UAV Detection Method Based on Residual Network Multi-View Feature Fusion

Abstract

This invention discloses a UAV detection method based on multi-view feature fusion using residual networks, comprising: constructing multi-view data: obtaining the time-domain plot, short-time Fourier transform plot, continuous wavelet transform plot, and Wegener-Will distribution plot of the measured signal; constructing a ResNet34, including an input structure, an intermediate structure, and an output structure; the input structure processes the input data through convolution and max pooling operations; the intermediate structure consists of four similar structural layers, each consisting of multiple residual blocks, each residual block containing three convolutional layers and a shortcut connection; starting from the second structural layer, the initial residual block of each structural layer also contains an up-dimensional sampling structure; constructing a multi-view feature fusion network model based on residual networks, and outputting the UAV detection accuracy after multi-view feature fusion at the output end. This invention improves the UAV detection efficiency by fusing features from the multi-view data of the signal.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) signal detection, and specifically to a UAV detection method based on residual network multi-view feature fusion. Background Technology

[0002] Due to the characteristics of drones, such as ultra-low altitude flight, slow speed, and small radar cross-section, accurate detection of drones remains challenging.

[0003] Deep learning can automatically acquire deep, essential feature information of targets, making it highly suitable for target detection. Therefore, introducing deep learning into the field of drone detection has become a research hotspot in recent years. However, traditional methods for drone detection using deep learning mostly rely on single-view data of the drone for target detection. Since single-view data cannot fully reflect the features and information of the data itself, traditional deep learning drone detection methods generally have low detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the limitations of traditional neural network methods that use single-view features of UAV echo signals to detect UAVs, and to provide a UAV detection method based on multi-view feature fusion using residual networks. This method improves the detection efficiency of UAVs by fusing features from multi-view data of the signal.

[0005] The technical solution to achieve the objective of this invention is:

[0006] A UAV detection method based on residual network multi-view feature fusion includes the following steps:

[0007] Step 1, Construct multi-view data:

[0008] Obtain the time-domain plot, short-time Fourier transform plot, continuous wavelet transform plot, and Wegener-Will distribution plot of the measured signal;

[0009] Step 2, construct the residual neural network model:

[0010] ResNet34 is constructed, consisting of three parts: input structure, intermediate structure, and output structure. The input structure processes the input data through convolution and max pooling operations. The intermediate structure consists of four similar structural layers, each consisting of multiple residual blocks. Each residual block contains three convolutional layers and a shortcut connection. In addition, starting from the second structural layer, the initial residual block of each structural layer also contains an up-dimensional sampling structure.

[0011] Step 3: Construct a multi-view feature fusion network model based on residual networks:

[0012] Step 3-1: Input the multi-view data constructed in Step 1 into the ResNet34 constructed in Step 2. At the input end of the network, set each view to RGB three-channel input, and set the length and width of the view to 64, that is, the input dimension of each view is 3×64×64; then, use a deep neural network to extract features from the temporal map, STFT map, CWT map, and WVD map respectively, and extract the network terminal feature vector with a dimension of 1×512.

[0013] Step 3-2: The extracted 1×512 network end feature vectors are stacked and fused. The fused 4×512 feature vectors are then input into the fully connected layer network for retraining. The fully connected layer structure consists of one fully connected layer and one activation function layer connected in series.

[0014] Step 3-3: Save and test the model, and finally output the UAV detection accuracy after multi-view feature fusion at the output end.

[0015] Experimental results show that the worst drone detection accuracy of single-view is 67.49%, and the best is 97.14%. The drone detection accuracy after multi-view feature fusion reaches 99.75%. Therefore, it can be concluded that the drone detection accuracy after multi-view feature fusion is better than that of single-view drone detection, proving the effectiveness of the method.

[0016] Effects or advantages of the present invention:

[0017] (1) This invention uses four views as multi-view data: time domain diagram, short-time Fourier transform diagram, continuous wavelet transform diagram and Wegener-Will distribution diagram of the measured signal. The four views are simple in principle and easy to obtain.

[0018] (2) In view of the limitations of traditional neural network methods for detecting UAVs using single-view features of UAV echo signals, this invention uses the ResNet34 network structure as a basis to fuse features of multi-view data of the signal, thereby improving the detection efficiency of UAVs. Attached Figure Description

[0019] Figure 1 A schematic diagram of the ResNet34 architecture;

[0020] Figure 2 This is a flowchart of a multi-view feature fusion model based on residual networks. Detailed Implementation

[0021] The present invention will be further described below with reference to the embodiments and 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.

[0022] Example:

[0023] A multi-view feature fusion method for UAV detection based on residual networks includes the following steps:

[0024] S1, Constructing multi-view data:

[0025] The measured signal x(t) was analyzed using MATLAB software to obtain different views of the measured signal:

[0026] S1-1, obtain the time-domain graph of the measured signal x(t), which can be plotted using MATLAB software;

[0027] S1-2, obtain the Short-Time Fourier Transform (STFT) graph of the measured signal x(t). The formula for the Short-Time Fourier Transform is as follows:

[0028]

[0029] In the formula, t represents time in seconds (s); ω represents angular frequency in radians per second (rad / s); x() represents the measured signal; w() represents the window function; and e -jωτ Let represent a complex exponential function, where e represents the base of the natural logarithm, j represents the imaginary unit, τ represents the time delay, and dτ represents the integration over τ. A Gaussian window function is used to perform a short-time Fourier transform on the signal. A common form of the Gaussian window function is:

[0030]

[0031] Where a is an arbitrary constant;

[0032] S1-3, obtain the continuous wavelet transform (CWT) graph of the measured signal x(t). The formula for the continuous wavelet transform is as follows:

[0033]

[0034] In the formula, b is the scaling parameter, c is the translation parameter, ψ(·) is the wavelet function, and * denotes complex conjugate; the signal is subjected to continuous wavelet transform using a complex Morlet wavelet with a bandwidth of 3 and a center frequency of 3, i.e., "cmor3-3". The transformation formula of the Cmor wavelet function is as follows:

[0035]

[0036] In the formula, f c f is the center frequency of the complex Morlet wavelet when analyzing the signal. d The wavelet bandwidth;

[0037] S1-4, obtain the Wigner-Ville Distribution (WVD) plot of the measured signal x(t). The autocorrelation function of the measured signal x(t) can be expressed as:

[0038]

[0039] The Wigner-Will distribution is the autocorrelation function R of the measured signal x(t). s The FFT transformation of (t,τ) is expressed as follows:

[0040]

[0041] S2, refer to Figure 1 Construct a residual neural network model using Table 1:

[0042] Table 1 ResNet34 Structure Parameters

[0043]

[0044] Figure 1 In the table, CBR represents a convolutional layer structure consisting of convolutional computation operations (Conv), batch normalization operations (BN), and ReLU activation function operations. Max Pool represents max pooling, Average Pool represents average pooling, and Fc represents a fully connected layer. In Table 1, Conv represents a convolutional operation, and the number before Conv represents the size of the convolutional kernel in the convolutional operation. For example, 7×7conv means that the height and width of the convolutional kernel in the convolutional operation are both 7. stride represents the kernel stride. MaxPool represents max pooling, and Average Pool represents average pooling. The numbers before Max Pool and Average Pool represent the size of the kernel in their respective operations. Softmax represents the activation function.

[0045] ResNet34 consists of three parts: input structure, intermediate structure, and output structure. The input structure processes the input data through convolution and max pooling operations. The intermediate structure consists of four similar structural layers, each consisting of multiple residual blocks. Each residual block contains three convolutional layers and a shortcut connection. In addition, starting from the second structural layer, the initial residual block of each structural layer also contains an up-dimensional sampling structure.

[0046] S3, according to Figure 2 Construct a multi-view feature fusion network model based on residual networks, using Table 2 as an example:

[0047] Table 2. Network structure parameters of the fully connected layer in the multi-view feature fusion network model based on residual networks.

[0048]

[0049] S3-1, input the multi-view data constructed in step 1) into the ResNet34 constructed in S2. At the input end of the network, set each view to be input as RGB three channels, and set the length and width of the view to 64, that is, the input dimension of each view is 3×64×64; then, use a deep neural network to extract features from the temporal map, STFT map, CWT map and WVD map respectively, and extract the network terminal feature vector with a dimension of 1×512.

[0050] S3-2, the extracted 1×512 dimension network end feature vectors are stacked and fused, and the fused 4×512 dimension features are input into the fully connected layer network for retraining. The fully connected layer structure consists of one fully connected layer and one activation function layer connected in series.

[0051] S3-3 saves and tests the model, and outputs the results at the output terminal.

[0052] This invention utilizes ResNet, which employs direct connections across two or more consecutive layers. Input data undergoes convolutional operations to obtain new feature maps, which are then residually connected to the input signal to achieve cross-layer mappings. These maps are then processed by activation functions before being output. Unlike other traditional neural networks, ResNet increases the network's depth and width by introducing residual blocks, allowing for more efficient use of information from each layer and further enhancing its ability to represent image features. Finally, the feature vectors are stacked and fused before being fed into the fully connected network for retraining, effectively improving the detection efficiency of drones.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A UAV detection method based on residual network multi-view feature fusion, characterized in that, Includes the following steps: Step 1, Construct multi-view data: Obtain the time-domain plot, short-time Fourier transform plot, continuous wavelet transform plot, and Wegener-Will distribution plot of the measured signal; Step 2, construct the residual neural network model: ResNet34 is constructed, consisting of three parts: input structure, intermediate structure, and output structure. The input structure processes the input data through convolution and max pooling operations. The intermediate structure consists of four similar structural layers, each consisting of multiple residual blocks. Each residual block contains three convolutional layers and a shortcut connection. In addition, starting from the second structural layer, the initial residual block of each structural layer also contains an up-dimensional sampling structure. Step 3: Construct a multi-view feature fusion network model based on residual networks: Step 3-1: Input the multi-view data constructed in Step 1 into the ResNet34 constructed in Step 2. At the input end of the network, set each view to RGB three-channel input, and set the length and width of the view to 64, that is, the input dimension of each view is 3×64×64; then, use a deep neural network to extract features from the temporal map, STFT map, CWT map, and WVD map respectively, and extract the network terminal feature vector with a dimension of 1×512. Step 3-2: The extracted 1×512 network end feature vectors are stacked and fused. The fused 4×512 feature vectors are then input into the fully connected layer network for retraining. The fully connected layer structure consists of one fully connected layer and one activation function layer connected in series. Step 3-3: Save and test the model, and finally output the UAV detection accuracy after multi-view feature fusion at the output end.

Images