Unmanned aerial vehicle detection method based on frequency modulated continuous wave radar and fusion features
By using a residual attention convolutional network and a feature fusion network based on frequency modulated continuous wave radar, the problems of UAV detection methods being insensitive to global structure and easily interfered with by local noise are solved. This enables deep-level correlation capture of deep local features and global structural features, thereby improving the accuracy of UAV detection.
Patent Information
- Application Number
- CN202511759803.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-02-13
AI Technical Summary
Existing UAV detection methods are insensitive to global structure, easily interfered with by local noise, and have difficulty capturing deep and complex relationships between features, resulting in poor detection performance.
A method based on frequency-modulated continuous wave radar is adopted. Deep local features of two-dimensional time-frequency maps are extracted through residual attention convolutional network, and global structural features are extracted using discrete Chebyshev moments. Deep fusion is performed by combining feature fusion network and the self-attention mechanism of Transformer architecture is used to adjust the feature fusion weights to achieve complementary advantages of deep local features and global structural features.
It significantly improves the accuracy of UAV detection, especially in complex environments with moving target interference, thus enhancing UAV detection performance.
Smart Images
Figure CN121522600A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of radar signal processing and deep learning, and particularly relates to a method for detecting unmanned aerial vehicles based on a frequency modulated continuous wave radar and fused features. BACKGROUND
[0002] Unmanned aerial vehicles (UAVs) have the characteristics of miniaturization, low cost, high maneuverability and convenient operation. In recent years, UAVs have made breakthroughs in endurance, autonomous flight and sensing performance, accelerating their application in multiple scenarios. However, UAV suicide attacks and "black flight" have also brought security risks to public safety and personal privacy. Therefore, it is of great significance to detect UAVs in safety-sensitive areas.
[0003] Due to the small size, low flight height, slow flight speed of UAVs and the existence of a large number of interferences in the environment, it is very difficult to reliably detect and identify UAVs. At present, the detection technology for UAVs mainly includes visual, acoustic and radar means. Among them, visual detection is easily affected by bad weather conditions such as rain and snow, and acoustic detection is limited in detection range and accuracy due to environmental noise interference. In contrast, radar detection technology stands out among various detection technologies due to its long range, high precision and all-weather working advantages. In particular, frequency modulated continuous wave (FMCW) radar is considered an ideal technology for detecting UAVs because of its high micro-motion resolution, no distance blind area and strong anti-interception capability. However, the robust detection of UAVs by FMCW radar still faces difficulties, i.e., the weak echo signal of UAVs and the low Doppler frequency make them easily submerged by background clutter, and the interference of other moving targets such as birds in the environment further increases the difficulty of detection and identification. Therefore, it is of great theoretical and practical significance to study an efficient and reliable FMCW radar detection method for UAVs.
[0004] The deep learning-based UAV detection method directly inputs the received FMCW radar echo signal or its time-frequency graph into a network model for training and detection. The advantages of this method are that it does not need to manually select features, set detection statistics and thresholds, and does not depend on model assumptions, but it needs a large amount of data for network model training, may not be sensitive to global structure, and is easily disturbed by local noise. Complementing the deep learning method with the precise expression ability of traditional mathematical feature descriptors on global and structural information can improve the performance of the UAV detection method. However, how to effectively and deeply fuse these features is a current problem. Existing methods achieve simple fusion through feature concatenation, which is difficult to capture deep complex correlations between features, thereby limiting the further improvement of UAV detection performance.
[0005] In summary, solving the problem that the existing unmanned aerial vehicle detection method is not sensitive to the global structure, is easily disturbed by local noise, and is difficult to capture the deep and complex correlation between features is the key to improve the unmanned aerial vehicle detection performance. SUMMARY
[0006] The purpose of the present application is to solve the problem of poor unmanned aerial vehicle detection performance due to the existing method being not sensitive to the global structure, being easily disturbed by local noise, and being difficult to capture the deep and complex correlation between features, and a unmanned aerial vehicle detection method based on frequency modulation continuous wave radar and fusion features is proposed.
[0007] The technical scheme adopted by the present application to solve the above technical problem is: a unmanned aerial vehicle detection method based on frequency modulation continuous wave radar and fusion features, the method specifically comprises the following steps:
[0008] Step one, using FMCW radar to emit Chirp signal to the area to be detected, then collecting the echo signal received by FMCW radar, mixing the transmitting signal of FMCW radar with the received echo signal to obtain beat frequency signal;
[0009] Step two, performing short-time Fourier transform on the beat frequency signal to obtain two-dimensional time-frequency graph;
[0010] Step three, taking the two-dimensional time-frequency graph as the input of the residual attention convolutional network, outputting the deep local features of the two-dimensional time-frequency graph through the residual attention convolutional network;
[0011] Step four, extracting the global structure features of the two-dimensional time-frequency graph by using discrete Chebyshev matrix;
[0012] Step five, taking the deep local features and the global structure features of the two-dimensional time-frequency graph as the input of the feature fusion network, and performing higher information density fusion on the deep local features and the global structure features through the feature fusion network to obtain the fusion features of the two-dimensional time-frequency graph;
[0013] Step six, taking the fusion features of the two-dimensional time-frequency graph as the input of the detector, and outputting the unmanned aerial vehicle detection result through the detector.
[0014] Further, the Chirp signal emitted by the FMCW radar is sawtooth wave, that is, the form of the transmitting signal is:
[0015]
[0016] wherein, represents the transmitting signal, is the amplitude of the transmitting signal, is the imaginary unit, is the starting frequency, is the frequency modulation slope, is a time, is a pulse repetition period, is a pulse width, is a natural number, .
[0017] Further, the form of the echo signal received by the FMCW radar is:
[0018]
[0019] wherein, is the echo signal received by the FMCW radar, is the radar reflection coefficient of the surface of the UAV, is the time delay of the echo signal of the UAV.
[0020] Further, the form of the beat signal is:
[0021]
[0022] wherein, is the beat signal, is the amplitude of the beat signal.
[0023] Further, the specific process of step two is:
[0024] Step two one, sliding the Hann window with a length of on the beat signal, the sliding step is , and obtaining the local signal corresponding to each position of the Hann window by sliding;
[0025] Step two two, multiplying the local signal corresponding to each position of the Hann window with the Hann window function respectively, and then respectively performing fast Fourier transform on each multiplication result to obtain the frequency domain amplitude of each frequency domain point in the fast Fourier transform result;
[0026] Step two three, using the frequency domain amplitudes of each frequency domain point to compose an initial two-dimensional time-frequency matrix, and then respectively performing logarithmic operation on each element in the initial two-dimensional time-frequency matrix, after global standardization of each element after logarithmic operation, using each standardized element to compose a final two-dimensional time-frequency matrix, that is, obtaining the two-dimensional time-frequency diagram.
[0027] Further, the working process of the residual attention convolutional network is:
[0028] Taking the two-dimensional time-frequency diagram as the input of the Stem Layer module, in the Stem Layer module, the two-dimensional time-frequency diagram is passed through the first convolutional layer;
[0029] The output of the first convolutional layer is taken as the input of the first set of normalization layers, and the output of the first set of normalization layers is taken as the input of the first ReLU activation function layer;
[0030] The output of the first ReLU activation function layer is taken as the input of the first regularization layer, and the output of the first regularization layer is taken as the output of the Stem Layer module.
[0031] The output of the Stem Layer module is taken as the input of the first residual channel attention mechanism module, and the first residual channel attention mechanism module includes a main path and a shortcut path, wherein:
[0032] In the main path, the output of the Stem Layer module is sequentially subjected to a second convolutional layer, a second set of normalization layers, a second ReLU activation function layer, a second regularization layer, a third convolutional layer, a third set of normalization layers, and a first attention module SE, and the output of the first attention module SE is taken as the output of the main path.
[0033] In the shortcut path, the output of the Stem Layer module is sequentially subjected to a fourth convolutional layer and a fourth set of normalization layers, and the output of the fourth set of normalization layers is taken as the output of the shortcut path.
[0034] The output of the main path and the output of the shortcut path are added, and the addition result is taken as the input of a fourth ReLU activation function layer, and the output of the fourth ReLU activation function layer is taken as the output of the first residual channel attention mechanism module.
[0035] The output of the first residual channel attention mechanism module is taken as the input of a second residual channel attention mechanism module, and the output of the second residual channel attention mechanism module is taken as the input of a third residual channel attention mechanism module.
[0036] The output of the third residual channel attention mechanism module is taken as the input of a feature aggregation module, and the feature aggregation module includes a first adaptive average pooling layer.
[0037] The output of the feature aggregation module is taken as the input of a feature head module, and in the feature head module, the output of the feature aggregation module is sequentially subjected to a flattening operation, a first layer normalization layer, a third regularization layer, a second fully connected layer, and a fifth ReLU activation function layer, and the output of the fifth ReLU activation function layer is taken as the output of the residual attention convolutional network.
[0038] Further, the first attention module SE includes a first global average pooling layer, a third ReLU activation function layer, a first fully connected layer, and a first sigmoid activation function layer in sequence.
[0039] Further, the specific process of step four is:
[0040] The value of the element at the i-th row and the j-th column in the two-dimensional time-frequency graph is denoted as The discrete Chebyshev matrix is obtained by using the discrete Chebyshev polynomials
[0041]
[0042] The absolute values of each element in the discrete Chebyshev matrix are calculated respectively, and the average value of the absolute values is calculated as a scaling factor; and each element value in the discrete Chebyshev matrix is divided by the scaling factor to obtain a new discrete Chebyshev matrix.
[0043] Each element value in the new discrete Chebyshev matrix is compressed to [-1, 1] by the tanh function to obtain a normalized discrete Chebyshev matrix, and finally the elements in the normalized discrete Chebyshev matrix are flattened into a one-dimensional feature vector by row, and the flattened one-dimensional feature vector is taken as the global structure feature.
[0044]
[0045] Further, the working process of the feature fusion network is as follows:
[0046] Step five one, taking the deep local feature as the input of the first linear projection layer, and outputting the deep local feature token through the first linear projection layer;
[0047] Step five two, taking the global structure feature as the input of the second linear projection layer, and outputting the global structure feature token through the second linear projection layer;
[0048] And introducing the classification token as an information aggregator;
[0049] Step five two, splicing the deep local feature token, the global structure feature token and the classification token in the sequence dimension to form a token sequence;
[0050] And introducing the position encoding matrix;
[0051] Step five three, element-wise addition is performed between the token sequence and the position encoding matrix, and an addition result is obtained;
[0052] Step five four, the addition result in step five three is taken as the input of the first multi-head self-attention layer, the output of the first multi-head self-attention layer is taken as the input of the second multi-head self-attention layer, the output of the second multi-head self-attention layer is taken as the input of the third multi-head self-attention layer;
[0053] The output of the third multi-head self-attention layer is further taken as the input of the second normalization layer, the output of the second normalization layer is taken as the input of the multi-layer perception, and the output of the multi-layer perception is taken as the fusion feature of the two-dimensional time-frequency graph.
[0054] Further, the working process of the detector is:
[0055] The fusion feature of the two-dimensional time-frequency graph is taken as the input of the third normalization layer, and the output of the third normalization layer is further subjected to nonlinear transformation by the hidden unit;
[0056] In the hidden unit, the output of the third normalization layer is sequentially subjected to the third linear projection layer, the sixth ReLU activation function layer and the fourth regularization layer, and the output of the fourth regularization layer is taken as the output of the hidden unit;
[0057] The output of the hidden unit is further taken as the input of the Softmax function, and the probability of detecting the unmanned aerial vehicle is output by the Softmax function;
[0058] If the probability of detecting the unmanned aerial vehicle is greater than 0.5, it indicates that the unmanned aerial vehicle is detected, otherwise, the unmanned aerial vehicle is not detected.
[0059] The beneficial effects of the present application are:
[0060] This invention utilizes FMCW radar to receive target echo signals, converts the received echo signals into a two-dimensional time-frequency map, and then uses a residual attention convolutional network to extract deep local features from the two-dimensional time-frequency map. Discrete Chebyshev moments are then used to extract global structural features from the two-dimensional time-frequency map, achieving an organic combination of automated feature learning capabilities and precise representation of global and structural information, thus realizing complementary advantages. Furthermore, a feature fusion method based on the Transformer architecture is used to deeply fuse the deep local features and global structural features of the two-dimensional time-frequency map. During the feature fusion process, a self-attention mechanism dynamically adjusts the fusion weights of deep local features and global structural features, making the feature fusion process more adaptive and the fused features more expressive. Compared to traditional methods, this invention's method is better able to capture the deep and complex correlations between deep local features and global structural features, significantly improving the detection accuracy of UAVs in complex moving target interference environments and enhancing UAV detection performance. Attached Figure Description
[0061] Figure 1 This is a flowchart of a UAV detection method based on frequency-modulated continuous wave radar and fused features according to the present invention;
[0062] Figure 2 A schematic diagram for acquiring echo signals received by FMCW radar;
[0063] Figure 3 This is a schematic diagram of the structure of a residual attention convolutional network;
[0064] Figure 4 This is a schematic diagram of the feature fusion network structure;
[0065] Figure 5 This is a schematic diagram of the detector's structure;
[0066] Figure 6 This is a graph showing the detection accuracy of the method of the present invention. Detailed Implementation
[0067] Specific implementation method one: Combining Figure 1 This embodiment describes a UAV detection method based on frequency-modulated continuous wave radar and fused features. The method specifically includes the following steps:
[0068] Step 1: Use FMCW radar to transmit Chirp signals to the area to be detected, then collect the echo signals received by FMCW radar, and mix the transmitted signals and received echo signals to obtain the beat frequency signal.
[0069] The chirp signal transmitted by the FMCW radar is a sawtooth wave, that is, the form of the transmitted signal is:
[0070]
[0071] wherein, represents a transmitted signal, is an amplitude of the transmitted signal, is an imaginary unit, is a starting frequency, is a frequency modulation slope, is time, is a pulse repetition period, is a pulse width, is a natural number, ;
[0072] The echo signal received by the FMCW radar is in the form of:
[0073]
[0074] wherein, is an echo signal received by the FMCW radar, is a radar reflection coefficient of a surface of the unmanned aerial vehicle, is a time delay of the echo signal of the unmanned aerial vehicle;
[0075] The beat signal is in the form of:
[0076]
[0077] wherein, is a beat signal, is an amplitude of the beat signal.
[0078] Step two, performing a short-time Fourier transform on the beat signal to obtain a two-dimensional time-frequency graph;
[0079] Step two one, sliding a Hanning window with a length of on the beat signal, and the sliding step is , and a local signal corresponding to each position of the Hanning window is obtained by sliding;
[0080] Step two two, multiplying the local signal corresponding to each position of the Hanning window with the Hanning window function respectively, and then performing a fast Fourier transform (FFT) on each multiplication result respectively to obtain a frequency domain amplitude of each frequency domain point in the fast Fourier transform result;
[0081] The specific process of the short-time Fourier transform is as follows:
[0082]
[0083] wherein, is a sliding window function, , is the window length (in the present application ), is the overlap degree (in the present application ), is the frequency, is a natural number.
[0084] The number of points of the fast Fourier transform is (in the present application , the value is 1024), so a group of frequency domain amplitudes can be obtained from the fast Fourier transform result corresponding to the local signal of each position, and each group of frequency domain amplitudes is taken as a column of a two-dimensional time-frequency matrix.
[0085] Step two, use the frequency domain amplitude of each frequency domain point to form an initial two-dimensional time-frequency matrix, and then perform logarithmic operation on each element in the initial two-dimensional time-frequency matrix (i.e., calculate the logarithmic value of each element with as the base), and after global standardization of each element, use each standardized element to form a final two-dimensional time-frequency matrix, i.e., obtain a two-dimensional time-frequency diagram.
[0086] In the present application, the global standardization process needs to first calculate the mean and global standard deviation of all elements in the matrix, then subtract the mean from each element in the matrix, and then divide each difference result by the global standard deviation to obtain the final elements.
[0087] Step three, taking the two-dimensional time-frequency diagram as the input of the residual attention convolutional network, outputting the deep local features of the two-dimensional time-frequency diagram through the residual attention convolutional network;
[0088] The structure of the residual attention convolutional network is shown in Figure 3 , which automatically learns the deep local features capable of representing target micro-Doppler, harmonic structure and other details from the two-dimensional time-frequency diagram. The working process of the residual attention convolutional network is as follows:
[0089] Taking the two-dimensional time-frequency diagram as the input of the Stem Layer module (this module adopts a ConvBlock structure), in the Stem Layer module, the two-dimensional time-frequency diagram is passed through a first convolutional layer (in the present application, a two-dimensional convolutional layer Conv2d is adopted);
[0090] Then the output of the first convolutional layer is taken as the input of the first group of normalization layers (GroupNorm), and the output of the first group of normalization layers is taken as the input of the first ReLU activation function layer (ReLU);
[0091] The output of the first ReLU activation function layer is taken as the input of a first regularization layer (two-dimensional space regularization Dropout in the present application), and the output of the first regularization layer is taken as the output of the Stem Layer module;
[0092] The output of the Stem Layer module is taken as the input of a first residual channel attention mechanism module (ResSEBlock), and the first residual channel attention mechanism module includes a main path (Main Path) and a shortcut path (Shortcut) therein, wherein:
[0093] In the main path, the output of the Stem Layer module is sequentially subjected to a second convolutional layer, a second group normalization layer, a second ReLU activation function layer, a second regularization layer, a third convolutional layer, a third group normalization layer, and a first attention module SE, and the output of the first attention module SE is taken as the output of the main path.
[0094] The first attention module SE includes a first global average pooling layer (Global Average Pooling, GAP), a third ReLU activation function layer (ReLU), a first fully connected layer (FC), and a first sigmoid activation function layer therein in sequence.
[0095] In the shortcut path, the output of the Stem Layer module is sequentially subjected to a fourth convolutional layer and a fourth group normalization layer, and the output of the fourth group normalization layer is taken as the output of the shortcut path.
[0096] The output of the main path and the output of the shortcut path are added, and the addition result is taken as the input of a fourth ReLU activation function layer, and the output of the fourth ReLU activation function layer is taken as the output of the first residual channel attention mechanism module.
[0097] The output of the first residual channel attention mechanism module is taken as the input of a second residual channel attention mechanism module, and the output of the second residual channel attention mechanism module is taken as the input of a third residual channel attention mechanism module.
[0098] The specific structure in each residual channel attention mechanism module is the same, and the three residual channel attention mechanism modules can output features with more abstract semantic information.
[0099] In the main path, the spatial information is compressed by global average pooling, and after the dimension reduction-up dimension transformation of the fully connected layer, the adaptive weight of each channel of the time-frequency graph is output, realizing the adaptive feature weight calibration; in the shortcut path, the dimension matching is realized by the convolutional layer and the group normalization layer. The outputs of the two paths are fused by element addition, and then the ReLU activation function is used to generate the output of the residual channel attention mechanism module.
[0100] The output of the third residual channel attention mechanism module is used as the input of the feature aggregation module, which includes the first adaptive average pooling layer (AdaptiveAvgPool).
[0101] The feature aggregation module compresses feature maps of arbitrary spatial dimensions to a uniform 1×1, thereby aggregating the spatial information of each feature channel into a single value, resulting in a channel count of [number missing]. The global structural feature vector (in this invention, The value is 64).
[0102] The output of the feature aggregation module is used as the input of the feature head module. Within the feature head module, the output of the feature aggregation module is sequentially passed through a flattening operation, a first normalization layer (LayerNorm), a third regularization layer, a second fully connected layer (FC), and a fifth ReLU activation function layer. The output of the fifth ReLU activation function layer is used as the output of the residual attention convolutional network.
[0103] The feature head module converts the tensor into a two-dimensional form through a flattening operation, and then maps the aggregated features to the final embedding space, finally outputting a D-dimensional deep local feature vector (in this invention, the value of D is 32).
[0104] Step 4: Extract the global structural features of the two-dimensional time-frequency graph using discrete Chebyshev moments; specifically:
[0105] The first in the two-dimensional time-frequency graph The value of each element is represented as Discrete Chebyshev moments are obtained using discrete Chebyshev polynomials:
[0106]
[0107] In the formula, and For discrete Chebyshev polynomials, for The order of for The order of For two-dimensional time-frequency plots in Magnitude in direction ( (corresponding timeline) For two-dimensional time-frequency plots in Magnitude in direction ( (corresponding frequency axis) and All are constants (in this invention, constants are constants) and The values are all 1. ), The th in the discrete Chebyshev moments Line number Iterate through the elements of the column. and After considering all possible values, the discrete Chebyshev moments can be obtained;
[0108] Calculate the absolute value of each element in the discrete Chebyshev moments, then calculate the average of the absolute values and use the average as the scaling factor; then divide each element value in the discrete Chebyshev moments by the scaling factor to obtain new discrete Chebyshev moments.
[0109] Then, the tanh function is used to compress each element value in the new discrete Chebyshev moments to the range [-1, 1] to obtain the normalized discrete Chebyshev moments. Finally, the elements in the normalized discrete Chebyshev moments are flattened row by row into one-dimensional feature vectors, and the flattened one-dimensional feature vectors are used as global structural features.
[0110] In step four of this invention, discrete Chebyshev moments are used to extract invariant features from a two-dimensional time-frequency graph that can describe its global shape and energy distribution.
[0111] Step 5: Use the deep local features and global structural features of the two-dimensional time-frequency graph as input to the feature fusion network. The feature fusion network fuses the deep local features and global structural features with higher information density to obtain the fused features of the two-dimensional time-frequency graph.
[0112] The structure of the feature fusion network is as follows Figure 4 As shown, the working process of the feature fusion network is as follows:
[0113] Step 51: Use the local depth features as input to the first linear projection layer (nn.Linear), and output the results through the first linear projection layer. Dimensional depth local feature token (STFT_Token);
[0114] The global structural features are used as input to the second linear projection layer, and the output is obtained through the second linear projection layer. Dimensional global structural feature token (Moms_Token);
[0115] And introduce a learnable Dimensional classification tokens (CLS_Token) serve as information aggregators;
[0116] Step five two, concatenate the deep local feature token, the global structure feature token and the classification token in the sequence dimension to form a token sequence with a shape of ;
[0117] and introduce a learnable positional embedding matrix;
[0118] Step five three, add the token sequence and the positional embedding matrix element by element through the broadcast mechanism of PyTorch to obtain an addition result, which gives the token sequence a position information in the addition result;
[0119] Step five four, take the addition result in step five three as the input of the first multi-head self-attention layer, take the output of the first multi-head self-attention layer as the input of the second multi-head self-attention layer, take the output of the second multi-head self-attention layer as the input of the third multi-head self-attention layer;
[0120] and take the output of the third multi-head self-attention layer as the input of the second layer normalization layer, take the output of the second layer normalization layer as the input of the multi-layer perception (MLP), and take the output of the multi-layer perception as the fusion feature of the two-dimensional time-frequency graph.
[0121] Step six, take the fusion feature of the two-dimensional time-frequency graph as the input of the detector, and output the unmanned aerial vehicle detection result through the detector.
[0122] The working process of the detector is as follows:
[0123] Take the fusion feature of the two-dimensional time-frequency graph as the input of the third layer normalization layer, and stabilize the input distribution through the layer normalization layer, and then perform nonlinear transformation on the output of the third layer normalization layer through the hidden unit;
[0124] In the hidden unit, the output of the third layer normalization layer is sequentially input into the third linear projection layer, the sixth ReLU activation function layer and the fourth regularization layer, and the output of the fourth regularization layer is taken as the output of the hidden unit;
[0125] Then take the output of the hidden unit as the input of the Softmax function, and the output of the Softmax function is the binary classification result, that is, the probability of detecting the unmanned aerial vehicle through the Softmax function;
[0126] If the probability of detecting the unmanned aerial vehicle is greater than 0.5, it means that the unmanned aerial vehicle is detected, otherwise, the unmanned aerial vehicle is not detected.
[0127] The process and results of the joint training of the residual attention convolutional network, the feature fusion network and the detector of the application will be described as follows in combination with the drawings:
[0128] (1) FMCW radar parameter configuration:
[0129] The parameters of the FMCW radar are configured through the mmWave Studio software, wherein the radar starting frequency , the frequency modulation slope , the pulse width , the transmission and reception setting is single transmission and single reception, the sampling point number , the sampling rate , the total number of single signal frames , the total number of single frames Chirp , and the transmission success rate .
[0130] (2) As shown in Figure 2 , the TI company's IWR1642 Boost FMCW radar is used to collect echo signals of four types of targets, i.e., unmanned aerial vehicles, bionic birds, pedestrians and bicycles, at different distances, speeds and attitudes;
[0131] The attitude of the unmanned aerial vehicle can be kept at a uniform speed, accelerated flight or hovering;
[0132] The bionic bird is randomly released to fly at any angle within the radar field of view after the spring is tightened;
[0133] The pedestrians and bicycles can keep accelerating, uniform motion or stationary state;
[0134] The distance r of each collection target relative to the radar is set to be ≤ 20 m, and the speed v is set to be ≤ 10 m / s. Each signal collection scene includes one of the four types of targets, i.e., unmanned aerial vehicles, bionic birds, pedestrians and bicycles, and at least 400 groups of data should be collected for each type of target. Each group of collected data includes 524288 sampling points, and the collected data is stored in the.bin format.
[0135] (3) Adding data labels
[0136] Each group of echo signals received by the FMCW radar is mixed with the radar transmission signal to obtain beat signals, and the beat signals are labeled to obtain a radar echo signal dataset. The labels of the unmanned aerial vehicle, bionic bird, pedestrian and bicycle are respectively marked as uav, bird, people and bike;
[0137] (4) Model training
[0138] The labeled beat signal dataset is divided into a training set and a test set. The proportion of data collected under various scenes in the training set and the test set is 3:1, and the initial learning rate is set to Cross-entropy loss function was selected as the loss function, AdamW was selected as the optimizer, and mixed precision training (AMP) and gradient pruning (max_norm=1.0) techniques were used to accelerate and stabilize the training process; the total number of training rounds was 50.
[0139] Two-dimensional time-frequency maps corresponding to each group of data in the training set are obtained respectively. Each two-dimensional time-frequency map is then input into a residual attention convolutional network. The residual attention convolutional network extracts the deep local features of each two-dimensional time-frequency map. The global structural features of each two-dimensional time-frequency map are extracted using discrete Chebyshev moments.
[0140] The feature fusion network based on the Transformer architecture is used to fuse deep local features and global structural features. That is, the global structural features and deep local features corresponding to the two-dimensional time-frequency map are input into the feature fusion network in batches:
[0141] 1. Obtain the dimension through feature projection. It uses deep local feature tokens and global structural feature tokens, and introduces learnable... Dimension Classification Token.
[0142] 2. Concatenate the deep local feature token, global structural feature token, and classification token along the sequence dimension to form a shape of... The token sequence, where, The batch size is indicated by adding the token sequence to the position encoding matrix element by element to assign position information to the token sequence.
[0143] 3. The token sequence, endowed with location information, is sequentially passed through three multi-head attention layers. Within each multi-head attention layer, the tokens are... The feature space of dimension is divided into An independent dimensional subspace, in which, Indicates the number of heads of attention.
[0144] Each attention head performs self-attention computation within its corresponding subspace. The self-attention mechanism calculates the association weights between each token in the sequence and all other tokens. Through the design of multi-head self-attention, the model can... We can examine and learn about the relationship between two heterogeneous features from different perspectives.
[0145] 4. Feature fusion
[0146] After deep information interaction within three multi-head self-attention mechanism layers, the classification token has fully gathered all the fused information from the two heterogeneous features and is output through the final attention mechanism layer. characteristics, and finally project the characteristics into dimensional fusion characteristics;
[0147] In the present application, 64, , , .
[0148] 5, the dimensional fusion characteristics are taken as the input of the detector, that is, the dimensional fusion characteristics are first subjected to layer normalization (LayerNorm) to stabilize the input distribution, and then subjected to a nonlinear transformation by a hidden unit to map the dimensional input characteristics to dimensional. Unlike the actual unmanned aerial vehicle detection process, as shown in Figure 5 , the hidden representation output by the hidden unit is mapped to a dimensional category logits space, and it is output which specific targets exist in the detection result.
[0149] In the present application, corresponding to the 4 types of targets in the data set, the detector adopts the Xavier uniform initialization strategy, and the output layer weight is subjected to scale reduction (gain = 0.7) to improve the numerical stability in the early stage of training.
[0150] The experimental results are shown in Figure 6 , as the number of training iteration rounds increases, the detection accuracy of the model on the test set steadily improves, and after 20-30 training rounds, it reaches convergence. The detection accuracy of the unmanned aerial vehicle on the test set is the highest when the iteration training round number of the present application method is 37, reaching 98.1%. In an environment with interference such as pedestrians, bionic birds, and bicycles, the present application method can still accurately detect unmanned aerial vehicles, verifying the effectiveness of the present application method.
[0151] The above examples of the present application are only to illustrate the calculation model and calculation process of the present application, and are not a limitation on the embodiments of the present application. For ordinary skilled persons in the art, other different forms of changes or variations can be made on the basis of the above description, and it is impossible to exhaust all embodiments here. Any obvious changes or variations derived from the technical solutions of the present application still fall within the protection scope of the present application.
Claims
1. A UAV detection method based on frequency-modulated continuous wave radar and fused features, characterized in that, The method specifically includes the following steps: Step 1: Use FMCW radar to transmit Chirp signals to the area to be detected, then collect the echo signals received by FMCW radar, and mix the transmitted signals and received echo signals to obtain the beat frequency signal. Step 2: Perform a short-time Fourier transform on the beat frequency signal to obtain a two-dimensional time-frequency diagram; Step 3: Use the two-dimensional time-frequency map as input to the residual attention convolutional network, and output the deep local features of the two-dimensional time-frequency map through the residual attention convolutional network; Step 4: Extract the global structural features of the two-dimensional time-frequency graph using discrete Chebyshev moments; Step 5: Use the deep local features and global structural features of the two-dimensional time-frequency map as input to the feature fusion network. The feature fusion network fuses the deep local features and global structural features to obtain the fused features of the two-dimensional time-frequency map. Step 6: Use the fused features of the two-dimensional time-frequency map as input to the detector, and output the UAV detection results through the detector.
2. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 1, characterized in that, The chirp signal transmitted by the FMCW radar is a sawtooth wave, that is, the form of the transmitted signal is: in, Indicates the transmission of a signal. The amplitude of the transmitted signal, The imaginary unit, The starting frequency, For frequency modulation slope, For time, The pulse repetition period, The pulse width. It is a natural number.
3. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 2, characterized in that, The echo signal received by the FMCW radar is in the following form: in, The echo signal received by the FMCW radar. The radar reflectance of the UAV surface. This represents the time delay of the drone's echo signal.
4. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 3, characterized in that, The beat frequency signal is in the form of: in, This is a beat frequency signal. This represents the amplitude of the beat frequency signal.
5. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 4, characterized in that, The specific process of step two is as follows: Step 21: Using a length of The Hanning window slides on the beat frequency signal, with a sliding step size of [missing information]. By sliding, the local signal corresponding to each position of the Hanning window can be obtained; Step 22: Multiply the local signal corresponding to each position of the Hanning window with the Hanning window function, and then perform a fast Fourier transform on each multiplication result to obtain the frequency domain amplitude of each frequency domain point in the fast Fourier transform result. Steps 2 and 3: Use the frequency domain amplitude values of each frequency domain point to form an initial two-dimensional time-frequency matrix. Then, perform logarithmic operations on each element in the initial two-dimensional time-frequency matrix. After global standardization, use the standardized elements to form the final two-dimensional time-frequency matrix, thus obtaining the two-dimensional time-frequency graph.
6. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 5, characterized in that, The operation process of the residual attention convolutional network is as follows: The two-dimensional time-frequency graph is used as the input to the Stem Layer module. Inside the Stem Layer module, the two-dimensional time-frequency graph is passed through the first convolutional layer. The output of the first convolutional layer is then used as the input of the first set of normalized layers, and the output of the first set of normalized layers is then used as the input of the first ReLU activation function layer. The output of the first ReLU activation function layer is used as the input of the first regularization layer, and the output of the first regularization layer is used as the output of the Stem Layer module. The output of the Stem Layer module is used as the input of the first residual channel attention mechanism module, which includes the main path and shortcut paths, wherein: Within the main path, the output of the Stem Layer module is sequentially passed through the second convolutional layer, the second set of normalization layers, the second ReLU activation function layer, the second regularization layer, the third convolutional layer, the third set of normalization layers, and the first attention module SE. The output of the first attention module SE is used as the output of the main path. Within the shortcut path, the output of the Stem Layer module passes through the fourth convolutional layer and the fourth set of normalization layers in sequence, and the output of the fourth set of normalization layers is used as the output of the shortcut path. The output of the main path and the output of the shortcut path are added together, and the result is used as the input of the fourth ReLU activation function layer. The output of the fourth ReLU activation function layer is then used as the output of the first residual channel attention mechanism module. Then, the output of the first residual channel attention mechanism module is used as the input of the second residual channel attention mechanism module, and the output of the second residual channel attention mechanism module is used as the input of the third residual channel attention mechanism module. The output of the third residual channel attention mechanism module is used as the input of the feature aggregation module, which includes a first adaptive average pooling layer. The output of the feature aggregation module is used as the input of the feature head module. Within the feature head module, the output of the feature aggregation module is passed sequentially through a flattening operation, a first normalization layer, a third regularization layer, a second fully connected layer, and a fifth ReLU activation function layer. The output of the fifth ReLU activation function layer is used as the output of the residual attention convolutional network.
7. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 6, characterized in that, The first attention module SE sequentially includes a first global average pooling layer, a third ReLU activation function layer, a first fully connected layer, and a first sigmoid activation function layer.
8. The UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 7, characterized in that, The specific process of step four is as follows: The first in the two-dimensional time-frequency graph The value of each element is represented as Discrete Chebyshev moments are obtained using discrete Chebyshev polynomials: In the formula, and For discrete Chebyshev polynomials, for The order of for The order of For two-dimensional time-frequency plots in Magnitude in direction For two-dimensional time-frequency plots in Magnitude in direction and All are constants. The th in the discrete Chebyshev moments Line number Column elements; Calculate the absolute value of each element in the discrete Chebyshev moments, then calculate the average of the absolute values and use the average as the scaling factor; then divide each element value in the discrete Chebyshev moments by the scaling factor to obtain new discrete Chebyshev moments. Then, the tanh function is used to compress each element value in the new discrete Chebyshev moments to the range [-1, 1] to obtain the normalized discrete Chebyshev moments. Finally, the elements in the normalized discrete Chebyshev moments are flattened row by row into one-dimensional feature vectors, and the flattened one-dimensional feature vectors are used as global structural features.
9. A UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 8, characterized in that, The working process of the feature fusion network is as follows: Step 51: Use the local depth features as input to the first linear projection layer, and output the local depth feature token through the first linear projection layer; The global structural features are used as input to the second linear projection layer, and the global structural feature token is output through the second linear projection layer. And class tokens are introduced as information aggregators; Step 52: Concatenate the deep local feature tokens, global structural feature tokens, and classification tokens along the sequence dimension to form a token sequence; And a position encoding matrix is introduced; Step 53: Add the token sequence to the position encoding matrix element by element to obtain the sum; Step 54: Use the summation result from Step 53 as the input to the first multi-head self-attention layer, then use the output of the first multi-head self-attention layer as the input to the second multi-head self-attention layer, and finally use the output of the second multi-head self-attention layer as the input to the third multi-head self-attention layer. The output of the third multi-head self-attention layer is then passed as input to the second normalization layer, and the output of the second normalization layer is used as input to the multilayer perceptron. The output of the multilayer perceptron is then used as the fusion feature of the two-dimensional time-frequency map.
10. A UAV detection method based on frequency-modulated continuous wave radar and fused features according to claim 9, characterized in that, The detector operates as follows: The fusion features of the two-dimensional time-frequency graph are used as the input of the third normalization layer, and the output of the third normalization layer is then subjected to nonlinear transformation through the hidden unit. Within the hidden unit, the output of the third normalization layer is passed sequentially through the third linear projection layer, the sixth ReLU activation function layer, and the fourth regularization layer, and the output of the fourth regularization layer is used as the output of the hidden unit. The output of the hidden unit is then used as the input to the Softmax function, and the probability of detecting the drone is output by the Softmax function. If the probability of detecting a drone is greater than 0.5, then a drone has been detected; otherwise, no drone has been detected.