A UAV Target Recognition Method Based on Micro-Doppler Feature Separation

Abstract

This invention belongs to the field of UAV target recognition and discloses a UAV target recognition method based on micro-Doppler feature separation. This method is based on an improved mD-PF-TBD tracking framework and a simplified observation model. It combines a constant false alarm rate (CFAR) detection algorithm to adaptively initialize micro-Doppler harmonics and obtain target state and micro-Doppler prior information. By constructing a Gaussian frequency domain mask, weak micro-Doppler signals are separated from strong airframe main echoes. High-resolution, fine-grained time-frequency maps are then generated through time-frequency analysis, and a pre-trained deep convolutional neural network is used to complete UAV target classification and recognition. Simultaneously, this invention uses UAV hovering samples to construct a small sample training set, significantly reducing data acquisition and training costs. It also combines a target trajectory binary accumulation optimization strategy to correct single-frame recognition errors. Furthermore, it requires no target prior information, has low data requirements, and can adapt to complex clutter and variable flight conditions, effectively improving the accuracy and stability of UAV recognition.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) target recognition technology, specifically to a UAV target recognition method based on micro-Doppler feature separation. Background Technology

[0002] With the rapid iteration of drone technology and its widespread adoption in the civilian sector, issues such as the control of illegal intrusion by low-altitude, slow-moving, small targets have become increasingly prominent, posing a severe challenge to key areas such as public safety. Traditional active radar suffers from inherent defects such as poor electromagnetic concealment, high deployment costs, and susceptibility to electronic interference, making it difficult to meet the routine low-altitude detection needs in complex electromagnetic environments. External radiation source radar, relying on third-party civilian radiation sources for target detection, does not require active electromagnetic wave emission and possesses significant advantages such as strong concealment, flexible deployment, outstanding anti-interference capabilities, and wide coverage, making it one of the mainstream technologies in the current field of low-altitude drone detection.

[0003] During flight, the rotating rotors of rotary-wing UAVs periodically modulate the radar echo, forming a unique micro-Doppler feature. This feature contains inherent structural information about the target, such as rotor speed, blade length, and number of rotors. It is unaffected by similar radar cross sections of the targets and is the core basis for distinguishing different types of UAVs and achieving accurate target identification.

[0004] However, the extraction and recognition of micro-Doppler features for UAVs using external radiation source radar still faces many technical bottlenecks. On the one hand, the target echo energy received by external radiation source radar is inherently weak, and the amplitude of the micro-Doppler signal generated by rotor scattering is usually significantly lower than that of the main body echo. The weak micro-Doppler signal is completely submerged in the strong body echo and background clutter, resulting in a significant decrease in the stability and accuracy of feature extraction. On the other hand, most existing micro-Doppler-based tracking and feature extraction methods require prior information such as the number of target rotors and rotor size to construct an observation model. However, the rotors of multi-rotor UAVs have similar geometric dimensions and small differences in angular velocity, and their micro-Doppler harmonics highly overlap in the Doppler spectrum. It is impossible to accurately invert rotor parameters from the number of harmonics, which severely limits the generalization ability of existing methods in unknown target scenarios. On the other hand, deep learning-based drone recognition methods rely on large-scale labeled datasets, requiring the collection of echo data from drones in hovering, level flight, climbing, and turning motions. However, acquiring measured data from external radiation source radar requires building a dedicated experimental platform and coordinating airspace resources, and the data labeling process is time-consuming and labor-intensive, resulting in extremely high dataset construction costs. In scenarios with small sample data, deep neural networks are prone to overfitting. When trained using only easily collected hovering data, the model's single-frame classification accuracy is difficult to meet the requirements of engineering applications, and the reliability of the recognition results cannot be guaranteed. Summary of the Invention

[0005] To address the above technical problems, this invention provides a UAV target recognition method based on micro-Doppler feature separation, the method comprising the following steps: S1. Based on the improved mD-PF-TBD tracking framework, the observation model is simplified according to the micro-physical characteristics of rotary-wing UAVs. Combined with the constant false alarm rate detection algorithm, micro-Doppler harmonic adaptive initialization is realized to obtain the complete state information and prior information of micro-Doppler frequency of the target. S2. Based on the target's complete state information and micro-Doppler frequency prior information, a Gaussian frequency domain mask is constructed in the Doppler dimension. Through the weighted operation of the Gaussian frequency domain mask and the original Doppler spectrum, the target's micro-Doppler features and the organism's main Doppler signal are separated. S3. Perform inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to reconstruct a clean time-domain micro-Doppler signal. Then, perform narrow-window short-time Fourier transform time-frequency analysis on the time-domain signal to extract refined micro-Doppler time-frequency features and generate a high-resolution refined time-frequency map. S4. The refined time-frequency map is fed into a pre-trained deep convolutional neural network for forward inference to obtain the probability distribution of each category. The category corresponding to the maximum probability is selected as the final classification and recognition result of the UAV target.

[0006] As a further improvement of the present invention, the specific process of micro-Doppler harmonic adaptive initialization in step S1 is as follows: S11. At the initial stage of the target trajectory, the cell average constant false alarm rate (CA-CFAR) algorithm is used to perform initial detection of micro-Doppler harmonics in the Doppler spectrum to obtain candidate harmonic traces. The detection threshold calculation method of the CA-CFAR algorithm is as follows: a preset number of reference cells are selected in the left and right adjacent regions of the cell to be detected, the average amplitude of the reference cells on both sides is calculated and the average value is taken to obtain the clutter power estimate, and the detection threshold is calculated based on the clutter power estimate and the preset threshold factor. If the sample amplitude of the cell to be detected is greater than or equal to the detection threshold, it is determined that the cell has a candidate harmonic trace. S12. The candidate harmonic points are screened a second time using the angle tolerance model, and the points that are consistent with the echo angle of the target body and conform to the symmetrical distribution characteristics of micro-Doppler harmonics are retained. S13. Select the group with the best signal-to-noise ratio from the filtered points as the real micro-Doppler harmonics, and use its micro-Doppler frequency as the initialization parameter of the tracking model observation model.

[0007] As a further improvement of the present invention, the specific process of micro-Doppler feature separation in step S2 is as follows: S21. Based on the tracking results at the current moment, obtain the target's state estimate, which includes the range cell, the body Doppler frequency, and the micro-Doppler frequency; S22. Extract the Doppler spectrum at the target's range cell, cross-correlate the reference signal and the echo signal, add a Hamming window, and obtain the original Doppler spectrum through fast Fourier transform. S23. Based on the acquired micro-Doppler frequency, construct a Gaussian soft mask in the Doppler dimension, and weight the Gaussian soft mask with the original Doppler spectrum point by point to obtain the separated micro-Doppler frequency domain signal. S24. Perform inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to reconstruct the time-domain micro-Doppler signal.

[0008] As a further improvement of the present invention, the Gaussian soft mask is constructed around each order of micro-Doppler harmonics, retains the preset maximum order of micro-Doppler harmonics, and each order of harmonics is matched with a corresponding preset frequency domain bandwidth. The narrow-window short-time Fourier transform uses a preset fixed window length and accumulation time to perform time-frequency analysis.

[0009] As a further improvement of the present invention, step S3 specifically includes the following steps: S31. Perform an inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to convert the frequency domain signal into a pure time domain micro-Doppler signal; S32. Apply a Hamming window to the time-domain micro-Doppler signal and perform segment-by-segment time-frequency analysis using narrow-window short-time Fourier transform to obtain the frequency distribution matrix of the micro-Doppler signal as a function of time. S33. Convert the frequency distribution matrix into a grayscale or color fine time-frequency map, which highlights the periodic modulation characteristics of rotor rotation and characterizes the differences in rotor structure parameters of different UAVs.

[0010] As a further improvement of the present invention, step S4 specifically includes the following steps: S41. Construct a UAV time-frequency chart annotation dataset containing different clutter backgrounds and various motion states, divide it into training set and validation set according to a preset ratio, perform end-to-end training on the deep convolutional neural network, update parameters using a preset optimization algorithm, and save the optimal model weights through an early stopping mechanism. S42. After preprocessing the generated fine time-frequency map by size normalization and pixel value standardization, input it into the pre-trained deep convolutional neural network, perform multi-layer convolution, pooling and fully connected operations, and output the original prediction scores of various UAV targets. S43. The original prediction scores are normalized into a probability distribution using the softmax function, and the category with the highest probability value is selected as the single-frame classification result of the UAV target corresponding to the time-frequency map of that frame.

[0011] As a further improvement of the present invention, a small sample data training step is also included, which specifically includes the following process. Based on the physical characteristics of rotor-type UAVs in hovering state, such as constant rotor speed, regular micro-Doppler harmonic structure and the highest feature recognition, a small sample training strategy is determined to use only hovering state samples to construct the training set. Collect radar echo data of various UAVs under standard hovering conditions, process the data to generate corresponding fine time-frequency maps, and construct a small sample training set after completing sample labeling. Radar echo data of various UAVs in various motion states such as level flight, climb, dive and turn are collected, and fine time-frequency maps are generated after micro-Doppler feature separation to construct a generalization performance test set; Based on the small sample training set, the deep convolutional neural network is trained end-to-end. The model's ability to recognize cross motion states is verified based on the generalization performance test set, thus completing the model training and optimization under small sample conditions.

[0012] As a further improvement of the present invention, a binary accumulation optimization step based on target trajectory information is also included, which specifically includes the following process. A binomial distribution probability model is constructed based on the statistical accuracy of single-frame target classification; Based on the complete target trajectory information output in step S1, the classification results of multiple consecutive single frames corresponding to the same target trajectory are temporally correlated. Using a sliding window of preset length, the classification results of multiple frames after association are statistically analyzed frame by frame, and the cumulative occurrence of each category within the window is calculated. When the cumulative number of occurrences of a certain category within the sliding window reaches a preset decision threshold, the final identification result of that category as the drone target is output.

[0013] Based on this, the present invention also provides a UAV target recognition system based on micro-Doppler feature separation, the system being used to implement the above-described method, comprising: The adaptive initialization module is used to improve the mD-PF-TBD tracking framework. It relies on the general physical characteristics of micro-Doppler harmonics of rotary-wing UAVs to complete the normalization and simplification of the observation model. It combines CA-CFAR detection and angle consistency constraints to realize the adaptive initialization of micro-Doppler harmonics and outputs the target micro-Doppler frequency prior information and the target complete state information. The micro-Doppler feature separation module is used to extract the Doppler spectrum of the target's distance cell, construct a Gaussian frequency domain mask based on the target state information obtained from tracking, and separate the micro-Doppler signal from the main Doppler signal of the organism through weighted operations; The time-frequency analysis module is used to reconstruct the time-domain signal of the separated micro-Doppler frequency domain signal and perform narrow-window short-time Fourier transform processing to generate a high-resolution fine time-frequency map. The single-frame classification module is used to feed the fine time-frequency map into the pre-trained deep convolutional neural network for inference and output the single-frame classification results and corresponding probability distributions of various UAV targets. The few-sample optimization module is used to perform few-sample data training based on hovering data and binary accumulation processing based on target trajectory information, and output the final identification result of the UAV target.

[0014] Based on this, the present invention also provides a computer-readable storage medium storing a computer program that, when executed, implements the method described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: First, this invention improves the mD-PF-TBD tracking framework by simplifying the observation model based on the general physical characteristics of micro-Doppler harmonics in rotary-wing UAVs, eliminating the dependence on prior information such as the number of target rotors and rotor size. Combined with a constant false alarm rate (CFAR) detection algorithm, it achieves adaptive initialization of micro-Doppler harmonics, significantly improving the algorithm's generalization ability in unknown target scenarios. Simultaneously, based on the target state information obtained from tracking, a Gaussian frequency domain mask is constructed to effectively separate weak micro-Doppler signals from strong main body echoes, solving the problem of micro-Doppler features being submerged and difficult to extract stably. This enhances the structural feature differences between different UAV models and provides more discriminative feature inputs for subsequent target classification.

[0016] Secondly, this invention proposes a small-sample training strategy based on hovering data, which can complete the training of deep neural networks using only easily collected UAV hovering state samples, significantly reducing the construction cost and cycle of UAV datasets for external radiation source radars. The accompanying binary accumulation optimization strategy based on track information effectively corrects single-frame recognition errors by fusing multi-frame classification results of the same target track for statistical decision-making, significantly improving the accuracy and stability of UAV target recognition under small-sample conditions.

[0017] In summary, this invention achieves accurate and stable target identification for UAVs under an external radiation source radar system without relying on prior target information and significantly reducing training data requirements. It can adapt to complex clutter backgrounds and the changing motion conditions of UAVs, providing strong support for the large-scale engineering application of external radiation source radar in the field of low-altitude UAV management and control. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the micro-Doppler harmonic adaptive initialization process in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the micro-Doppler feature separation process in Embodiment 1 of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] First, some technical terms used in this invention will be explained: External radiation source radar, also known as passive radar, is a type of radar that does not actively emit electromagnetic waves. It relies on a third-party civilian radiation source as an illumination source and achieves target detection and tracking by receiving radiation source signals reflected from the target. It has the characteristics of strong electromagnetic concealment, low deployment cost, and outstanding anti-interference capability.

[0023] Micro-Doppler characteristics: When a target exhibits micro-motions such as rotation and vibration in addition to overall translation, its radar echo will generate periodic modulation sidebands on both sides of the main Doppler frequency. The frequency, amplitude, phase and other characteristics of these sideband signals are called micro-Doppler characteristics, which contain information about the target's structure and motion parameters.

[0024] An improved mD-PF-TBD tracking framework: Please refer to the applicant's published paper, titled "Micro-Doppler-Assisted Particle Filtering for Tracking of Doppler-Blind-Zone Targets with Passive Radar." This paper details the improved mD-PF-TBD tracking framework, which relies on the DTMB passive radar system. It integrates the micro-Doppler characteristics of the UAV rotor to construct an observation model and introduces a variable particle allocation strategy based on the Logistic function. When the target falls into the Doppler blind zone, the framework adaptively adjusts the particle tracking to include micro-Doppler harmonic components unaffected by clutter. Combined with measured angle constraints, it optimizes the target presence probability, effectively solving the problems of track fragmentation and degraded tracking accuracy in the Doppler blind zone of traditional algorithms.

[0025] Cell Average Constant False Alarm Rate Detection (CA-CFAR): By calculating the average power of reference cells around the cell to be detected to estimate the clutter level, the detection threshold is adaptively set to achieve target detection while maintaining a constant false alarm rate.

[0026] Gaussian frequency domain mask: Construct a weight matrix with a Gaussian function distribution centered on the frequency of the target signal, perform weighted processing on the original frequency domain signal, retain the signal components near the target frequency, and suppress interference and clutter at other frequencies.

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Example 1 As can be seen from the background technology, existing methods for identifying UAV targets using external radiation source radar suffer from difficulties in extracting micro-Doppler features, reliance on prior target information, and high training costs due to small sample sizes.

[0029] Please see Figures 1-3 This invention provides a method for UAV target recognition based on micro-Doppler feature separation, which includes the following steps: S1. Based on the improved mD-PF-TBD tracking framework, the observation model is simplified according to the micro-physical characteristics of rotary-wing UAVs. Combined with the constant false alarm rate detection algorithm, micro-Doppler harmonic adaptive initialization is realized to obtain the complete state information and prior information of micro-Doppler frequency of the target. First, assuming the target time delay, Doppler frequency, and rotor angular velocity are within a time interval... With the internal parameters unchanged, the system's measurement function model can be expressed as: In the formula Indicates the measurement noise of the system. This represents the spread function of the target echo. The receiver measurement result can be approximated by the Gaussian point target spread function, therefore representing the target body echo. and micro-Doppler harmonics It can be represented as: In the formula, and These represent the spread of the echo in the distance dimension and the Doppler dimension, respectively.

[0030] It should be noted that rotary-wing UAV targets contain at least one rotor structure, therefore their Doppler spectrum contains at least one set of micro-Doppler harmonics. Furthermore, for multi-rotor UAVs, the rotors typically have similar geometric dimensions, and the differences in rotor angular velocities are relatively small, meaning the frequencies of the micro-Doppler harmonics of the same target are close. Therefore, determining only one set of harmonics is sufficient for effective separation of micro-Doppler features.

[0031] Based on this physical characteristic, the number of rotors can be further included in the observation model. Fixed at 1, thus simplifying to: Accordingly, representing the target micro-Doppler harmonics Simplified to: However, in the absence of prior information about the target, the micro-Doppler frequency corresponding to the target to be tracked cannot be directly determined. To address this issue, a constant false alarm rate (CFAR) detection method is introduced at the initial stage of the target trajectory to detect micro-Doppler harmonics. Considering that the UAV rotor speed is relatively high and its micro-Doppler frequency is usually high, the probability of strong interference near this frequency is relatively low in low-altitude observation environments, and the background clutter can be approximated as uniformly distributed. Therefore, the cell-averaging-CFAR (CA-CFAR) algorithm from the mean-based CFAR method is selected to achieve initial harmonic detection.

[0032] The basic idea of ​​CA-CFAR is to select regions adjacent to the left and right sides of the unit to be detected. For each reference cell, the amplitude average of the left and right reference cells is calculated separately. The average of these two averages is then calculated to obtain the final clutter power estimate. The detection threshold based on the estimated value is In the formula This is the threshold factor. If the sample amplitude of the cell to be detected is greater than or equal to the threshold, the cell is determined to contain a target. To prevent target energy leakage to the reference cell and affecting clutter estimation, several protection cells are set on both sides of the cell to be detected to improve the accuracy of the detection process.

[0033] Meanwhile, due to the low SNR of micro-Doppler harmonics, the CFAR threshold should be lowered to achieve effective harmonic detection. However, the false alarm rate expression for CA-CFAR... It is evident that lowering the threshold inevitably leads to an increase in the false alarm rate of the system. Therefore, if only amplitude information is used for detection, it is difficult to distinguish between genuine harmonic traces and false alarm traces caused by clutter.

[0034] The decision criteria for real micro-Doppler harmonics have been given in the previous text. Therefore, the candidate point traces of CFAR output are screened a second time using the angle tolerance model.

[0035] Specifically, based on the angle consistency constraint, candidate points that are consistent with the echo angle of the target body and conform to the symmetrical distribution characteristics are retained. Then, the group with the highest SNR is selected as the true micro-Doppler harmonics, and its micro-Doppler frequency is obtained. As a measurement model The prior initialization parameters are used to achieve adaptive initialization of micro-Doppler harmonics during the initial stage of the trajectory. The specific process is as follows: Figure 2 As shown.

[0036] Furthermore, in step S1, the specific process of micro-Doppler harmonic adaptive initialization is as follows: S11. At the initial stage of the target trajectory, the cell average constant false alarm rate (CA-CFAR) algorithm is used to perform initial detection of micro-Doppler harmonics in the Doppler spectrum to obtain candidate harmonic traces. The detection threshold calculation method of the CA-CFAR algorithm is as follows: Select a preset number of reference cells in the left and right adjacent regions of the cell to be detected, calculate the average amplitude of the reference cells on both sides and take the average value to obtain the clutter power estimate, and calculate the detection threshold based on the clutter power estimate and the preset threshold factor. If the sample amplitude of the cell to be detected is greater than or equal to the detection threshold, it is determined that the cell has a candidate harmonic trace. S12. The candidate harmonic points are screened a second time using the angle tolerance model, and the points that are consistent with the echo angle of the target body and conform to the symmetrical distribution characteristics of micro-Doppler harmonics are retained. S13. Select the group with the best signal-to-noise ratio from the filtered points as the real micro-Doppler harmonics, and use its micro-Doppler frequency as the initialization parameter of the tracking model observation model.

[0037] S2. Based on the target's complete state information and prior information on micro-Doppler frequencies, a Gaussian frequency domain mask is constructed in the Doppler dimension. Through weighted operations of the Gaussian frequency domain mask and the original Doppler spectrum, the target's micro-Doppler features and the organism's main Doppler signal are separated. Traditional micro-Doppler separation algorithms typically start from the frequency variation characteristics of different signal components, separating the slowly varying main Doppler component from the rapidly varying micro-Doppler component. For example, they may use short-time fractional Fourier transform to perform sparse signal representation, and then separate and extract the micro-Doppler signal in the fractional domain. Such methods have good universality in the absence of prior information, but their high computational complexity limits their application in real-time engineering. In contrast, the micro-Doppler feature separation method proposed in this invention is guided by tracking results, does not require complex optimization, and is suitable for practical engineering applications. Based on the prior information of micro-Doppler frequencies obtained from the tracking results, a Gaussian frequency domain mask is constructed in the Doppler dimension, the expression of which is: In the formula, To retain the maximum micro-Doppler harmonic order, It is used to control the frequency domain bandwidth of each harmonic. The mask assigns higher weights to the vicinity of the micro-Doppler harmonics and smoothly attenuates other components, thereby effectively suppressing the main Doppler component of the body and other interferences, and retaining only the micro-Doppler characteristics of the target.

[0038] Please see Figure 3 In step S2, the specific process of micro-Doppler feature separation is as follows: S21. Based on the tracking results at the current moment, obtain the target's state estimate, which includes the range cell, the body Doppler frequency, and the micro-Doppler frequency; Specifically, according to The tracking results at each moment are used to obtain the target's state estimate, including the range cell. Doppler frequency With microDoppler frequency .

[0039] S22. Extract the Doppler spectrum at the target's range cell, cross-correlate the reference signal and the echo signal, add a Hamming window, and obtain the original Doppler spectrum through fast Fourier transform. Specifically, extract the distance unit of the target location. The Doppler spectrum at that point was obtained. In the formula, the reference signal is The echo signal is , This is the Hamming window. The Doppler frequency is calculated using the Fast Fourier Transform, where the Hamming window suppresses sidelobe leakage; its expression is: S23. Based on the obtained micro-Doppler frequency, construct a Gaussian soft mask in the Doppler dimension, and weight the Gaussian soft mask and the original Doppler spectrum point by point to obtain the separated micro-Doppler frequency domain signal. Based on microDoppler frequency A Gaussian soft mask is constructed using Doppler imaging, and weighted processing is applied to obtain the frequency domain expression after micro-Doppler feature separation. S24. Perform inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to reconstruct the time-domain micro-Doppler signal.

[0040] More preferably, the Gaussian soft mask is constructed around each order of micro-Doppler harmonics, retaining the preset maximum order of micro-Doppler harmonics, and each order of harmonics is matched with the corresponding preset frequency domain bandwidth; Narrow-window short-time Fourier transform uses a preset fixed window length and accumulation time to perform time-frequency analysis.

[0041] S3. Perform inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to reconstruct a clean time-domain micro-Doppler signal. Then, perform narrow-window short-time Fourier transform time-frequency analysis on the time-domain signal to extract refined micro-Doppler time-frequency features and generate a high-resolution refined time-frequency map. More specifically, step S3 includes the following steps: S31. Perform an inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to convert the frequency domain signal into a pure time domain micro-Doppler signal. ; S32. Apply a Hamming window to the time-domain micro-Doppler signal and perform segment-by-segment time-frequency analysis using narrow-window short-time Fourier transform to obtain the frequency distribution matrix of the micro-Doppler signal as a function of time. S33. Convert the frequency distribution matrix into a grayscale or color fine time-frequency map. The fine time-frequency map highlights the periodic modulation characteristics of the rotor rotation and characterizes the differences in rotor structure parameters of different UAVs.

[0042] S4. Feed the detailed time-frequency map into the pre-trained deep convolutional neural network for forward inference to obtain the probability distribution of each category. Select the category corresponding to the maximum probability as the final classification and recognition result of the UAV target.

[0043] Specifically, step S4 includes the following steps: S41. Construct a UAV time-frequency chart annotation dataset containing different clutter backgrounds and various motion states, divide it into training set and validation set according to a preset ratio, perform end-to-end training on the deep convolutional neural network, update parameters using a preset optimization algorithm, and save the optimal model weights through an early stopping mechanism. S42. After preprocessing the generated fine time-frequency map by size normalization and pixel value standardization, input it into the pre-trained deep convolutional neural network, perform multi-layer convolution, pooling and fully connected operations, and output the original prediction scores of various UAV targets. S43. The original prediction scores are normalized into a probability distribution using the softmax function, and the category with the highest probability value is selected as the single-frame classification result of the UAV target corresponding to the time-frequency map of that frame.

[0044] To verify the performance improvement of the micro-Doppler feature separation algorithm described in this invention on UAV target recognition, several mainstream deep convolutional neural network models were selected for comparative experiments, including AlexNet, VGG-16, ResNet-18, ResNet-50, and DenseNet. AlexNet consists of multiple convolutional layers and fully connected layers, using ReLU as the activation function and introducing a random deactivation mechanism to alleviate overfitting. VGG-16 uses a continuously stacked 3×3 convolutional kernel structure, resulting in a regular network structure. ResNet-18 and ResNet-50 are based on residual learning, using skip connections to address gradient degradation in deep network training; ResNet-50 further employs a bottleneck structure to enhance feature extraction capabilities. DenseNet achieves feature reuse and efficient gradient propagation through a dense connection mechanism, exhibiting good learning capabilities in scenarios with small sample data.

[0045] All network models were trained and tested using a general-purpose deep learning platform. Parameter optimization was performed using a stochastic gradient descent algorithm with a preset learning rate, and a reasonable number of training epochs was set to ensure model convergence. The experimental dataset contains time-frequency image samples of UAVs under various clutter backgrounds, including constant speed, acceleration, turning, climbing, and hovering motions. These samples were divided into training and testing sets according to a preset ratio.

[0046] Furthermore, considering that the micro-Doppler features of UAV targets change significantly with variations in flight speed, attitude, and other motion states, and that the feature distribution differs considerably across different motion states, deep learning-based recognition methods require the establishment of large-scale datasets covering all motion states. However, acquiring and manually annotating measured data from external radiation source radar is costly, and directly collecting UAV data across all motion states is difficult. Therefore, it is necessary to research practical and effective simplified training strategies under limited sample conditions.

[0047] The time-frequency map after micro-Doppler feature separation shows some similarity to the target's feature distribution in DBZ. Under conditions without strong clutter interference, the results of separating micro-Doppler features are similar to those after suppressing the aircraft echo. A typical special case of Doppler blind zone scenarios is when the UAV is hovering. Such data samples are easy to collect in real-world environments, as the UAV's rotor speed is relatively stable, and its micro-Doppler harmonic structure is clear and representative. Based on the above analysis, a simplified training strategy is proposed while employing the micro-Doppler separation algorithm: establishing a training set using only hovering UAV data samples. This method does not require large-scale labeled data, potentially significantly reducing training data requirements while maintaining high recognition accuracy.

[0048] The method in the preferred embodiment of the present invention further includes a small sample data training step, which specifically includes the following steps: Based on the physical characteristics of rotor-type UAVs in hovering state, such as constant rotor speed, regular micro-Doppler harmonic structure and the highest feature recognition, a small sample training strategy is determined to use only hovering state samples to construct the training set. Collect radar echo data of various UAVs under standard hovering conditions, process the data to generate corresponding fine time-frequency maps, and construct a small sample training set after completing sample labeling. Radar echo data of various UAVs in various motion states such as level flight, climb, dive and turn are collected, and fine time-frequency maps are generated after micro-Doppler feature separation to construct a generalization performance test set; Based on a small sample training set, the deep convolutional neural network is trained end-to-end. The model's ability to recognize cross motion states is verified based on the generalization performance test set, thus completing the model training and optimization under small sample conditions.

[0049] In the above steps, in order to address the problem that deep neural networks are prone to overfitting and insufficient generalization ability in small sample data scenarios, we prioritize deep convolutional neural networks (such as AlexNet, DenseNet, etc.) that have good adaptability to small sample data as classification models. Among them, DenseNet achieves feature reuse and efficient gradient transfer through dense connection mechanism, and shows better learning performance under small sample conditions.

[0050] Based on the constructed small sample training set, the network is trained end-to-end using a stochastic gradient descent algorithm with a preset learning rate. A reasonable number of training rounds is set to balance the model's convergence speed and the risk of overfitting. After training, the model is fully validated using a generalization performance test set containing multiple motion states to evaluate its ability to recognize cross motion states under different flight conditions. The model training and optimization under small sample conditions are completed by iteratively adjusting the network parameters.

[0051] Furthermore, considering that deep neural networks have limitations in learning complex time-frequency domain micro-motion features when training only with hovering state samples, and that single-frame classification results are easily affected by clutter interference and feature fluctuations, resulting in random errors; in practical engineering applications, target recognition tasks are usually based on continuous tracks formed by stable tracking, and the classification results of the same target in consecutive frames have natural temporal continuity, which can effectively correct single-frame recognition deviations by fusing multi-frame temporal information, thereby improving the reliability and stability of the final recognition results.

[0052] Specifically, for object detection problems, the binary accumulation method can often be used to improve the reliability of the decision results, that is, in In this decision The target is considered to exist only when the detection conditions are met. This idea can also be extended to target recognition tasks, by statistically accumulating the classification results of multiple frames within the same track to improve the accuracy of the final recognition result.

[0053] Let the probability of accurate target classification in a single frame be... ,but In this experiment The probability of accumulating successes for In the formula In target tracking, track confirmation typically employs a "3 / 5" detection method, meaning a track can be formed if 3 out of 5 frames contain valid detection results. A similar "3 / 5" voting strategy is used in target recognition, where 3 out of 5 adjacent classification results of the track are considered the final recognition result. Substituting this into the formula, we can obtain the accumulated recognition accuracy. The accuracy rate is approximately 94.21%, indicating that multi-frame information fusion based on flight tracks can significantly improve the recognition accuracy, thereby meeting the needs of practical engineering applications.

[0054] Therefore, the method in the preferred embodiment of the present invention further includes a binary accumulation optimization step based on target trajectory information, which specifically includes the following process. A binomial distribution probability model is constructed based on the statistical accuracy of single-frame target classification; Based on the complete target trajectory information output in step S1, the classification results of multiple consecutive single frames corresponding to the same target trajectory are temporally correlated. Using a sliding window of preset length, the classification results of multiple frames after association are statistically analyzed frame by frame, and the cumulative occurrence of each category within the window is calculated. When the cumulative number of occurrences of a certain category within the sliding window reaches a preset decision threshold, the final identification result of that category as the drone target is output.

[0055] In the above steps, by using a statistical decision mechanism based on a binomial probability model, the temporal continuity characteristics of the target trajectory are fully utilized, and the complementary information of classification results from multiple consecutive frames is integrated. This effectively corrects the random errors and misjudgments in single-frame classification, significantly improves the accuracy and robustness of UAV target recognition under small sample conditions, and enables the recognition results to meet the reliability requirements of practical engineering applications.

[0056] First, this invention improves the mD-PF-TBD tracking framework by simplifying the observation model based on the general physical characteristics of micro-Doppler harmonics in rotary-wing UAVs, eliminating the dependence on prior information such as the number of target rotors and rotor size. Combined with a constant false alarm rate (CFAR) detection algorithm, it achieves adaptive initialization of micro-Doppler harmonics, significantly improving the algorithm's generalization ability in unknown target scenarios. Simultaneously, based on the target state information obtained from tracking, a Gaussian frequency domain mask is constructed to effectively separate weak micro-Doppler signals from strong main body echoes, solving the problem of micro-Doppler features being submerged and difficult to extract stably. This enhances the structural feature differences between different UAV models and provides more discriminative feature inputs for subsequent target classification.

[0057] Secondly, this invention proposes a small-sample training strategy based on hovering data, which can complete the training of deep neural networks using only easily collected UAV hovering state samples, significantly reducing the construction cost and cycle of UAV datasets for external radiation source radars. The accompanying binary accumulation optimization strategy based on track information effectively corrects single-frame recognition errors by fusing multi-frame classification results of the same target track for statistical decision-making, significantly improving the accuracy and stability of UAV target recognition under small-sample conditions.

[0058] In summary, this invention achieves accurate and stable target identification for UAVs under an external radiation source radar system without relying on prior target information and significantly reducing training data requirements. It can adapt to complex clutter backgrounds and the changing motion conditions of UAVs, providing strong support for the large-scale engineering application of external radiation source radar in the field of low-altitude UAV management and control.

[0059] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0060] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, it should be understood that the sequence number of each step in the above embodiments does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention. The actions or steps recorded in the claims can be performed in a different order than that in the above embodiments and can still achieve the desired result. In addition, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0061] Example 2 Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides a UAV target recognition system based on micro-Doppler feature separation. The system includes an adaptive initialization module for improving the mD-PF-TBD tracking framework, completing the normalization and simplification of the observation model based on the general physical characteristics of micro-Doppler harmonics of rotary-wing UAVs, and realizing adaptive initialization of micro-Doppler harmonics by combining CA-CFAR detection and angle consistency constraints, and outputting the target micro-Doppler frequency prior information and the target complete state information. The micro-Doppler feature separation module is used to extract the Doppler spectrum of the target's distance cell, construct a Gaussian frequency domain mask based on the target state information obtained from tracking, and separate the micro-Doppler signal from the main Doppler signal of the organism through weighted operations; The time-frequency analysis module is used to reconstruct the time-domain signal of the separated micro-Doppler frequency domain signal and perform narrow-window short-time Fourier transform processing to generate a high-resolution fine time-frequency map. The single-frame classification module is used to feed the fine time-frequency map into the pre-trained deep convolutional neural network for inference and output the single-frame classification results and corresponding probability distributions of various UAV targets. The few-sample optimization module is used to perform few-sample data training based on hovering data and binary accumulation processing based on target trajectory information, and output the final identification result of the UAV target.

[0062] The system described in the above embodiments is used to implement the corresponding UAV target recognition method based on micro-Doppler feature separation in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0063] It should be noted that the aforementioned UAV target recognition system based on micro-Doppler feature separation is embodied in the form of functional units. The term "module" here can be implemented in software and / or hardware, without specific limitations.

[0064] For example, a "module" can be a software program, hardware circuitry, or a combination of both that implements the above-described functions. Hardware circuitry may include application-specific integrated circuits, electronic circuitry, a processor (e.g., a shared processor, a proprietary processor, or a group processor) and memory for executing one or more software or firmware programs, integrated logic circuitry, and / or other suitable components that support the described functions.

[0065] Example 3 Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the UAV target recognition method based on micro-Doppler feature separation as described in any of the above embodiments.

[0066] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0067] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the UAV target recognition method based on micro-Doppler feature separation as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0068] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0069] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuitry) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0070] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0071] Therefore, the units of the various examples described in the embodiments of this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0072] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for UAV target recognition based on micro-Doppler feature separation, characterized in that, The method includes the following steps: S1. Based on the improved mD-PF-TBD tracking framework, the observation model is simplified according to the micro-physical characteristics of rotary-wing UAVs. Combined with the constant false alarm rate detection algorithm, micro-Doppler harmonic adaptive initialization is realized to obtain the complete state information and prior information of micro-Doppler frequency of the target. S2. Based on the target's complete state information and micro-Doppler frequency prior information, a Gaussian frequency domain mask is constructed in the Doppler dimension. Through the weighted operation of the Gaussian frequency domain mask and the original Doppler spectrum, the target's micro-Doppler features and the organism's main Doppler signal are separated. S3. Perform inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to reconstruct a clean time-domain micro-Doppler signal. Then, perform narrow-window short-time Fourier transform time-frequency analysis on the time-domain signal to extract refined micro-Doppler time-frequency features and generate a high-resolution refined time-frequency map. S4. The refined time-frequency map is fed into a pre-trained deep convolutional neural network for forward inference to obtain the probability distribution of each category. The category corresponding to the maximum probability is selected as the final classification and recognition result of the UAV target.

2. The method according to claim 1, characterized in that, In step S1, the specific process of micro-Doppler harmonic adaptive initialization is as follows: S11. At the initial stage of the target trajectory, the unit average constant false alarm rate algorithm is used to perform initial detection of micro-Doppler harmonics in the Doppler spectrum to obtain candidate harmonic points. The detection threshold calculation method of the unit average constant false alarm rate algorithm is as follows: select a preset number of reference units in the left and right adjacent areas of the unit to be detected, calculate the average amplitude of the reference units on both sides and take the average value to obtain the clutter power estimate, calculate the detection threshold based on the clutter power estimate and the preset threshold factor, and if the sample amplitude of the unit to be detected is greater than or equal to the detection threshold, it is determined that the unit has a candidate harmonic trace. S12. The candidate harmonic points are screened a second time using the angle tolerance model, and the points that are consistent with the echo angle of the target body and conform to the symmetrical distribution characteristics of micro-Doppler harmonics are retained. S13. Select the group with the best signal-to-noise ratio from the filtered points as the real micro-Doppler harmonics, and use its micro-Doppler frequency as the initialization parameter of the tracking model observation model.

3. The method according to claim 1, characterized in that, In step S2, the specific process of micro-Doppler feature separation is as follows: S21. Based on the tracking results at the current moment, obtain the target's state estimate, which includes the range cell, the body Doppler frequency, and the micro-Doppler frequency; S22. Extract the Doppler spectrum at the target's range cell, cross-correlate the reference signal and the echo signal, add a Hamming window, and obtain the original Doppler spectrum through fast Fourier transform. S23. Based on the obtained micro-Doppler frequency, construct a Gaussian soft mask in the Doppler dimension, and weight the Gaussian soft mask and the original Doppler spectrum point by point to obtain the separated micro-Doppler frequency domain signal. S24. Perform inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to reconstruct the time-domain micro-Doppler signal.

4. The method according to claim 3, characterized in that, The Gaussian soft mask is constructed around each order of micro-Doppler harmonics, retaining the preset maximum order of micro-Doppler harmonics, and each order of harmonics is matched with the corresponding preset frequency domain bandwidth. Narrow-window short-time Fourier transform uses a preset fixed window length and accumulation time to perform time-frequency analysis.

5. The method according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Perform an inverse fast Fourier transform on the separated micro-Doppler frequency domain signal to convert the frequency domain signal into a pure time domain micro-Doppler signal; S32. Apply a Hamming window to the time-domain micro-Doppler signal and perform segment-by-segment time-frequency analysis using narrow-window short-time Fourier transform to obtain the frequency distribution matrix of the micro-Doppler signal as a function of time. S33. Convert the frequency distribution matrix into a grayscale or color fine time-frequency map. The fine time-frequency map highlights the periodic modulation characteristics of the rotor rotation and characterizes the differences in rotor structure parameters of different UAVs.

6. The method according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Construct a UAV time-frequency chart annotation dataset containing different clutter backgrounds and various motion states, divide it into training set and validation set according to a preset ratio, perform end-to-end training on the deep convolutional neural network, update parameters using a preset optimization algorithm, and save the optimal model weights through an early stopping mechanism. S42. After preprocessing the generated fine time-frequency map by size normalization and pixel value standardization, input it into the pre-trained deep convolutional neural network, perform multi-layer convolution, pooling and fully connected operations, and output the original prediction scores of various UAV targets. S43. The original prediction scores are normalized into a probability distribution using the softmax function, and the category with the highest probability value is selected as the single-frame classification result of the UAV target corresponding to the time-frequency map of that frame.

7. The method according to claim 1, characterized in that, It also includes a small sample data training step, which specifically includes the following process: Based on the physical characteristics of rotor-type UAVs in hovering state, such as constant rotor speed, regular micro-Doppler harmonic structure and highest feature recognition, a small sample training strategy is determined to use only hovering state samples to construct the training set. Collect radar echo data of various UAVs under standard hovering conditions, process the data to generate corresponding fine time-frequency maps, and construct a small sample training set after completing sample labeling. Radar echo data of various UAVs in various motion states such as level flight, climb, dive and turn are collected, and fine time-frequency maps are generated after micro-Doppler feature separation to construct a generalization performance test set; Based on the small sample training set, the deep convolutional neural network is trained end-to-end. The model's ability to recognize cross motion states is verified based on the generalization performance test set, thus completing the model training and optimization under small sample conditions.

8. The method according to claim 1, characterized in that, It also includes a binary accumulation optimization step based on target trajectory information, which specifically includes the following process. A binomial distribution probability model is constructed based on the statistical accuracy of single-frame target classification; Based on the complete target trajectory information output in step S1, the classification results of multiple consecutive single frames corresponding to the same target trajectory are temporally correlated. Using a sliding window of preset length, the classification results of multiple frames after association are statistically analyzed frame by frame, and the cumulative occurrence of each category within the window is calculated. When the cumulative number of occurrences of a certain category within the sliding window reaches a preset decision threshold, the final identification result of that category as the drone target is output.

9. A UAV target recognition system based on microDoppler feature separation, the system being used to implement the method according to any one of claims 1-8, characterized in that, It includes The adaptive initialization module is used to improve the mD-PF-TBD tracking framework. It relies on the general physical characteristics of micro-Doppler harmonics of rotary-wing UAVs to complete the normalization and simplification of the observation model. It combines CA-CFAR detection and angle consistency constraints to realize the adaptive initialization of micro-Doppler harmonics and outputs the target micro-Doppler frequency prior information and the target complete state information. The micro-Doppler feature separation module is used to extract the Doppler spectrum of the target's distance cell, construct a Gaussian frequency domain mask based on the target state information obtained from tracking, and separate the micro-Doppler signal from the main Doppler signal of the organism through weighted operations; The time-frequency analysis module is used to reconstruct the time-domain signal of the separated micro-Doppler frequency domain signal and perform narrow-window short-time Fourier transform processing to generate a high-resolution fine time-frequency map. The single-frame classification module is used to feed the fine time-frequency map into the pre-trained deep convolutional neural network for inference and output the single-frame classification results and corresponding probability distributions of various UAV targets. The few-sample optimization module is used to perform few-sample data training based on hovering data and binary accumulation processing based on target trajectory information, and output the final identification result of the UAV target.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method as described in any one of claims 1 to 8.

Images