An unmanned aerial vehicle model recognition method based on image processing

Abstract

The present application relates to the technical field of image processing, in particular to a kind of unmanned aerial vehicle model identification method based on image processing.The method comprises: obtaining the image containing unmanned aerial vehicle, and detecting the image obtains multiple candidate key points and corresponding confidence;Based on random sampling consistency algorithm, the matching structure key point graph with the highest matching score of the matching structure key point graph of the preset standard model is parsed from the candidate key point;The matching structure key point graph is rotated at a plurality of different rotation angles around its geometric center, the matching error between the key point graph after rotation at each rotation angle and the matching structure key point graph is calculated, and a group of feature vectors consisting of the matching error at multiple different rotation angles is obtained;The feature vector is input into pre-trained classifier, and the model of unmanned aerial vehicle is obtained.The scheme of the present application can accurately identify the specific model of unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a method for drone model identification based on image processing. Background Technology

[0002] In modern airspace management and security monitoring, automatically identifying drone models using image processing technology is crucial. Currently, deep learning-based methods, particularly convolutional neural networks (CNNs), have become the mainstream technology in this field. These methods, through end-to-end training on large drone image datasets, can automatically learn high-dimensional abstract features in images and use these features for classification, thereby achieving drone model identification. Compared to traditional manually designed feature methods, deep learning has achieved a significant improvement in recognition accuracy, demonstrating its powerful capabilities in complex image classification tasks.

[0003] Despite the success of deep learning-based methods, they suffer from a fundamental flaw: the models lack an explicit understanding of the drone's physical structure. Convolutional neural networks are essentially powerful function fitters that learn classification boundaries from the statistical distribution of massive pixel data. However, the features they learn are often abstract and non-intuitive, failing to establish clear correspondences with the drone's physical components (such as the fuselage, arms, and rotors). This "black box" learning mechanism makes the model extremely sensitive to variations not fully represented in the training data (especially changes in attitude and viewpoint). When a drone appears at an unusual angle, or when some of its key components are briefly obscured, the feature responses of the entire network can change drastically, leading to a sharp drop in recognition performance.

[0004] Therefore, it is particularly important to be able to identify drone models more accurately. Summary of the Invention

[0005] The purpose of this invention is to propose a drone model recognition method based on image processing to solve the problem that the prior art cannot accurately identify drone models; to this end, this invention provides a solution in one aspect.

[0006] This invention provides a method for drone model identification based on image processing, comprising:

[0007] Acquire images of the drone under test, and detect multiple candidate key points and their corresponding confidence levels in the images;

[0008] Based on the random sampling consensus algorithm, the matching structural key point map with the highest matching score to the preset standard model is parsed from the candidate key points; wherein, the matching score is positively correlated with the sum of the confidence scores of the key points contained in the structural key point map at each iteration, and negatively correlated with the sum of the distance errors between the coordinates of the key points and the coordinates of the corresponding projection points;

[0009] The key point map of the matching structure is rotated around its geometric center by multiple different rotation angles. The matching error between the rotated key point map and the matching key point map at each rotation angle is calculated to obtain a set of feature vectors composed of the matching errors at multiple different rotation angles.

[0010] The feature vector is input into a pre-trained classifier to obtain the drone model.

[0011] The aforementioned scheme detects key physical points such as the drone's fuselage and rotor, then uses a random sample consensus algorithm to parse the drone's geometry from candidate points containing noise and false positives. Finally, it creates feature vectors insensitive to in-plane rotation by analyzing the symmetry of this structure under rotation. This method bases the recognition process on observable and interpretable physical structures, rather than abstract pixel statistical features, thereby greatly enhancing robustness to different drone postures, partial occlusion, and changes in shooting angle, significantly improving the accuracy and reliability of recognition.

[0012] Optionally, the matching score is:

[0013] ;

[0014] in, The matching score at the k-th iteration. Let be the set of interior points at the k-th iteration. The i-th interior point in the set of interior points at the k-th iteration. Confidence level, The j-th interior point in the set of interior points at the k-th iteration. The normalized distance between its corresponding projection points These are preset weighting coefficients used to balance the importance of confidence level and geometric error.

[0015] The above scheme ensures that the final selected "matching structure key point map" is not only geometrically aligned with the standard model (small distance error), but also that the constituent points themselves are confirmed by the detection model with high confidence (high confidence), thus effectively avoiding geometric pseudo-matches caused by accidental low-confidence noise points.

[0016] Optionally, the process of obtaining the matching error is as follows:

[0017] Calculate the square of the distance between each rotated keypoint in the keypoint map at each rotation angle and the keypoint with the smallest distance in the matching structure keypoint map. Sum the squares of all distances as the matching error.

[0018] The above scheme enables the final generated feature vectors to more clearly and discriminatively reflect the unique symmetry characteristics of different drone models, thus facilitating more accurate identification by the subsequent classifier.

[0019] Optionally, the specific process of the parsing includes:

[0020] a. Randomly select a predetermined number of initial points from the candidate key points, and establish a one-to-one correspondence between each initial point and the center coordinates of each target component in the standard model;

[0021] b. Based on the relationship, obtain the transformation parameters for transforming the standard model to the image viewpoint; use the transformation parameters to project all key points of the standard model into the image to obtain the corresponding projection points; if there are candidate key points within the set range of each projection point, mark the candidate key point as an interior point;

[0022] c. Calculate the matching score of the keypoint graph formed by the set of interior points in the current iteration;

[0023] d. Repeat steps a to c until the preset number of iterations is reached, and construct the key point graph of the matching structure from the set of interior points with the highest matching scores.

[0024] The above scheme avoids exhaustive search of all possible point combinations by adopting the strategy of "random sampling-hypothesis construction-verification-iteration". The structure can be analyzed by a limited number of random sampling and verification. It can efficiently find the optimal structure interpretation with controllable computational cost in complex cases containing a large number of false positives and false negatives. This not only greatly improves the running efficiency of the algorithm, enabling it to achieve near real-time processing, but also enhances the tolerance for detection errors.

[0025] Optionally, the set range is an area centered on the projection point with a radius of 10% of the distance between the selected initial points.

[0026] The adaptive setting of the above-mentioned range significantly improves the algorithm's universality and matching robustness for UAV images at different distances and scales.

[0027] Optionally, the range of the multiple different rotation angles is [0, 2π].

[0028] Optionally, the image is detected to obtain multiple candidate key points, including:

[0029] The image is preprocessed, including size normalization, grayscale conversion, and Gaussian filtering;

[0030] A pre-trained keypoint detection model is used to analyze the pre-processed image and output the candidate keypoints in the image and the confidence score of each candidate keypoint. The candidate keypoints include the position coordinates of the UAV fuselage center and rotor center.

[0031] Optionally, the key point detection model is the YOLOv-Key points model.

[0032] Optionally, the pre-trained classifier is a support vector machine or a multilayer perceptron.

[0033] Optionally, the training process of the classifier includes:

[0034] Obtain standard models of various known drone types;

[0035] Generate corresponding structural symmetry feature vectors for each standard model and pair them with model labels to form a training set;

[0036] The classifier is trained using the training set.

[0037] The beneficial effects of this invention are as follows:

[0038] Compared to existing deep learning methods, the present invention first locates key physical parts of the UAV from the image, such as the fuselage center and the centers of each rotor. Second, using a robust matching algorithm (based on RANSAC), it efficiently filters the combination of points that best matches the standard structure of a certain type of UAV from candidate points containing false positives and noise, thereby reconstructing the matching structure key point map of the UAV. Finally, by analyzing the self-matching degree of this matching structure key point map at different rotation angles, a feature vector reflecting its geometric symmetry is generated for the identification of the UAV model under test. The present invention's solution is based on observable key components and their spatial relationships, unlike deep learning which relies on difficult-to-interpret internal features, making the identification process more transparent and traceable. Attached Figure Description

[0039] Figure 1 The flowchart illustrating the steps of an image processing-based drone model recognition method in this embodiment is shown in the illustration. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0041] This invention provides an image processing-based method for drone model recognition, which aims to solve the problem of low model recognition accuracy caused by the variable attitude and shooting angle of drones in the prior art. Specifically, by constructing a structural symmetry feature that is robust to rotation and combining it with a high-precision key point matching strategy, accurate identification of drone models is achieved.

[0042] Specifically, such as Figure 1 As shown in this embodiment, a drone model recognition method based on image processing includes the following steps:

[0043] Step S1: Image acquisition and preprocessing.

[0044] In this embodiment, an image containing the drone to be identified is first acquired using an image acquisition device (e.g., a fixed camera or a handheld device).

[0045] To eliminate the differences in images acquired from different devices and environments, and to provide high-quality input for subsequent keypoint detection, preprocessing of the original images is necessary.

[0046] Specifically, the preprocessing steps include:

[0047] 1. Size Normalization: The original image is uniformly scaled to a preset size. For example, to balance computational efficiency and detection accuracy, the preferred image size is... Pixels. If the size is too small, information about tiny key points may be lost, while if the size is too large, the computational burden of subsequent processing will be significantly increased.

[0048] 2. Grayscale processing: Convert the color image to a grayscale image to reduce the complexity of subsequent calculations and eliminate the interference of color information on structural feature extraction.

[0049] 3. Gaussian Filtering: Gaussian filtering is applied to the grayscale image to smooth the image and suppress noise. A small kernel size cannot effectively filter out noise, while an excessively large kernel size may blur the edges of key points, affecting positioning accuracy. In this embodiment, the kernel size for Gaussian filtering is preferably... Standard deviation Preferred .

[0050] Step S2: Detection of candidate key points.

[0051] The preprocessed image is analyzed to detect multiple candidate key points of the UAV and their corresponding confidence levels. These candidate key points are points that characterize the structural features of the UAV; in this embodiment, they specifically include the position coordinates of the UAV fuselage center, the centers of each rotor, etc.

[0052] In one embodiment, a pre-trained keypoint detection model is used to obtain candidate keypoints in an image of the drone to be identified.

[0053] The keypoint detection models mentioned above are either the YOLOv-Keypoints model or the EfficientDet model.

[0054] Taking the YOLOv-Keypoints model as an example, this model is an extension of the YOLO (You Only Look Once) object detection framework. It can not only output the bounding box of the object, but also predict the location coordinates and confidence scores of predefined key points.

[0055] Specifically, the preprocessed image is input into the YOLOv-Keypoints model, which extracts multi-scale features through a deep convolutional neural network backbone and outputs three types of information in parallel in the detection head: target category, target bounding box coordinates, and coordinates and confidence scores of each keypoint.

[0056] When training the YOLOv-Keypoints model, the initial learning rate can be set to... The training batch size can be set to [value missing]. Cosine annealing is used for attenuation. The total number of training epochs can be set to Rounds are used to ensure that the model converges fully.

[0057] Since the training process of the YOLOv-Keypoints model is based on existing technology, it will not be described in detail here.

[0058] Step S3: Obtain the key point map of the matching structure.

[0059] Since the detected candidate keypoints may include false positives (i.e., points in the background that are incorrectly identified as keypoints) or false negatives, directly using all keypoints for analysis would affect the robustness of the identification. Therefore, this embodiment uses the Random Sample Consensus (RANSAC) algorithm to parse the matching structure keypoint map with the highest matching score to the preset standard model from the candidate keypoints, so as to effectively eliminate the interference of outliers.

[0060] The preset standard model is a pre-built database containing two-dimensional key point coordinates of various known UAV models.

[0061] In this embodiment, the process of obtaining the matching structure key point map is as follows:

[0062] a. From the acquired set of candidate key points, randomly select a predetermined number of initial points and construct a one-to-one correspondence between each initial point and the coordinates of the corresponding target component in the preset standard model.

[0063] The target components are rotors, arms, fuselages, etc., in the standard model. For example, there can be two initial points, that is, the two initial points are matched with the center coordinates of the fuselage and the second rotor (numbering the four rotors).

[0064] b. Based on the aforementioned relationship, obtain the transformation parameters for transforming the standard model to the image viewpoint; use the transformation parameters to project all keypoints of the standard model into the image to obtain corresponding projection points; for each projection point, determine whether there are candidate keypoints within its defined neighborhood. If so, mark the candidate keypoint as an interior point.

[0065] In this embodiment, based on the one-to-one correspondence between the coordinates of the initial point and the center coordinates of the target component in the standard model, the transformation parameters, namely the scale, rotation and translation parameters required for the standard model to be transformed to the current image viewpoint, are calculated, i.e., the projection matrix.

[0066] The projection matrix described above describes the transformation relationship of transforming the 3D standard model to the current image viewpoint. The projection matrix can be obtained using the PnP (Perspective-n-Point) algorithm.

[0067] In this embodiment, the projection matrix described above can be used to project the three-dimensional coordinates of all key points in the standard model onto the image plane, resulting in a set of projection points.

[0068] The defined range is a circular area centered on the projection point and bounded by a defined radius. As a preferred embodiment, this defined radius can be adaptively set to a value equal to the average distance between the selected initial points. .

[0069] Of course, as in other implementations, when the drone occupies a large area in the image and the initial point spacing is large, the radius is increased accordingly to tolerate greater positioning errors; conversely, when the drone is far away, the radius is decreased to require more precise matching. Preferably, this scaling factor can be set to... between.

[0070] c. After obtaining the set of interior points for the current iteration, it is necessary to calculate the matching score of the key point graph formed by them.

[0071] To evaluate the quality of the interior point set in the k-th iteration, the following matching score calculation formula is constructed:

[0072] ;

[0073] in, For the first Matching score at the next iteration; For the first The set of interior points at the next iteration; , Each of these is a key point in the set of interior points; For the first In the next iteration, the internal point Confidence level; For the first In the next iteration, the internal point The normalized distance between its corresponding projection points; This is a preset weighting coefficient used to balance the importance of confidence level and geometric error.

[0074] In this embodiment, a preset weighting coefficient is used. Preferred .

[0075] In one embodiment, the normalized distance is the ratio of the Euclidean distance between each interior point and its corresponding projection point at each iteration to the maximum Euclidean distance at the corresponding iteration, wherein the maximum Euclidean distance is the maximum value among the Euclidean distances corresponding to all interior points at the corresponding iteration.

[0076] As can be seen from the above formula, the matching score Both the confidence level and geometric matching accuracy of the interior points were considered. When the set of interior points... When it contains more high-confidence key points, This will increase, thus improving the score. In other words, a hypothesis model supported by multiple high-confidence keypoints will score higher than a model supported by a large number of low-confidence keypoints, making the scoring results more reliable. Simultaneously, when these inliers are closer to their corresponding projected points, i.e., the distance error... Even smaller, It will decrease, and it will also increase the score. The introduction of this distance error ensures that not only a sufficient number of inliers are found, but also that the alignment of these inliers with the model is accurate enough. That is, a model with many inliers but loose matching will have its score reduced.

[0077] d. Repeat steps a to c above until the preset number of iterations is reached to obtain the matching structure key point map corresponding to the set of interior points with the highest matching score.

[0078] In one embodiment, the preset number of iterations is: Second-rate.

[0079] In this embodiment, the optimal solution is retained through multiple iterations. That is, after one iteration, the matching score calculated in the k-th iteration is used. Compared to the highest recorded global score Comparison. If Then update And save the current set of interior points and transformation parameters. This is the current optimal solution.

[0080] go through In the next iteration, the algorithm finally outputs the set of inliers with the highest scores. This set is the most reliable UAV geometry parsed from the noisy candidates. This process not only efficiently solves the combinatorial explosion problem, but also enhances robustness against false detections and occlusions of key points by re-estimating the scale factor and using all inliers for comprehensive scoring in each iteration.

[0081] This embodiment introduces the Random Sample Consensus (RANSAC) algorithm, which does not perform exhaustive search. Instead, it generates structural hypotheses by randomly sampling a small number of points multiple times and verifies the degree of conformity between these hypotheses and the overall candidate points. Compared with the existing technology that directly matches the structural templates (i.e., standard models) of various UAV models to determine which type of UAV appears in the image, this method can reduce the amount of computation. At the same time, for the candidate points that contain a large number of false detection points, the scoring method can be used to filter them, which can avoid misidentification and reduce noise interference.

[0082] Step S4: Extraction of feature vectors.

[0083] To construct a feature representation that is insensitive to in-plane rotation of the UAV, this embodiment proposes a feature vector. This feature vector is generated by analyzing the self-matching error variation law of the key point map of the matching structure during rotation.

[0084] Specifically, the process of obtaining the feature vector is as follows:

[0085] First, obtain the geometric center of the key point map of the matching structure.

[0086] The geometric center mentioned above can be the center point of the set of interior points in the matching structure keypoint diagram, or it can be the arithmetic mean of the coordinates of all interior points in the set of interior points corresponding to the matching structure keypoint diagram.

[0087] Secondly, the key point map of the matching structure is rotated around its geometric center by multiple different rotation angles to obtain the rotated key point map at each rotation angle.

[0088] In this embodiment, the rotation angle The range of values ​​is For example, it can be... The interval is divided into Divide into equal parts, for example Then the angle of each rotation is ,in .

[0089] For each rotation angle This yields a rotated keypoint map.

[0090] Then, the matching error between each rotated keypoint map and the matching structure keypoint map is calculated.

[0091] Specifically, the matching error is:

[0092] ;

[0093] in, The rotation angle is Matching error at that time; Key point diagram after rotation The j-th key point in To match the structural key point diagram The i-th key point in; Let represent the square of the Euclidean distance, and min() is the minimum value.

[0094] in, This represents the minimum sum of squares of the distances between each keypoint after rotation and the closest point in the original keypoint set. This ensures that even if the index of the keypoint changes (e.g., rotor 1 rotates to the position of rotor 2), the distance value will be very small as long as the shapes can overlap.

[0095] The above formula calculates the sum of squares of the Euclidean distances from each keypoint in the rotated keymap to its nearest keypoint in the original matching structure keymap. This calculation is performed when the rotation angle... It is a symmetrical angle, so that the key point diagram after rotation Key point diagram of the matching structure with the original The structures are highly overlapping (e.g., for a quadcopter drone, when rotating...). When the value is an integer multiple of the original key point (and their structures are basically overlapping), each key point in the rotated key point map can find a very close or even overlapping key point in the original matching key point map. At this point, the matching error is minimal. It will be very small, close to zero. Conversely, if the rotation angle... If it is not a symmetrical angle, the rotated keymap will be misaligned with the original matching structure keymap. Each keypoint will not be able to find a keypoint that overlaps with it, resulting in a large sum of squared distances and thus a large error value.

[0096] Finally, the matching errors under all rotation angles are combined sequentially to form a... 3D eigenvectors .

[0097] The aforementioned eigenvectors describe the changes in the key point structure of the UAV as it rotates, i.e., its structural symmetry, and therefore can also be called structural symmetry eigenvectors.

[0098] The reason for obtaining the structural symmetry feature vector is that if the original positions of key points in the image are used directly for recognition, the same drone might be mistaken for a different model simply by changing the angle or attitude. For example, the rotor positions in frontal and side views are completely different, which can easily lead to misjudgment. Therefore, in this embodiment, the absolute positions of these key points in the image are not the concern; instead, the relative relationships between them are the focus, especially the repetitive patterns exhibited by the entire structure during rotation (e.g., the same structure every 90 degrees for a quadcopter). Through this degree of matching under rotation, the representation will not change due to the drone's attitude and distance in the image, thus obtaining a feature that only reflects the structural characteristics of the drone itself and is unaffected by the shooting angle, making the recognition results more stable and reliable.

[0099] Step S5: Identification of the model of the drone to be tested.

[0100] In this embodiment, the structural symmetry feature vector of the UAV under test is input into a pre-trained classifier, and the classifier outputs the final UAV model.

[0101] The classifier can be a support vector machine (SVM) or a multilayer perceptron (MLP).

[0102] Taking SVM as an example, a radial basis function kernel can be used. When using SVM for classification, the penalty coefficient... Preferred Parameters of the radial basis function kernel Grid search is available The optimal hyperplane is found within the range to divide the feature vectors corresponding to different UAV models into different category spaces.

[0103] Taking SVM as an example, its training process is as follows:

[0104] First, obtain the training set.

[0105] The process of obtaining the training set is as follows:

[0106] Obtain standard models of various known UAV models and obtain the standard structural symmetry feature vectors of each standard model. Pair these standard structural symmetry feature vectors with their model labels to form a training set.

[0107] The method for obtaining the standard structural symmetry feature vector of the standard model mentioned above is the same as the method for obtaining the structural symmetry feature vector of the UAV to be identified, and will not be repeated here.

[0108] Secondly, train the SVM using the training set until the model converges.

[0109] In this embodiment, after obtaining the trained classifier, the structural symmetry feature vector of the UAV to be identified is input into the trained classifier, and the model of the UAV is output.

[0110] The solution of this invention extracts the symmetry information of the structure itself, and the generated features are not affected by the orientation or distance of the UAV in the image. It can effectively cope with the changes caused by various flight attitudes and shooting angles, and realize the accurate and robust identification of the specific model of the UAV from a single two-dimensional image, providing strong technical support for applications in related fields.

[0111] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0112] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. An unmanned aerial vehicle model identification method based on image processing, characterized in that, The method comprises the following steps: Obtain an image of a to-be-detected unmanned aerial vehicle, and detect the image to obtain a plurality of candidate key points and corresponding confidence levels; The candidate key points include position coordinates of the center of the unmanned aerial vehicle body and the center of the rotor; Based on a random sample consensus algorithm, a matching structure key point graph with the highest matching score with a preset standard model is parsed from the candidate key points, wherein the matching score is positively correlated with the sum of the confidence levels of the key points contained in the structure key point graph in each iteration, and is negatively correlated with the sum of distance errors between the coordinates of the key points and the coordinates of the corresponding projection points; the preset standard model is a database containing two-dimensional key point coordinates of a plurality of known models of unmanned aerial vehicles; The matching structure key point graph is rotated at a plurality of different rotation angles around its geometric center, the matching error between the rotated key point graph at each rotation angle and the matching structure key point graph is calculated, and a structure symmetry feature vector composed of a plurality of matching errors at different rotation angles is obtained; The feature vector is input into a pre-trained classifier to obtain the model of the unmanned aerial vehicle; The matching error is obtained by calculating the square of the distance between each rotated key point in the rotated key point graph at each rotation angle and the key point with the smallest distance in the matching structure key point graph, and taking the sum of all distance squares as the matching error; The parsing process comprises: a. randomly selecting a predetermined number of initial points from the candidate key points, and constructing a one-to-one correspondence between the initial points and the center coordinates of each target component in the standard model; b. based on the correspondence, obtaining transformation parameters for transforming the standard model to the image perspective; using the transformation parameters to project all key points of the standard model into the image to obtain corresponding projection points; if there is a candidate key point within the set range of each projection point, the candidate key point is marked as an internal point; c. calculating the matching score of the key point graph composed of the internal point set in the current iteration; d. repeating steps a to c until a preset number of iterations is reached, and the matching structure key point graph is composed of the internal point set with the highest matching score. 2.The unmanned aerial vehicle model identification method based on image processing of claim 1, wherein, The matching score is: ； in, The matching score at the k-th iteration. Let be the set of interior points at the k-th iteration. The i-th interior point in the set of interior points at the k-th iteration. Confidence level, The j-th interior point in the set of interior points at the k-th iteration. The normalized distance between its corresponding projection points These are preset weighting coefficients used to balance the importance of confidence level and geometric error. 3.The unmanned aerial vehicle model recognition method based on image processing of claim 1, wherein, The set range is a region with the projection point as the center and the distance between the selected initial points as the radius. 4.The unmanned aerial vehicle model recognition method based on image processing of claim 1, wherein, The plurality of different rotation angles are in the range of [0, 2π).

5. The unmanned aerial vehicle model identification method based on image processing according to claim 1, characterized in that, Detecting the image to obtain a plurality of candidate key points comprises: Preprocessing the image, which includes size normalization, grayscale and Gaussian filtering; Using a pre-trained key point detection model to analyze the preprocessed image to output the candidate key points in the image and the confidence levels of each candidate key point.

6. The unmanned aerial vehicle model identification method based on image processing according to claim 5, characterized in that, The key point detection model is a YOLOv-Keypoints model.

7. The unmanned aerial vehicle model identification method based on image processing according to claim 1, characterized in that, The pre-trained classifier is a support vector machine or a multilayer perceptron. 8.The method of claim 7, wherein, The training process of the classifier comprises: Obtain standard models of a plurality of known models of unmanned aerial vehicles; Generate corresponding structure symmetry feature vectors for each standard model, and pair them with model labels to form a training set; Train the classifier using the training set.

Images