A UAV target recognition method based on deep convolution and normalization

Abstract

The present invention belongs to the technical field of unmanned aerial vehicle (UAV) target recognition, and particularly relates to a UAV target recognition method based on deep convolution and normalization. First, the radar echo data sequence of a multi-rotor UAV is preprocessed, and then input into a comprehensive network of deep convolution features and normalization features for classification and recognition. The comprehensive network of convolution features and normalization features includes 5 convolutional subnets, a convolutional feature integration subnet, and a normalization feature integration subnet. The convolutional feature integration subnet arranges the convolutional feature maps of the latter 3 convolutional subnets in hierarchical order, followed by a convolutional layer with a kernel size of 1×1 and a fully connected layer. The normalization feature integration subnet arranges the normalization feature maps of the latter 3 convolutional subnets in hierarchical order, also followed by a convolutional layer with a kernel size of 1×1 and a fully connected layer. Then, the fully connected layers in the convolutional feature integration subnet and the normalization feature integration subnet are input into a softmax classification layer to complete the classification and recognition of the target.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) target recognition, and particularly relates to a UAV target recognition method based on deep convolution and normalization. Background Art

[0002] Due to the characteristics of small size, slow flight speed, low flight altitude, and easy operation of UAVs, UAVs have been widely used in military, civilian and other fields. At the same time, they have brought great troubles to the management of low-altitude areas and national defense. Therefore, accurately identifying the type of UAV has very important application value.

[0003] In recent years, with the progress of deep learning technology, methods based on deep convolutional network models have also been widely applied to the automatic recognition of multi-rotor UAVs. The main feature is that they can automatically learn high-order non-linear features of targets beneficial to recognition from radar echo data. However, in conventional deep convolutional network model methods, only the convolutional features of the last layer can be used for classification and recognition, resulting in limited feature representation ability and thus reducing the recognition performance. Therefore, there is room for further improvement in the recognition rate of the multi-rotor UAV recognition method based on the deep convolutional network model. Summary of the Invention

[0004] The purpose of the present invention is to propose a UAV target recognition method based on deep convolution and normalization, which can better describe the feature information in radar echo data and improve the recognition rate of targets by effectively integrating convolutional features and normalization features.

[0005] The technical solution of the present invention is as follows:

[0006] A UAV target recognition method based on deep convolution and normalization, comprising the following steps:

[0007] S1. Define the radar echo data sequence of the obtained multi-rotor UAV as , where represents the length of the sequence, represents the -th data point, , and perform the following processing on the data sequence :

[0008]

[0009] where represents the modulus of the vector;

[0010] S2. Construct a comprehensive network model for deep convolutional features and normalized features, including five multi-convolution sub-networks, a convolutional feature synthesis sub-network, a normalized feature synthesis sub-network, a fully connected layer, and a classification layer. Among them, the five multi-convolution sub-networks are cascaded in sequence and are respectively defined as the first multi-convolution sub-network, the second multi-convolution sub-network, the third multi-convolution sub-network, the fourth multi-convolution sub-network, and the fifth multi-convolution sub-network. Each convolutional sub-network contains a convolutional layer, a pooling layer, and a normalization layer; the convolutional feature synthesis sub-network arranges the convolutional feature maps of the third multi-convolution sub-network, the fourth multi-convolution sub-network, and the fifth multi-convolution sub-network in hierarchical order, and the normalized feature synthesis sub-network arranges and splices the normalized feature maps of the third multi-convolution sub-network, the fourth multi-convolution sub-network, and the fifth multi-convolution sub-network in hierarchical order; both the convolutional feature synthesis sub-network and the normalized feature synthesis sub-network are followed by a convolutional layer with a kernel size of and a fully connected layer; the output of the fully connected layer is connected to the input of the classification layer, and the classification layer outputs class labels;

[0011] S3. Use the obtained in S1 to train the comprehensive network model constructed in S2. Specifically, use the BP method to train the model parameters of the entire deep network. The loss function is the mean squared error function, the optimization method is the steepest gradient descent method, the activation function is ReLU, and the optimal number of iterations and learning rate are determined by experiments;

[0012] S4. Input the obtained multi-rotor UAV radar echo data sequence into the trained comprehensive network model, and use the label corresponding to the maximum component in the output vector of the softmax classification layer as the target recognition category.

[0013] The beneficial effects of the present invention are that the two feature synthesis sub-networks of the present invention not only extract the original convolutional features but also contain normalized features, further enhancing the ability to describe the feature information in the target data. The simulation experiment results for 4 types of multi-rotor UAVs verify the effectiveness of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is the overall flow schematic diagram of the present invention;

[0015] Figure 2 is the structural block diagram of the convolutional sub-network;

[0016] Figure 3 is the structural block diagram of the convolutional feature synthesis sub-network;

[0017] Figure 4 is the structural block diagram of the normalized feature synthesis sub-network. DETAILED DESCRIPTION OF THE INVENTION

[0018] The following is a simulation to prove the effectiveness and progress of the present invention:

[0019] As Figure 1 shown, in the recognition process of the present invention, first, the radar echo data sequence of the multi-rotor UAV is preprocessed, and then input into the comprehensive network of deep convolutional features and normalized features for classification and recognition. The comprehensive network of convolutional features and normalized features includes 5 convolutional subnets, a convolutional feature synthesis subnet, and a normalized feature synthesis subnet. Each convolutional subnet contains a convolutional layer, a pooling layer, and a normalization layer. The convolutional feature synthesis subnet arranges the convolutional feature maps of the last 3 convolutional subnets in hierarchical order, followed by a convolutional layer with a kernel size of and a fully connected layer. The normalized feature synthesis subnet arranges the normalized feature maps of the last 3 convolutional subnets in hierarchical order, and also followed by a convolutional layer with a kernel size of and a fully connected layer. Then, the fully connected layers in the convolutional feature synthesis subnet and the normalized feature synthesis subnet are input into the softmax classification layer to complete the classification and recognition of the target.

[0020] Each convolutional subnet contains a convolutional layer, a pooling layer, and a normalization layer, and its structure is as Figure 2 shown. The convolutional feature synthesis subnet arranges the convolutional feature maps of the last 3 convolutional subnets in hierarchical order, followed by a convolutional layer with a kernel size of and a fully connected layer, and its structure is as Figure 3 shown. The normalized feature synthesis subnet arranges and splices the normalized feature maps of the last 3 convolutional subnets in hierarchical order, and also followed by a convolutional layer with a kernel size of and a fully connected layer, and its structure is as Figure 4 shown.

[0021] The simulation experiment designs 4 types of UAVs, including tri-rotor UAV, quad-rotor UAV, hex-rotor UAV, and octa-rotor UAV, and their simulation parameters are shown in Table 1. The simulation radar parameters include: radar carrier frequency is 24 GHz; pulse repetition frequency is 100 KHz; the distance between the target and the radar is 200 m; the pitch angle of the UAV relative to the radar is 10°, and the azimuth angle is 30°

[0022] Each type of target records 10 s of radar echo signals, which are divided into segments with a fixed length of 0.05 s (at least containing one rotation period), and the overlap between segments is 50%. Each segment contains 0.05×100000 = 5000 radar echo sampling data points, and each type has a total of 400 segments. Randomly select 200 segments from the 400 segments as the training data set, and the remaining 200 segments as the test data set. Then, the training data sets of the 4 types of targets include a total of 800 segments, and the test data set includes 800 segments. In order to reduce the number of input nodes of the network, the PCA method is first used to reduce the dimension of the training and test samples to 200. For the selected training data set, train the comprehensive network model of the deep convolutional features and normalized features in this paper. Then, use the trained deep network to identify the test data set, and the average correct recognition rate for the 4 types of multi-rotor UAVs is 97%. Among them, the number of iterations is 2000 times, the learning rate is 0.1, and noise is added to the UAV echo data with a signal-to-noise ratio of 10 dB.

[0023] Table 1 Simulation Parameters of Four UAVs

[0024] UAV code name Three-rotor UAV Four-rotor UAV Six-rotor UAV Eight-rotor UAV Number of rotors 3 4 6 8 Rotor speed r / min 1200 1200 1200 1200 Blade length (m) 0.3 0.3 0.3 0.3 Distance from axis to origin (m) 0.8 0.8 0.8 0.8

Claims

1. A method for drone target recognition based on deep convolution and normalization, characterized in that, Including the following steps: S1. Define the radar echo data sequence obtained from the multi-rotor UAV as \(x = [x_1\ x_2\ \cdots\ x n \), where \(n\) represents the length of the sequence, and \(x i \) represents the \(i\)-th data point, \(i = 1, 2, \cdots, n\). Process the data sequence \(x\) as follows: Where, ||·|| represents the norm of a vector; S2. Construct a comprehensive network model of deep convolutional features and normalized features, including five multi-convolution sub-networks, a convolutional feature synthesis sub-network, a normalized feature synthesis sub-network, a fully connected layer, and a classification layer. The five multi-convolution sub-networks are cascaded in sequence and are respectively defined as the first multi-convolution sub-network, the second multi-convolution sub-network, the third multi-convolution sub-network, the fourth multi-convolution sub-network, and the fifth multi-convolution sub-network. Each convolutional sub-network contains a convolutional layer, a pooling layer, and a normalization layer; the convolutional feature synthesis sub-network arranges the convolutional feature maps of the third multi-convolution sub-network, the fourth multi-convolution sub-network, and the fifth multi-convolution sub-network in a hierarchical order, and the normalized feature synthesis sub-network arranges and splices the normalized feature maps of the third multi-convolution sub-network, the fourth multi-convolution sub-network, and the fifth multi-convolution sub-network in a hierarchical order; both the convolutional feature synthesis sub-network and the normalized feature synthesis sub-network are followed by a convolutional layer with a kernel size of 1×1 and a fully connected layer; the output of the fully connected layer is connected to the input of the classification layer, and the classification layer outputs class labels; S3. Using what is obtained in S1 Train the comprehensive network model constructed in S2. Specifically, use the BP method to train the model parameters of the entire deep network. The loss function is the mean squared error function, the optimization method is the steepest gradient descent method, the activation function is ReLU, and the optimal number of iterations and learning rate are determined by experiments; S4. Input the obtained multi-rotor UAV radar echo data sequence into the trained comprehensive network model, and use the label corresponding to the maximum component in the output vector of the softmax classification layer as the target recognition category.

Images