A low, slow and small target detection method in a clutter environment based on a neural network

Abstract

The present application belongs to the technical field of radar data processing, and proposes a low, slow and small target detection method in clutter environment based on neural network. The present application preprocesses the point track information of the clutter and low, slow and small target collected by the radar, establishes a convolutional neural network model, extracts the aperiodic features of the single frame data of the radar, constructs an LSTM recurrent neural network model, and obtains the periodic features of the inter-frame data. The obtained feature vectors are fused to realize the classification of the clutter and low, slow and small target, and then the clutter starting is suppressed, and a good clutter filtering effect is obtained. The present application overcomes the shortcomings of low, slow and small target detection radar, such as low classification accuracy of clutter, poor clutter filtering effect, and poor detection effect of low, slow and small target. Without increasing additional hardware, the present application realizes the classification of radar clutter and low, slow and small target in a single radar antenna cycle, and then better filters the radar clutter at the data processing level, and improves the detection capability of radar low, slow and small target.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar data processing. BACKGROUND

[0002] Due to low cost, wide application and potential danger to public safety, unmanned aerial vehicles have become an increasingly important radar target. Small unmanned aerial vehicles belong to typical low, slow and small targets, which have the characteristics of low flight height, slow flight speed, small target radar echo intensity, easy to be blocked by the surrounding environment, and easy to be submerged in environmental clutter.

[0003] In a complex electromagnetic environment or a strong clutter environment, there are quite a few irregular clutter regions composed of ground clutter groups, point sources and distributed clutter in the radar coverage range, which may be static clutter such as ground buildings, slopes and power poles that are relatively stationary to the radar, or dynamic clutter such as clouds, rain, sea waves and wind-swayed trees that are relatively moving to the radar. Since the radar echo intensity of low, slow and small targets is weak, the Doppler frequency is low. In order to better realize the detection and tracking of low, slow and small targets, it is necessary to reduce the Doppler frequency threshold and the signal-to-noise ratio detection threshold, but the environmental clutter and interference targets are also easily detected, forming radar false alarms, and the formation of low, slow and small target tracks will be affected, which may produce clutter tracks on the display interface, making it difficult to distinguish between clutter and targets. Therefore, filtering out clutter tracks and suppressing clutter can better extract targets, reduce clutter generation and clutter track generation, and realize better detection and tracking processing of low, slow and small targets. SUMMARY

[0004] In order to overcome the shortcomings of low recognition rate of clutter classification by prior art point track prior information processing at the level of traditional data processing, poor clutter filtering effect, and low detection capability of radar for low, slow and small targets, the application provides a low, slow and small target detection method in a clutter environment based on a neural network.

[0005] In order to solve the above technical problems, the application provides a low, slow and small target detection method in a clutter environment based on a neural network, and the technical scheme comprises:

[0006] (1) The point track data of radar clutter and low, slow and small targets are taken as training samples, and the true class labels of low, slow and small targets are confirmed by photoelectric equipment;

[0007] (2) The point track data in step (1) are preprocessed to eliminate the influence caused by different dimensions of different dimensions;

[0008] (3) The data in step (2) are processed by a convolutional neural network to extract the non-periodic features of a single frame of radar data through a model;

[0009] (4) Considering the time sequence of inter-frame track data, LSTM neural network is constructed to train the model and obtain the periodic characteristics of inter-frame data.

[0010] (5) The feature vectors output by the neural network are fused to obtain the classification results of radar clutter and low, slow and small targets, and then the clutter is filtered out, the clutter is suppressed, and the detection ability of radar low, slow and small targets is improved.

[0011] Further, the features of the clutter and target track data in step (1) mainly include: period, time, track azimuth, track distance, track elevation, Doppler velocity, track amplitude, elevation quality and other feature dimensions. The original data set is: Where n and m are the number of data and the number of feature attributes, respectively.

[0012] Further, the minimum-maximum normalization method of step (2) for track data is: Where x n is the normalized value, x min and x max are the minimum and maximum values of each attribute value in the data set, respectively. The influence of different dimensions of different dimensions of data is eliminated.

[0013] Further, step (3) processes the data set by convolutional neural network to extract the non-periodic features F a of the radar single frame data. Convolutional neural network (CNN, Convolutional Neural Networks) is a deep learning framework that processes and interprets high-dimensional input data through convolution and pooling. The structure is input layer, convolution layer, pooling layer, activation layer, full connection layer and output layer. CNN processes high-dimensional data through learning kernel, parameter sharing strategy and pooling operation, effectively reduces the number of learned parameters during training, and prevents the network from overfitting. Convolution layer is the basis of CNN, including learning kernel, feature filtering, and convolution layer generates feature maps of input data through convolution operation. The convolution process is Where is the i-th output feature map obtained by the l-th convolution operation, is the i-th output feature map obtained by the l-1-th convolution operation, is the bias value of the k-th convolution kernel of the l-th network, where W is the weight matrix preset for the current layer, and σ(·) is the Sigmoid activation function. The convolution process is a linear transformation, and the non-linear problem is handled by the activation function. After the convolution operation, the input feature map of the current convolution layer and the neuron are dot product operations, and the non-linear problem is converted and processed by the activation function. Batch normalization layer: used to accelerate the running process of the neural network, which normalizes the output of the previous layer. The output matrix of the pooling layer is where a(u,v) is the value of the u-th row and v-th column in the input matrix of the pooling layer. The pooling layer is used to realize dimension reduction, which reduces the spatial size of the feature map, and dimension reduction helps the network to maintain the invariance of the conversion process when learning important features from the original data. Fully connected layer: the fully connected layer is a basic artificial feedforward neural network, which converts the multi-dimensional feature map generated by the convolution layer to a one-dimensional vector. The output layer is after the fully connected layer, which evaluates the possibility of each input belonging to a certain class, and provides the prediction function of the network. Neural network learning involves an optimization problem, and the optimal algorithm adjusts the weight and learning rate to minimize the prediction loss.

[0014] Further, step (4) considers the time sequence of the inter-frame track data, constructs an LSTM neural network, trains the model, and obtains the periodicity feature F b of the inter-frame data. LSTM is a recurrent neural network algorithm that learns sequence dependencies through feedback connections. Compared with traditional recurrent neural network methods, better results are obtained by storing useful information in memory cells and discarding useless information. LSTM includes memory cells, forget gates, input gates, and output gates. The forget gate f j = σ(w f [h j-1 ,X j ]+b f ), b f is the bias of the forget gate weight, X j is the input data, and h j-1 is the state value of the previous hidden layer. The input gate i j = σ(w i [h j-1 ,X j ]+b i ), w i and w c are the weights of the input gate, and the corresponding biases are b i and b c . The Sigmoid function determines the new information that should be retained, and tanh(·) creates all possible values of input X j . The new cell state is updated by adding the previous cell state c j-1 . The output gate o j = σ(w o[h j-1 ,X j ]+b o ), w o is the weight of the output layer, b o is the bias of the weight.

[0015] Further, step (5) fuses the feature vector output by the network F = [F a , F b ] into the full connection layer y = ωF + b, where F is the input of the full connection layer, ω is the weight matrix, and b is the bias vector. Then the classification results of clutter and low, slow and small targets are obtained, and the clutter is filtered out, and the automatic start of the clutter data is suppressed.

[0016] The method overcomes the shortcomings of low detection capability of radar for low, slow and small targets and poor clutter filtering effect, and is a real-time, universal and efficient low, slow and small target detection method based on neural network in clutter environment. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the flowchart of the present application. DETAILED DESCRIPTION

[0018] Figure 1 is the implementation process of the present application, which will be described below with examples.

[0019] The preferred embodiments of the present application include:

[0020] 1. The clutter and low, slow and small target track information is collected by a Ku band low, slow and small target detection radar, and the low, slow and small target data includes unmanned aerial vehicle (Dajiang Ling 4Pro, Dajiang M600, Dajiang Yu 2, fixed-wing unmanned aerial vehicle), bird and other target track information. The real class label of the low, slow and small target is confirmed by photoelectric equipment. The target track data features mainly include: period, time, track direction, track distance, track elevation, Doppler velocity, track amplitude, elevation quality and other feature dimensions.

[0021] 2. The data is preprocessed, in order to make the data properties of the input model not affected by the dimension, all input data is standardized to the range of 0 to 1 using the min-max standardization method.

[0022] 3. The data set is processed by a convolutional neural network to extract the aperiodic features of a single frame of radar data. The convolutional neural network processes and interprets high-dimensional input data through convolution and pooling. The convolution kernel size is 3x1, the convolution step is 1, and the convolution output layer activation function is ReLU. Compared with the Sigmoid function, the ReLU function can prevent gradient dispersion and greatly improve the calculation efficiency. Pooling layers are used to achieve dimensionality reduction, which reduces the size of the feature mapping space. Dimensionality reduction helps the network maintain the invariance of the transformation process when learning important features from the original data. The stride of the pooling layer is 2.

[0023] 4. Considering the temporal nature of the inter-frame point data, an LSTM neural network is constructed, and the model is trained to obtain the periodic features of the inter-frame data. The LSTM has 4 layers and 16 neurons. The Sigmoid function determines the contribution of the LSTM memory to the output, and the value is mapped to the (-1, 1) interval by the nonlinear function tanh(·).

[0024] 5. The feature vectors output by the network are fused and input into the fully connected layer to obtain the classification results of clutter and low-speed small targets, thereby filtering out and suppressing the initiation of clutter.

[0025] To verify the effectiveness of the method in the embodiments, the low-speed, small target point data used in the experiment came from data collected by a certain type of anti-drone Ku radar at a test range on land. The point dataset includes clutter and low-speed, small target point information. The low-speed, small target data includes target point information such as drones (DJI Phantom 4 Pro, DJI M600, DJI Mavic 2, and fixed-wing drones) and birds. The target categories of the collected data were confirmed by optoelectronic equipment. Clutter and low-speed, small target types are represented by Class 1 and Class 2, respectively.

[0026] The table below shows the classification and identification results of clutter and low-speed, small targets for radar spot data selected at five different time periods. In time period 1, the cooperative target was a DJI Phantom 4 Pro. The table shows that the accuracy rate for clutter classification in time period 1 was 0.926, and the accuracy rate for low-speed, small target classification was 0.922. In time period 2, the cooperative targets were a DJI Phantom 4 Pro and a DJI M600. The table shows that the accuracy rate for clutter classification in time period 2 was 0.858, and the accuracy rate for low-speed, small target classification was 0.945. In time period 3, the cooperative targets were a DJI Phantom 4 Pro and a DJI Mavic 2. The table shows that the accuracy rate for clutter classification in time period 3 was 0.896, and the accuracy rate for low-speed, small target classification was 0.926. In time period 4, the cooperative targets were a DJI Phantom 4 Pro and a fixed-wing drone. The table shows that the accuracy rate for clutter classification in time period 4 was 0.924, and the accuracy rate for low-speed, small target classification was 0.90. During time period 5, the target in flight was a bird. As shown in the table below, the accuracy rate for clutter classification in time period 5 was 0.883, and the accuracy rate for low-speed, small target classification was 0.812. Experiments demonstrate that this embodiment significantly improves the accuracy of clutter and low-speed, small target classification, enhances radar clutter filtering, and improves the radar's ability to detect low-speed, small targets, thus providing a better guarantee for automatic initiation in subsequent radar data processing.

[0027]

[0028] This invention provides a method for detecting low-speed, small targets in cluttered environments based on neural networks. Application results show that this method achieves high clutter classification accuracy, good clutter filtering effect, and high detection capability for low-speed, small targets. It requires no additional hardware, has low implementation cost, and can be widely applied to related product fields, showing great potential for widespread application.

Claims

1. A neural network-based low, slow and small target detection method in a clutter environment, characterized in that: (1) the point track data of radar clutter and low, slow and small targets are taken as training samples, and the true class labels of low, slow and small targets are confirmed by photoelectric equipment; (2) the point track data in step (1) is preprocessed to eliminate the influence caused by different dimensions of different data; (3) the data in step (2) is processed by a convolutional neural network to extract the aperiodic features of radar single-frame data through the model; (4) the time sequence of inter-frame point track data is considered, an LSTM neural network is constructed, the model is trained, and the periodic features of inter-frame data are obtained; (5) the feature vectors output by the neural network are fused to obtain the classification results of radar clutter and low, slow and small targets, and then the clutter is filtered out, the clutter is suppressed, and the detection capability of radar low, slow and small targets is realized. The point track data of clutter and targets in step (1) includes period, time, point track azimuth, point track distance, point track elevation, Doppler velocity, point track amplitude and elevation quality. The data preprocessing in step (2) is standardization processing. ​ ​ ​ 2. The low, slow and small target detection method in a clutter environment based on a neural network according to claim 1, characterized in that: ​ 3. The low, slow and small target detection method in a clutter environment based on a neural network according to claim 1, characterized in that: ​ 4. The low, slow and small target detection method in a clutter environment based on a neural network according to claim 1, characterized in that: The step (3) performs convolutional neural network processing on each frame of radar track data to extract non-periodic features F of the radar single frame data a The convolution process is Wherein is the i-th output feature map obtained by the l-th layer convolution operation, is the bias value of the k-th convolution kernel of the l-th layer network, is the preset weight matrix of the current layer, and the pooling layer output matrix is 5. The low, slow and small target detection method in a clutter environment based on a neural network according to claim 1, characterized in that: The step (4) constructs an LSTM neural network model to extract periodic features F between radar multi-frame data b where the model forget gate f j = σ(w f [h j-1 ,X j ]+b f ), b f is the bias of the forget gate weight, X j is the input data, h j-1 is the state value of the previous hidden layer, w f is the weight matrix; the input gate is i j = σ(w i [h j-1 ,X j ]+b i ), w i and w c are the weights of the input gate, and the corresponding biases are b i and b c ; the output gate is o j = σ(w o [h j-1 ,X j ]+b o ), w o is the weight of the output layer, and b o is the bias of the weight.

6. The low, slow and small target detection method in a clutter environment based on a neural network according to claim 1, characterized in that: The step (5) fuses the feature vector output by the neural network F = [F a ,F b ] and inputs it to a full connection layer y = ωF + b, where F is the input of the full connection layer, ω is a weight matrix, and b is a bias vector.

Images