefficient. This set of indicators provides a feature space for subsequent clustering analysis, where each dynamic target's series of motion states during observation can be mapped to a point or trajectory in this space. The core of classifying the abnormal behavior categories of dynamic targets is to perform spatial density analysis on the motion deviation indicator set using density clustering algorithms. Density clustering algorithms such as DBSCAN (Density-Based Spatial Clustering with Noise) do not require pre-specifying the number of categories and can automatically form clusters based on the distribution density of data samples in the feature space. This algorithm divides data points into three categories: core points (located inside dense regions), boundary points (located at the edges of dense regions), and noise points (far from any dense region). By finding high-density regions separated by low-density regions, the algorithm can identify different clusters of abnormal behavior patterns. The core data points in high-density regions are identified and mapped to keyframe sequences. In the DBSCAN algorithm, if a point's neighborhood contains a sufficient number of other points (reaching a set minimum), that point is marked as a core point. The connected regions formed by these core points represent data-dense patterns in the feature space, i.e., typical abnormal behavior patterns. Each core data point corresponds to a specific moment on the original data acquisition timeline. By backtracking through timestamps, the frame in the original image sequence or trajectory data where the abnormal behavior occurred can be accurately located. These marked frames constitute the keyframe sequence.

[0040] The system categorizes anomalous behaviors into severity categories based on the distribution density of core data points. After clustering, each discovered cluster (composed of a series of core points and their connected boundary points) represents a type of anomalous behavior. Information such as the density of core points within a cluster, the cluster diameter, and the location of the cluster center in the feature space (i.e., the magnitude of its average deviation coefficient) can be used to define the severity level or anomaly type of that behavior. For example, a cluster with generally high first and second deviation coefficients and dense point distribution might be labeled as a "severe anomaly"; while another cluster with relatively low deviation coefficients and more dispersed point distribution might be labeled as a "minor anomaly." This categorization of severity labels allows the system to differentiate and prioritize identified anomalous behaviors. The process relies on the rationality and accuracy of the baseline motion parameters; inappropriate parameters can lead to excessive false positives or false negatives. The deviation coefficient design needs to have good discriminative power, effectively amplifying the differences between anomalous and normal behaviors.

[0041] The parameter settings of density clustering algorithms, especially the neighborhood radius and minimum number of points, have a significant impact on the clustering results. The implementation process of density clustering algorithm parameters begins with an in-depth analysis of the data characteristics of the motion bias index set. The motion bias index typically includes multiple dimensions such as the first bias coefficient and the second bias coefficient. The distribution characteristics of these data points in the high-dimensional feature space directly determine the rationality of parameter selection. The algorithm needs to be able to adapt to the differences in data distribution under different task scenarios. The neighborhood radius Eps is usually determined using the k-distance graph method as an auxiliary tool. This method calculates the distance from each data point to its k-th nearest neighbor and plots the sorted distance curve. Observing this curve, abrupt changes in distance can be found. The distance value corresponding to this abrupt change is often used as a candidate value for Eps. This position means that points with a distance less than this value will form a relatively dense region. In practice, the value of k is generally about twice the number of data dimensions, but it needs to be adjusted according to the specific data distribution.

[0042] The minimum number of points (MinPts) is interrelated with and needs to be determined collaboratively with the Eps value. This parameter essentially defines the minimum number of samples required to form a cluster. A MinPts that is too small may lead to misclassifying noise points as clusters, while a MinPts that is too large may treat real small clusters as noise. Generally, the value of MinPts should not be less than the data dimension plus one, while also considering the minimum size of the desired cluster. The dynamic adjustment mechanism for parameter settings needs to consider changes in the actual application environment, as the motion deviation characteristics of different spatial domains and different types of targets may vary significantly. The system can maintain multiple parameter configuration templates and automatically select the most suitable parameter set based on the real-time monitored target type and motion characteristics. When the monitoring environment changes significantly, the system can also activate the parameter self-learning module to recalculate the optimal parameters based on newly acquired data samples.

[0043] Parameter optimization often requires manual verification incorporating domain knowledge. While methods like k-distance maps can provide reference values, the final parameter determination must consider specific task requirements. For example, in scenarios requiring high sensitivity for detecting anomalous behavior, the Eps value can be appropriately relaxed while the MinPts value can be reduced; conversely, in scenarios demanding high accuracy, stricter parameter settings should be adopted. This trade-off needs to be continuously refined through extensive experiments and feedback from practical applications. The parameter validation stage assesses the rationality of parameter settings by checking the quality of clustering results. Good clustering results should result in high similarity among samples within the same cluster, while samples from different clusters should show significant differences. It is also necessary to check whether the proportion of noise points is within an acceptable range and whether the clustering results have a reasonable correspondence with subsequent keyframe labeling results. When the clustering effect is unsatisfactory, the parameters need to be readjusted and the above process repeated. By transforming numerical motion deviation indicators into semantically meaningful anomalous behavior categories and levels, and associating them with specific time frames, the system achieves a deep understanding and structured description of dynamic target behavior patterns.

[0044] Example 4: Reconstructing the local motion feature map of a dynamic target based on a keyframe sequence. This begins by extracting the region of interest (ROI) containing the dynamic target from a keyframe sequence already labeled with anomalous behavior categories. The keyframe sequence itself contains time points where the target's behavior undergoes significant changes, such as a sudden turn, rapid ascent, or irregular maneuver by an aerial target. Extracting the ROI is not simply about preserving the entire frame image; rather, it requires precisely cropping an image region containing only the target itself, based on the target contour boundaries or detection boxes determined in previous steps. This process removes a large amount of irrelevant background information, concentrating computational resources on the target itself. The cropping operation may involve coordinate transformations and image interpolation to ensure that the ROI of the target remains spatially aligned even if the target undergoes scale changes or rotation due to motion in different frames.

[0045] Optical flow field calculations are performed on multiple extracted regions of interest (ROIs). The optical flow field reflects the motion vector of each pixel in the image across consecutive frames. The method for calculating the optical flow field is based on the assumptions of constant brightness and small motion, specifically employing the Lucas-Kanade pyramid optical flow algorithm with the following parameters: 3 pyramid layers, a scaling factor of 0.5 per layer, and a window size of 15×15 pixels. The algorithm terminates when the change in the optical flow vector is less than 0.01 pixels or the maximum number of iterations reaches 20. To compensate for the constant brightness assumption, an illumination variation model is introduced, and histogram matching is performed between each frame. After calculating the motion vector, a mean-mode filter (3×3 window) is applied to remove outliers, and a motion vector amplitude threshold of 0.5-5 pixels / frame is set; vectors outside this range are considered noise and discarded. By comparing the ROIs of adjacent keyframes, the algorithm assigns a two-dimensional motion vector to each pixel within the region. This vector describes the direction and distance of movement of the pixel from the previous frame to the current frame.

[0046] The motion vectors of all pixels together form a motion vector distribution map, which visualizes the relative motion of different parts of the target surface, such as wing twisting, rotor rotation, or overall translation. Frequency domain feature components corresponding to morphological change parameters are extracted from the motion vector distribution map. While the time-domain motion vector sequence contains rich motion information, it is difficult to directly capture periodic and regular morphological change patterns. By performing a frequency domain transformation on the time series of the motion vector distribution map (i.e., the motion vector diagram corresponding to multiple consecutive keyframes), such as using a two-dimensional Fourier transform, the motion pattern can be converted from the time domain to the frequency domain for analysis. In the frequency domain, specific frequency components with concentrated energy often correspond to the inherent, periodic motion patterns of the target, such as the rotation frequency of a propeller or the vibration frequency of the fuselage. Extracting these frequency domain feature components, such as the amplitude, phase, and frequency value of the main frequencies, yields a set of quantified parameters characterizing the local morphological changes of the target.

[0047] The condition for activating the supplementary recognition process is that the confidence score output by the pre-trained target recognition network is lower than a preset threshold, indicating that the recognition result based on the current motion characteristics and macroscopic shape has high uncertainty. The threshold is usually set according to the tolerance for misidentification in practical applications, for example, it can be set to 0.7 or 0.8. When the network's confidence in its judgment is lower than this level, the system will trigger the supplementary recognition process to obtain more refined information. Controlling the high-resolution imaging equipment to perform zoom tracking and imaging of the dynamic target is the core action of the supplementary recognition process. The multispectral imaging equipment mainly used typically performs wide-area surveillance with a high frame rate but relatively low spatial resolution, while the high-resolution imaging equipment is usually a telephoto camera or electro-optical pod with a motorized zoom lens. The system needs to transmit the real-time three-dimensional coordinates of the target to the servo control system of the high-resolution imaging equipment to drive the gimbal rotation and lens zoom, ensuring that the target remains stably in the center of the imaging field of view. Zoom tracking and imaging aims to acquire high-resolution images of the target surface, sufficient to distinguish microscopic details such as hatches, windows, paintwork, and wing surface connection structures.

[0048] Super-resolution reconstruction is performed on high-resolution image sequences. Even with high-resolution cameras, images acquired at long distances can become blurry due to factors such as atmospheric turbulence and optical diffraction limits. Super-resolution reconstruction utilizes multiple frames with sub-pixel displacements within an image sequence, fusing these subtly different images through complex algorithms to reconstruct a clear image with a spatial resolution higher than a single frame. Reconstruction algorithms may be based on interpolation, regularization constraints, or deep learning models, aiming to recover high-frequency detail information from the image. Microstructural features from high-resolution texture data are extracted using convolutional neural networks (CNNs). The super-resolution-enhanced image is then fed into a specially designed or pre-trained CNN. The network's deep convolutional layers activate and respond to local texture patterns of the target (e.g., regular patterns of skin seams, rivet arrangements, specific logos on paint). Feature vectors extracted from specific layers of the network (usually before the last fully connected layer) represent the high-dimensional features characterizing the microstructure of the target surface. These features are crucial for distinguishing targets that look similar but are of different models. Table 1 shows the basic information of three regions of interest extracted from a keyframe sequence and their calculated main morphological change parameters (frequency domain feature components). It should be noted that the actual number and dimensionality of the parameters will be higher.

[0049] Table 1: Region of Interest in Keyframes and Extracted Morphological Change Parameters , Taking the identification of a maneuvering drone as an example, a keyframe sequence might capture the moment it hovers or changes course. Regions of interest extracted from these keyframes focus on the drone's body. Optical flow calculations might reveal strong, vortex-like motion vector patterns in the rotor region, as well as relatively smooth translational vectors of the body. Frequency domain analysis of several consecutive frames of rotor region motion vector maps can extract the dominant frequency component consistent with the rotor speed; its frequency value (e.g., 120Hz) and amplitude (characterizing the prominence of rotor motion in the image) constitute part of the morphological change parameters. If the target recognition network cannot determine with high confidence whether the target is a quadcopter or a hexacopter drone based solely on these motion parameters (confidence score 0.65 < threshold 0.75), the system activates a supplementary recognition process. Upon receiving the instruction, a high-resolution imaging device quickly zooms in and locks onto the target, capturing a series of close-up images.

[0050] After super-resolution reconstruction, these images clearly show the shape of the drone's landing gear and the specific structure of the gimbal camera. A convolutional neural network extracts features from these details, analyzes them to confirm the landing gear is of a specific style, thus refining the identification from the general term "drone" to the "DJIMatrice300" model, generating a high-confidence final identification label. The accuracy and adaptability of the baseline motion parameters need to cover possible variations in normal operation to avoid false alarms. The design of the bias coefficients needs to balance sensitivity and robustness, enabling timely capture of real anomalies while also tolerating a certain degree of tolerance for measurement noise. Parameter tuning of the density clustering algorithm is a highly empirical process, potentially requiring learning from historical anomaly data or the involvement of domain experts. By associating abstract numerical biases with specific time frames and assigning them semantic category and level labels, the system completes the transformation from low-level motion data to high-level behavioral understanding.

[0051] Example 5: Updating the feature database of dynamic targets based on the final identification label. This process begins by matching the newly generated final identification label with the historical identification records stored in the database. The final identification label is a structured data object containing information such as target type, model, confidence level, key feature vectors, and timestamps. Historical identification records are collections of successfully identified and verified target feature entries, forming the system's existing knowledge base. Similarity matching is not a simple string or type name comparison, but rather a distance calculation based on feature vectors, such as using cosine similarity or Euclidean distance to measure the closeness of the new label's feature representation to the feature representations of existing entries in the database. The matching process may require traversing all records under the same major category in the database, or using indexing techniques to accelerate the query.

[0052] When the matching degree is lower than the preset standard, the system determines that the currently identified target may be a new variant, a previously insufficiently recorded model, or an unprecedented high-quality feature data obtained due to excellent observation conditions. In this case, the system will not force it into an existing category, but will create a new target feature entry. This new entry will be assigned a unique identifier and will record all relevant information about this identification event in detail: including but not limited to the final identification label, the original feature data used to generate the label (such as multispectral features, motion trajectory features, high-resolution texture features), keyframe sequence index, calculated morphological change parameters, and environmental context information during identification. The creation of new entries enriches the system's knowledge base, enabling it to identify more diverse targets. The kernel size of the target recognition network is optimized based on the new target feature entries; model optimization is a careful process. The convolutional kernel is the basic weight unit in a convolutional neural network used to extract local features, and its size directly affects the size of the network's receptive field and feature extraction capability.

[0053] Once the system learns new target patterns through added feature entries, it may need to adjust the convolutional kernel size to better adapt the network to these new features. For example, if the new target entries show certain discriminative features as larger-scale texture patterns, appropriately increasing the convolutional kernel size of shallower networks may help capture this information; conversely, if the new features are very fine-grained, smaller, denser convolutional kernels may be needed to extract details. The optimization process is typically not performed on a real-time recognition line, but rather periodically or after accumulating a sufficient number of new samples, initiating a network retraining or fine-tuning process. In this process, an expanded feature database containing the new entries is used as the training set, and the network is trained on top of the original network weights, allowing parts of the network structure (such as the convolutional kernel size of certain layers) to adaptively adjust according to the learning objective.

[0054] To illustrate with a concrete example: Suppose the system identifies an aerial target. Through multispectral analysis, trajectory tracking, and high-resolution detail verification, it ultimately identifies it as a "Tianying-2B" UAV and generates a final identification tag. The system then performs a similarity match between the feature vector of this tag and existing records under the "UAV" category in the feature database. The database contains records for models such as "Tianying-1A," "Tianying-2A," and "Xunfeizhe-3." The calculation shows that the new tag has a similarity of 0.75 with the "Tianying-2A" record and 0.65 with the "Xunfeizhe-3." Although it is relatively close to "Tianying-2A," it fails to reach the system's set merging threshold (e.g., 0.85). The system determines that this is a new model that is similar to "Tianying-2A" but has significant differences.

[0055] The system creates a new feature entry named "Tianying-2B" under the "UAV" category in the database. This entry stores the unique features identified, such as the shape of its winglets differing from "Tianying-2A," subtle differences in its spectral response in specific infrared bands, and unique vibration frequency characteristics exhibited during specific maneuvers. All these feature data are encoded as feature vectors and stored in the database. In subsequent model optimization cycles, the system fine-tunes the target recognition network using the updated database (now containing "Tianying-2B" samples). Analysis revealed that a key feature distinguishing "Tianying-2B" from "Tianying-2A" is a relatively large area of ​​skin texture pattern in the mid-fuselage. During fine-tuning, the size of a convolutional kernel in an early convolutional layer of the network was slightly increased, allowing the network to better integrate contextual information from a larger area, thus more effectively capturing this new discriminative feature.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for accurate identification of dynamic aerial targets, characterized in that, Includes the following steps: Multidimensional spectral data of dynamic targets in the air are acquired using a multispectral imaging device, and the spatial distribution characteristics of the multidimensional spectral data are extracted. A three-dimensional motion trajectory model of the dynamic target is established based on the spatial distribution characteristics, and the instantaneous velocity vector of each trajectory point in the three-dimensional motion trajectory model is calculated. By combining the instantaneous velocity vector with preset baseline motion parameters, a set of motion deviation indicators for the dynamic target is generated; Cluster analysis is performed on the set of motion deviation indicators to classify the abnormal behavior categories of dynamic targets, and the keyframe sequences corresponding to the abnormal behavior categories are marked. Based on the keyframe sequence, a local motion feature map of the dynamic target is reconstructed, and morphological change parameters in the local motion feature map are extracted. The morphological change parameters are input into a pre-trained target recognition network, which outputs the type recognition result and confidence score of the dynamic target. When the confidence score is lower than a preset threshold, a supplementary identification process is activated to obtain high-resolution texture data of the dynamic target. By fusing the high-resolution texture data with the type recognition results, a final identification label for the dynamic target is generated; The feature database of the dynamic target is updated based on the final identification label, and the weight parameters of the target identification network are adjusted.

2. The method for accurate identification of dynamic aerial targets according to claim 1, characterized in that, The extraction of spatial distribution features from the multidimensional spectral data includes: The multidimensional spectral data is time-domain aligned to generate a standardized spectral sequence; Detect energy abrupt change regions in the normalized spectral sequence and calculate the spectral gradient distribution of the energy abrupt change regions; The outline boundary of the dynamic target is defined based on the spectral gradient distribution, and the geometric topology corresponding to the spatial distribution features is generated.

3. The method for accurate identification of dynamic aerial targets according to claim 2, characterized in that, The step of establishing a three-dimensional motion trajectory model of the dynamic target based on the spatial distribution characteristics includes: The centroid coordinate sequence of the dynamic target is determined based on the aforementioned geometric topology. The centroid coordinate sequence is smoothed using Kalman filtering to generate a continuous trajectory curve; Discrete trajectory points are sampled on the continuous trajectory curve, and the tangent direction angle of each discrete trajectory point is calculated.

4. The method for accurate identification of dynamic aerial targets according to claim 3, characterized in that, The calculation of the instantaneous velocity vector of each trajectory point in the three-dimensional motion trajectory model includes: Obtain the timestamp difference and spatial displacement of adjacent discrete trajectory points; The instantaneous velocity of the dynamic target is determined based on the ratio of the timestamp difference to the spatial displacement. The polar coordinate expression corresponding to the instantaneous velocity vector is generated by combining the tangent direction angle.

5. The method for accurate identification of dynamic aerial targets according to claim 1, characterized in that, The set of motion deviation indicators for generating dynamic targets includes: Extract typical velocity ranges and acceleration thresholds from preset baseline motion parameters; The degree of deviation between the instantaneous velocity vector and the typical velocity range is compared to generate a first deviation coefficient; Detect acceleration abrupt events of dynamic targets, count the frequency of acceleration abrupt events exceeding the acceleration threshold, and generate a second deviation coefficient; The first deviation coefficient and the second deviation coefficient are combined to form the set of motion deviation indicators.

6. The method for accurate identification of dynamic aerial targets according to claim 5, characterized in that, The classification of abnormal behaviors of dynamic targets includes: The spatial density analysis of the motion deviation index set was performed using a density clustering algorithm. Identify core data points in high-density areas and map the core data points to the key frame sequence; The abnormal behavior categories are classified into level labels based on the distribution density of core data points.

7. The method for accurate identification of dynamic aerial targets according to claim 1, characterized in that, The reconstruction of the local motion feature map of the dynamic target based on the keyframe sequence includes: Extract the region of interest (ROI) of the dynamic target from the keyframe sequence; Optical flow field calculations are performed on the region of interest to generate a motion vector distribution map; Based on the motion vector distribution map, the frequency domain feature components corresponding to the morphological change parameters are extracted.

8. The method for accurate identification of dynamic aerial targets according to claim 1, characterized in that, The activation and supplementary identification process includes: Control high-resolution imaging equipment to perform zoom tracking and shooting of dynamic targets; Super-resolution reconstruction is performed on the captured image sequence to generate the high-resolution texture data; Microstructural features in the high-resolution texture data are extracted using a convolutional neural network.

9. The method for accurate identification of dynamic aerial targets according to claim 1, characterized in that, The step of updating the feature database of dynamic targets based on the final identification label includes: The final identification label is then matched with historical identification records for similarity. When the matching degree is lower than the preset standard, a new target feature entry is created. The kernel size of the target recognition network is optimized based on the new target feature entries.

10. A precise aerial dynamic target identification system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for accurate identification of aerial dynamic targets as described in any one of claims 1 to 9.

Images