A radar automatic recognition method and device for low-altitude small targets

Abstract

The present invention discloses a radar automatic recognition method and device for low-altitude small targets. The method first performs radar data preprocessing; then uses a deep neural network to extract the features of the processed data, and then respectively uses supervised learning and unsupervised learning to perform binary classification and anomaly detection; finally, determines whether it is a drone, a bird or other low-altitude floating objects according to the results of binary classification and anomaly detection. This method solves the problem of identifying low-altitude floating objects, has a very wide range of applications, strong generalization ability, less computation, lower cost, and higher recognition accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal feature extraction and target recognition and classification, and particularly relates to a radar automatic recognition method and device for low-altitude small targets. Background Art

[0002] Low-altitude small targets refer to light and small unmanned aerial vehicles, birds, floating balloons, plastic bags, etc. They pose different degrees of threats to military and civilian airports and various military facilities. Identifying which type of target in the air has become a very urgent task.

[0003] The discovery of air targets mainly relies on radar, which has the working ability of all-weather and all-time, and has a long detection range. As the main means of current target detection, obtaining the ability of autonomous recognition and classification is of irreplaceable importance. However, after the current radar outputs the point traces and tracks of targets, it cannot further identify whether the target poses a military threat. In particular, light and small unmanned aerial vehicles are typical "low, slow, small" targets, and their RCS information, motion feature information, etc. are similar to those of birds and other low-altitude floating targets, and they are easily interfered by clutter such as people, vehicles, trees, and houses. Distinguishing low-altitude small targets based on radar data is a persistent problem.

[0004] Chinese Patent CN201610896005.6 distinguishes birds and light and small unmanned aerial vehicles detected by radar by manually extracting motion features. However, manually extracting features will result in poor generalization ability and low recognition rate. The solution of Chinese Patent CN202011419393.1 requires image recognition of targets, with complex calculations and high costs. The method of Chinese Patent CN201810851549.X for identifying unmanned aerial vehicles requires multiple means such as spectrum detection and optoelectronic detection. In the papers "UAV Classification and Recognition Based on Dual-Radar Micro-Motion Feature Fusion" by Zhang Pengfei, Li Gang, Huo Chaoying, etc. and "Drone Classification Using Convolutional Neural Networks With Merged Doppler Images" by BK Kim, HS Kang, SO Park, unmanned aerial vehicles and birds are identified through the micro-Doppler effect, but in many cases, micro-Doppler parameters are difficult to estimate. In addition, all of these literatures do not consider the small floating object type of low-altitude small targets that exist in actual applications. Summary of the Invention

[0005] The object of the present invention is to overcome the deficiencies existing in the prior art, and to provide a radar automatic recognition method and device for low-altitude small targets. The method first performs radar data preprocessing; then uses a deep neural network to extract the features of the processed data, and then uses supervised learning and unsupervised learning to perform binary classification and anomaly detection respectively; finally, it determines whether it is a drone, a bird or other low-altitude floating objects according to the binary classification and anomaly detection results. This method solves the problem of identifying low-altitude floating objects, has a very wide range of applications, strong generalization ability, less computational complexity, lower cost, and higher recognition accuracy.

[0006] In order to achieve the above technical effects, the present invention adopts the following technical solutions:

[0007] A radar automatic recognition method for low-altitude small targets, characterized by comprising the following steps:

[0008] a) Radar data preprocessing;

[0009] b) Using a deep neural network to extract the features of the processed data;

[0010] c) Using supervised learning and unsupervised learning to perform binary classification and anomaly detection;

[0011] d) Identifying whether the input data is a drone, a bird or other low-altitude floating objects according to the binary classification and anomaly detection results.

[0012] In a further technical solution, the radar data is high-dimensional time series data composed of target track and radar cross-section information, and the preprocessing includes data interpolation and extrapolation, data cleaning, and data segmentation.

[0013] In a further technical solution, the data segmentation is specifically to set the length of the sub-time series, and for a time series exceeding this length, it is cut into several sub-time series using a sliding time window of a fixed size.

[0014] In a further technical solution, step b) is specifically to input the sub-time series into the trained neural network to obtain automatically extracted features, where the training method of the neural network is: collecting historical radar data of drones and birds, and using it as a training set after being processed by step a) to train the parameter values of the deep neural network, and the deep neural network is an autoencoder neural network.

[0015] In a further technical solution, step c) specifically includes the following steps:

[0016] c1). Sending the features output in step b) into the trained supervised learning model for binary classification, and outputting the probability of belonging to a drone or a bird;

[0017] c2). Since features have been extracted using a deep neural network, the supervised learning model described in c1 is a traditional binary classification model or a deep learning model;

[0018] c3). Send the features output in step b into the trained unsupervised learning model for anomaly detection to determine whether it is an abnormal sample.

[0019] In a further technical solution, the training method of the supervised learning model is to collect historical radar data of drones and birds. After the processing of steps a and b, label the obtained features, and then use the feature data with labels as the training set to supervised train the binary classification model.

[0020] In a further technical solution, the training method of the unsupervised learning model is to collect historical radar data of drones and birds. After the processing of steps a and b, use the obtained feature data as the training set to unsupervised train the anomaly detection model.

[0021] In a further technical solution, step d includes the following steps:

[0022] d1). When the anomaly detection described in step c determines that the input sample is a normal sample, the final category recognition result is output according to the binary classification result described in step c;

[0023] d2). When the anomaly detection described in step c determines that the input sample is an abnormal sample, if the output probability of the binary classification described in step c is close to 0.5, then the final category recognition result is other low-altitude floating objects, otherwise the final category recognition result is output according to the binary classification result described in step c.

[0024] The present invention also provides a radar automatic recognition device for low-altitude small targets. The device consists of a radar, a memory, a processor, and a communication bus. The communication bus is respectively connected to the radar, the memory, and the processor; the radar is used to detect low-altitude small targets and send the radar data into the memory and the processor through the communication bus; the memory stores a computer program, and the computer program causes the processor to execute the radar automatic recognition method for low-altitude small targets.

[0025] Compared with the prior art, the present invention has the following beneficial effects: Based on the historical radar data of unmanned aerial vehicles (UAVs) and flying birds, the present invention solves the problem of identifying UAVs, flying birds, and other low-altitude floating objects that may appear in practical applications. Since it is difficult or even impossible to collect data on low-altitude floating objects, current relevant patent documents only consider the binary identification of UAVs and flying birds and cannot identify other low-altitude floating objects; the present invention has a wide range of applications and can be used for any radar that can obtain target track and radar cross-section information; the present invention segments radar data using a sliding time window, increasing the amount of training data; the present invention uses a deep neural network to automatically extract features, which has stronger expression and generalization capabilities and more accurate recognition results compared to manually designed features; the present invention does not require time-consuming image recognition, consumes less computing power, and operates faster; the present invention also does not require auxiliary means such as optoelectronic detection and does not add any additional hardware costs during application; the present invention improves the radar's ability to identify small low-altitude targets, making the radar system more intelligent. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flowchart of the radar automatic identification method for small low-altitude targets of the present invention;

[0027] Figure 2 shows the track diagrams of several targets on a certain radar;

[0028] Figure 3 is a structural diagram of the radar automatic identification device for small low-altitude targets of the present invention;

[0029] Figure 4 is a structural diagram of a deep neural network in an embodiment of the present invention;

[0030] Figure 5 is the actual recognition effect diagram of an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Figure 1 The flowchart of the radar automatic identification method for small low-altitude targets of the present invention is shown in, including:

[0033] a) Radar data preprocessing;

[0034] b) Extracting features of the processed data using a deep neural network;

[0035] c) Performing binary classification and anomaly detection using supervised learning and unsupervised learning;

[0036] a) Identifying whether the input data is a UAV, a flying bird, or other low-altitude floating objects according to the results of binary classification and anomaly detection.

[0037] The radar data includes high-dimensional time series data composed of target track and radar cross-section information. Figure 2 Track charts of several targets on a certain radar are given. The track information includes the distance, azimuth, pitch, and speed of the target. One track and the corresponding radar cross-section information constitute a time series. In this embodiment, the normal interval of the time series, i.e., the radar period, is 3 seconds, the dimension is 5, and each dimension is the spatial position coordinate, speed, and radar cross-section information respectively.

[0038] The preprocessing includes data interpolation and extrapolation, data cleaning, and data segmentation steps.

[0039] In this embodiment, when the time interval between two adjacent points in the time series is more than two radar periods, the linear interpolation method is used to add interpolation points; when the time interval between two adjacent points at the end of the time series is zero, the linear extrapolation method is used to fit and extrapolate the values of the next radar period for each dimension of the time series. Through interpolation and extrapolation, the interval of the time series is made 3 seconds.

[0040] In this embodiment, data cleaning is performed to check the consistency of each point in the time series in sequence; that is, the mean and standard deviation of each dimension up to the currently checked point in the time series are statistically calculated. If the values of each dimension at the currently checked point are all outside the interval of the mean plus or minus three standard deviations, the time series is disconnected at this point. For historical radar data, the data cleaning step also deletes time series with a length less than 6 to avoid the interference of clutter and noise; during actual recognition, track segments with a length less than 6 and no next point appearing in three consecutive radar periods are deleted.

[0041] The data segmentation step in the preprocessing refers to setting the length of the sub-time series. For a time series longer than this length, it is cut into several sub-time series using a sliding time window of a fixed size.

[0042] In this embodiment, the length of the sub-time series is set to 6, the size of the sliding time window is 6, and the step size is 2, that is, the sub-time series are sampled with overlap; the significance of a single point in the radar data time series is not great, so analysis is carried out by sampling the sub-time series; at the same time, for a time series with a length of n, [(n + 1 - 6) / 2] + 1 sub-time series can be obtained through overlapping sampling, greatly enhancing the data.

[0043] In this embodiment, during actual recognition, if the target track has accumulated 6 points and has not been deleted by the aforementioned data cleaning, the corresponding time series is used as input data for further preprocessing; if this track has accumulated 2 more points and has not been deleted, the time series corresponding to the 3rd to 8th points is used as the second input data for processing; and so on.

[0044] Using the deep neural network to extract the features of the processed data means inputting the sub-time series into the trained neural network to obtain the automatically extracted features.

[0045] The deep neural network can be an encoder-decoder structure, such as an autoencoder neural network.

[0046] In this embodiment, the autoencoder is a denoising autoencoder with 5 hidden layers, and the dimensionality of the features encoded by the middle hidden layer is 17. Figure 3 The corresponding network structure diagram is given.

[0047] Training the deep neural network means collecting the historical radar data of the unmanned aerial vehicle and the bird, and after being processed by step a, using it as the training set to train the parameter values of the deep neural network.

[0048] In this embodiment, there are 73 pieces of historical radar data of the unmanned aerial vehicle and 580 pieces of birds. After being processed by step a, a training set is obtained, including 5832 samples of the unmanned aerial vehicle and 12416 samples of the bird. The initial parameter values are obtained by using layer-by-layer greedy training and then optimized by the bp algorithm. After all the training is completed, the decoder is discarded, and only the encoder part is retained for feature extraction.

[0049] Performing binary classification and anomaly detection using supervised learning and unsupervised learning includes:

[0050] c1). Feeding the features output in step b into the trained supervised learning model for binary classification, and outputting the probability of belonging to the unmanned aerial vehicle or the bird;

[0051] c2). Since the features have been extracted using the deep neural network, the supervised learning model described in c.1) can be a traditional binary classification model or a deep learning model such as a convolutional neural network;

[0052] c3). Feeding the features output in step b into the trained unsupervised learning model, such as a one-class support vector machine model, for anomaly detection to determine whether it is an abnormal sample.

[0053] In this embodiment, the binary classification model of supervised learning selects a logistic regression model, and the output is the probability that the target belongs to the unmanned aerial vehicle. If the output probability is greater than or equal to 0.5, it is classified as the unmanned aerial vehicle; otherwise, it is classified as the bird.

[0054] In this embodiment, the anomaly detection model of supervised learning selects a one-class support vector machine model.

[0055] The training of the supervised learning model refers to collecting the historical radar data of drones and birds, processing them through step a and step b, labeling the obtained features, and then using the feature data with labels as the training set to train a binary classification model in a supervised manner.

[0056] In this embodiment, 5832 drone samples and 12416 bird samples, after going through step b, obtain 5832 drone feature samples and 12416 bird feature samples. Label the drone feature samples with label 1 and the bird feature samples with label 0, and input them into logistic regression to calculate the parameter values.

[0057] The training of the unsupervised learning model refers to collecting the historical radar data of drones and birds, processing them through step a and step b, using the obtained feature data as the training set, and training an anomaly detection model in an unsupervised manner.

[0058] In this embodiment, the 5832 drone feature samples and 12416 bird feature samples obtained after being processed through step a and step b are all normal samples; send them into a one-class support vector machine, select the RBF kernel function, and train the separation surface.

[0059] The identification of the input data as a drone, a bird, or other low-altitude floating objects according to the binary classification and anomaly detection results includes:

[0060] d1). When the anomaly detection described in step c determines that the input sample is a normal sample, the final category identification result is output according to the binary classification result described in step c;

[0061] d2). When the anomaly detection described in step c determines that the input sample is an abnormal sample, if the output probability of the binary classification described in step c is close to 0.5, then the final category identification result is other low-altitude floating objects, otherwise the final category identification result is output according to the binary classification result described in step c.

[0062] In this embodiment, when the one-class support vector machine determines that the input data is normal data, if the output probability of logistic regression is greater than or equal to 0.5, the input data is identified as a drone, if the output probability of logistic regression is less than 0.5, the input data is identified as a bird; when the one-class support vector machine determines that the input data is abnormal data, if the output probability of logistic regression is between 0.45 and 0.55, the input data is identified as other low-altitude floating objects, if the output probability of logistic regression is greater than 0.55, the input data is identified as a drone, if the output probability of logistic regression is less than 0.45, the input data is identified as a bird.

[0063] It should be noted that the radar period, the total dimension of the time series, the length of the sub-time series, the size and step of the sliding time window, the deep neural network structure, the supervised binary classification model, the unsupervised anomaly detection model, and the threshold range for judging other floating objects can all be determined according to the actual situation, and are not limited to the specific values listed in this embodiment.

[0064] This embodiment also provides a radar automatic recognition device for low-altitude small targets, as Figure 4 shown, which consists of a radar, a memory, a processor, and a communication bus. The communication bus connects the radar, the memory, and the processor. The radar detects low-altitude small targets and sends the radar data to the memory and the processor through the communication bus. The memory stores a computer program, and the computer program causes the processor to execute the radar automatic recognition method for low-altitude small targets.

[0065] Figure 5 The effect diagram during actual recognition of this embodiment is given.

[0066] Although the present invention has been described here with reference to the illustrative embodiments of the present invention, the above embodiments are only the preferred embodiments of the present invention. The embodiments of the present invention are not limited by the above embodiments. It should be understood that those skilled in the art can design many other modifications and embodiments, and these modifications and embodiments will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A radar automatic recognition method for low-altitude small targets, characterized in that, It includes the following steps: a) Radar data preprocessing; b) Using a deep neural network to extract features from the processed data; c) Performing binary classification and anomaly detection using supervised learning and unsupervised learning; Step c) specifically includes the following steps: c1). Feeding the features output in step b into a trained supervised learning model for binary classification and outputting the probability of belonging to a drone or a bird; c2). Since features have been extracted using a deep neural network, the supervised learning model described in c1 is a traditional binary classification model or a deep learning model; c3). Feeding the features output in step b into a trained unsupervised learning model for anomaly detection to determine whether it is an abnormal sample; d) Identifying whether the input data is a drone, a bird, or other low-altitude floating objects based on the binary classification and anomaly detection results; Step d) includes the following steps: d1). When the anomaly detection described in step c determines that the input sample is a normal sample, the final category recognition result is output according to the binary classification result described in step c; d2). When the anomaly detection described in step c determines that the input sample is an abnormal sample, if the binary classification output probability described in step c is between 0.45 and 0.55, then the final category recognition result is other low-altitude floating objects, otherwise the final category recognition result is output according to the binary classification result described in step c.

2. The radar automatic recognition method for low-altitude small targets according to claim 1, characterized in that, The radar data is high-dimensional time series data composed of target track and radar cross-section information, and the preprocessing includes data interpolation and extrapolation, data cleaning, and data segmentation.

3. The radar automatic recognition method for low-altitude small targets according to claim 2, characterized in that, The data segmentation is specifically to set the length of the sub-time series. For a time series longer than this length, it is sliced into several sub-time series using a sliding time window of a fixed size.

4. The radar automatic recognition method for low-altitude small targets according to claim 1, characterized in that, Step b) is specifically to input the sub-time series into a trained neural network to obtain automatically extracted features. The training method of the neural network is: collecting historical radar data of drones and birds, and after being processed through step a), it is used as a training set to train the parameter values of the deep neural network. The deep neural network is an autoencoder neural network.

5. The radar automatic recognition method for low-altitude small targets according to claim 1, characterized in that, The training method of the supervised learning model is to collect historical radar data of drones and birds, process it through steps a and b, label the obtained features, and then use the labeled feature data as a training set to supervised train the binary classification model.

6. The radar automatic recognition method for low-altitude small targets according to claim 1, characterized in that, The training method of the unsupervised learning model is to collect historical radar data of drones and birds, process it through steps a and b, and use the obtained feature data as a training set to unsupervised train the anomaly detection model.

7. A radar automatic recognition device for low-altitude small targets, characterized in that, The device consists of a radar, a memory, a processor, and a communication bus. The communication bus is respectively connected to the radar, the memory, and the processor; the radar is used to detect low-altitude small targets and send the radar data to the memory and the processor through the communication bus; the memory stores a computer program, and the computer program causes the processor to execute the radar automatic recognition method for low-altitude small targets.

Images