KR102557458B1 - Method and Apparatus for Detecting Radar Target Based on Artificial Intelligence - Google Patents

Abstract

Disclosed is an AI-based radar target detection method and an apparatus therefor.
A radar target detection device according to an embodiment of the present invention,
a radar antenna for transmitting a radar waveform signal and receiving a reflection signal obtained by reflecting the radar waveform signal on a target; a receiver configured to generate a sum channel signal based on the reflected signal; a Doppler processor generating a distance-Doppler based first signal map by performing Doppler processing based on the sum channel signal; a clutter removal unit to remove clutter from the first signal map by applying a pre-learned neural network model and to generate a second signal map from which the clutter is removed; and a target detection unit for detecting a target using the second signal map.

Description

AI-based radar target detection method and apparatus therefor {Method and Apparatus for Detecting Radar Target Based on Artificial Intelligence}

The present invention relates to a method for performing radar target detection in a range-Doppler signal map of a real environment based on artificial intelligence and an apparatus therefor.

The contents described in this section simply provide background information on the embodiments of the present invention and do not constitute prior art.

Radar is a detection system that uses radio waves to measure the range, angle, and speed of a target. In a radar detection system, it is very important to distinguish the target from the environmental scene.

A radar detection system may use a range-Doppler signal map to detect a target. A method of detecting a target from a range-Doppler signal map generally uses a constant false alarm rate (CFAR) method, and the CFAR method predicts a target when the signal difference between the reflector and the surroundings exceeds a threshold value.

In general, detecting a target from a range-Doppler signal map using artificial intelligence is difficult in two respects.

First, it is difficult to detect a target in a range-Doppler signal map in a radar detection system because the spatial information of the range-Doppler signal map data represents a specific distance and frequency compared to a general image. Target detection in the range-Doppler signal map must be processed in consideration of spatial information in the radar detection system because the two axes of the range-Doppler signal map image are related to distance and frequency.

In order to solve this problem, a network that does not use max-pooling was proposed and the performance of the classification task was improved. performance will deteriorate.

Second, the synthetic range-Doppler signal map used in the radar detection system does not consider the existence of various clutters such as seawater spikes and wind, making it difficult to detect targets in real environments.

Recently, there has been a study on a CNN approach that detects targets from synthetic range-Doppler signal maps. The network of this CNN approach performs pixel-level classification using a sliding window to detect targets in the synthetic RD map.

Fig. 1(a) shows a synthesized distance-Doppler signal map, and Fig. 1(b) shows a distance-Doppler signal map of a real environment. In radar detection systems, synthetic range-Doppler signal map data shows a prominent target distribution in certain noise, but real environment data makes it difficult to detect target distribution from background noise and clutter.

The present invention generates an artificial intelligence-based target detection model through learning to detect a target in a real environment range-Doppler signal map, and removes clutter from a range-Doppler signal map for a real environment based on the target detection model. The main purpose is to provide an artificial intelligence-based radar target detection method and a device for detecting a target by using

According to one aspect of the present invention, a radar target detection device for achieving the above object includes a radar antenna for transmitting a radar waveform signal and receiving a reflection signal in which the radar waveform signal is reflected on a target; a receiver configured to generate a sum channel signal based on the reflected signal; a Doppler processor generating a distance-Doppler based first signal map by performing Doppler processing based on the sum channel signal; a clutter removal unit to remove clutter from the first signal map by applying a pre-learned neural network model and to generate a second signal map from which the clutter is removed; and a target detection unit for detecting a target using the second signal map.

In addition, according to another aspect of the present invention, a radar target detection method for achieving the above object includes a radar transmitting and receiving step of transmitting a radar waveform signal and receiving a reflection signal in which the radar waveform signal is reflected on the target; a receiving step of generating a sum channel signal based on the reflected signal; a Doppler processing step of generating a distance-Doppler based first signal map by performing Doppler processing based on the sum channel signal; a clutter removal step of removing clutter from the first signal map by applying a pre-learned neural network model and generating a second signal map from which the clutter is removed; and a target detection step of detecting a target using the second signal map.

As described above, the present invention has an effect of overcoming the limitations of feature extraction methods based on heuristics algorithms used in existing machine learning methods.

In addition, the present invention has an effect that it is possible to develop a high-capacity model using high computing power using a GPU, and to obtain generality of the model through the use of a vast dataset.

In addition, the present invention can improve target detection capability compared to conventional methods, minimize false detection of targets due to a dramatic increase in clutter removal performance, and enable detection of targets with a small radar cross section (RCS). It works.

1 is a diagram showing general synthesis-based radar data and actual acquired radar data.
2 is a schematic block diagram of a radar target detection device according to an embodiment of the present invention.
3 is a schematic block diagram of a clutter processing unit of a radar target detection device according to an embodiment of the present invention.
4 is a diagram illustrating a target detection model for clutter processing according to an embodiment of the present invention.
5 is a flowchart illustrating a learning model processing operation for AI-based radar target detection according to an embodiment of the present invention.
6 is a flowchart illustrating a method for detecting a radar target based on artificial intelligence according to an embodiment of the present invention.
7 to 9 are exemplary diagrams illustrating target detection experiments and test results of a radar target detection apparatus according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description will be omitted. In addition, although preferred embodiments of the present invention will be described below, the technical idea of the present invention is not limited or limited thereto and can be modified and implemented in various ways by those skilled in the art. Hereinafter, an AI-based radar target detection method and an apparatus therefor proposed in the present invention will be described in detail with reference to drawings.

2 is a schematic block diagram of a radar target detection device according to an embodiment of the present invention.

The radar target detection apparatus 200 according to the present embodiment includes a radar antenna 210, a receiver 220, a Doppler processing unit 230, a clutter removal unit 240, a target detection unit 250, and a learning model processing unit 260. ). The radar target detection device 200 of FIG. 2 is according to an embodiment, and all blocks shown in FIG. 2 are not essential components, and in another embodiment, some blocks included in the radar target detection device 200 are added. , may be changed or deleted.

The radar target detection apparatus 200 performs an operation of detecting a target in a range-Doppler signal map of a real environment using an AI-based target detection model. Here, the target detection model is preferably composed of a deep neural network (DNN), but is not necessarily limited thereto. Hereinafter, each component of the radar target detection device 200 will be described.

The radar antenna 210 transmits (radiates) a radar waveform signal for target detection and receives a reflected signal obtained by reflecting the radar waveform signal from targets. A radar waveform signal, a reflected signal, etc. transmitted and received by the radar antenna 210 may be a Radio Frequency (RF) signal, and the radar antenna 210 may be implemented with various types of antennas.

The receiving unit 220 generates a sum channel signal based on the reflected signal received based on the radar waveform.

The receiver 220 may generate a sum channel signal through the received plurality of channel signals.

The receiver 220 according to an embodiment of the present invention may apply weights to a plurality of channel signals received by the radar antenna 210 and sum them up. For example, the receiver 220 may include a plurality of hybrid couplers, and may generate a sum channel signal obtained by applying weights to a plurality of channel signals using the hybrid couplers and summing the plurality of channel signals.

The Doppler processing unit 230 performs Doppler processing based on the sum channel signal to generate a distance-Doppler based first signal map.

The Doppler processing unit 230 generates a first signal map, which is a distance-Doppler signal map for a real environment, by performing a Doppler direction Fast Fourier Transform (FFT) on the sum channel signal in which the specified number of pulse signals are collected. The first signal map may include a plurality of clutter and background noise of a real environment.

The clutter removal unit 240 removes clutter from the first signal map by applying the learned neural network model and generates a second signal map from which the clutter is removed. Here, the pre-trained neural network model may be a target detection model, and may be a target detection model pre-learned in the learning model processing unit 260 .

The clutter remover 240 receives the first signal map, performs downsampling, classifies objects related to the target within the first signal map, and generates first feature data.

In addition, the clutter remover 240 may generate target feature compressed data by filtering at least one feature data generated in the process of generating the first feature data.

The clutter remover 240 receives the first feature data and the compressed target feature data, performs upsampling, restores the size of the first signal map including objects related to the target, and generates a second signal map.

The target detection unit 250 performs an operation of detecting a target using the second signal map.

The target detection unit 250 displays target candidates from which clutter is removed on the radar map based on the second signal map, and finally detects the target. Here, the target detection unit 250 preferably detects the target using CA-CFAR (Cell Average-Constant False Alarm Rate) method, but is not necessarily limited thereto.

The learning model processing unit 260 performs an operation of generating a target detection model (neural network model) by learning a neural network model for clutter removal.

The learning model processing unit 260 generates a range-Doppler signal map generated based on the received radar signal obtained through the field test as learning input data.

The learning model processing unit 260 inputs learning input data into a preset learning model and repeatedly performs a step of labeling target data and clutter data.

The learning model processing unit 260 generates a target detection model (neural network model) based on the determined weight matrix by repeatedly performing the step of labeling the target data and the clutter data.

The learning model processor 260 transmits the generated target detection model (neural network model) to the clutter remover 240 to generate a second signal map from which clutters included in the first signal map are removed.

3 is a schematic block diagram of a clutter processing unit of a radar target detection device according to an embodiment of the present invention.

The clutter removal unit 240 removes clutter from the first signal map by applying the learned neural network model and generates a second signal map from which the clutter is removed. The clutter removal unit 240 according to the present embodiment includes an encoder 242, a target feature compression unit 244, and a decoder 246.

The encoder 242 receives the first signal map, performs downsampling, classifies objects related to the target within the first signal map, and generates first feature data. The encoder 242 may use the first signal map having a size of 32 × 32 as an input.

The encoder 242 according to this embodiment may be composed of three convolution layers, a max pooling layer, and a pulley connect layer.

The target feature compression unit 244 may generate target feature compressed data by filtering at least one feature data generated in the process of generating the first feature data.

The target feature compression unit 244 filters outputs of the first and second convolution layers of the encoder 242 to generate first target feature compressed data and second target feature compressed data.

On the other hand, the target feature compression unit 244 filters the output of each of the first and second convolution layers of the encoder 242, and sets different weights (first weight and second weight) according to the order of the convolution layers. ) is applied to each of the filtered results to generate first target feature compressed data and second target feature compressed data. Here, the weights (first weight and second weight) set differently may be different values previously set by the user, but are not necessarily limited thereto, and are automatically mutually related based on the number of clutters included in the first signal map. It may be a different set value. In addition, the target feature compression unit 244 calculates a first weight based on the number of clutters included in the first signal map, and automatically calculates a second weight by applying a preset arithmetic process to the calculated first weight. can also be calculated.

The first target feature compression data generated by the target feature compression unit 244 is added to the second deconvolution layer of the decoder 246 . In addition, the second target feature compression data generated through the target feature compression unit 244 is added to the first deconvolution layer of the decoder 246 .

The decoder 246 receives the first feature data and the compressed target feature data, performs upsampling, restores the size of the first signal map including objects related to the target, and generates a second signal map.

4 is a diagram illustrating a target detection model for clutter processing according to an embodiment of the present invention.

The radar target detection apparatus 200 according to this embodiment uses a target detection model (neural network model) to detect a target in a range-Doppler map for a real environment.

This target detection model is a segmentation network for target detection that can make full use of spatial information, and the range-Doppler signal map data for the real environment makes the target detection model robust.

Since the spatial information of the range-Doppler signal map data indicates a specific distance and frequency, the radar target detection device 200 needs to consider the spatial information. Therefore, in the radar target detection apparatus 200, it is preferable to use a segmentation network that acquires the boundary of an object at the pixel level rather than a detection method that obtains a bounding box around the object. Here, the target detection model may be a U-Net based segmentation network.

The target detection model of the radar target detection apparatus 200 according to the present embodiment has computational efficiency and simultaneously calculates outputs of several layers to improve target positioning and situational awareness.

A real environment distance-Doppler signal map has more clutter and background noise than a synthetic distance-Doppler signal map. In other words, the synthetic range-Doppler signal map contains the target signal that stands out from the specific noise, but the real environment range-Doppler signal map contains the target signal together with many clutters such as wind, seawater spikes, background noise, etc. there is. Therefore, it is reasonable for the radar target detection apparatus 200 according to the present embodiment to use a data set for a real environment for robust target detection.

In other words, the radar target detection apparatus 200 according to the present embodiment uses a target detection model, which is a computationally efficient architecture that considers spatial information in a real environment range-Doppler signal map, and uses a real environment distance where clutter and background noise exist. - A robust target detection model can be implemented by constructing a data set based on a Doppler signal map. In addition, the radar target detection apparatus 200 may improve target detection accuracy and recall performance compared to conventional target detection methods (eg, CFAR techniques).

The real environment range-Doppler signal map used for target detection in the radar target detection apparatus 200 is a two-dimensional matrix including relative distance and power information according to Doppler bins. For example, in a real environment distance-Doppler signal map, the distance axis contains information on targets ranging from about 4 km to 35 km at 12.5 m intervals via radar, and the Doppler bin axis contains 32 DFT (Discrete Fourier Transform) components. includes

The radar target detection apparatus 200 divides each range-Doppler signal map into 32 x 32 patches in order to easily process the real environment range-Doppler signal map in the target detection model (neural network).

As shown in (b) of FIG. 1 , in the patch of the real environment distance-Doppler signal map, there is a unique distribution for the target around the target, and the remaining part is clutter and background data. In a real environment distance-Doppler signal map, it is difficult to distinguish the clutter from the target because the clutter varies in size and magnitude. thus,

The radar target detection apparatus 200 may learn the distribution of targets using the target detection model, and may use a real environment range-Doppler signal map patch as initial input data of the target detection model.

Hereinafter, the structure of the target detection model of the radar target detection device 200 will be described. Here, the operation of the target detection model may be performed in the clutter removal unit 240 and the learning model processing unit 260 of the radar target detection device 200, but is not necessarily limited thereto.

In the target detection model, a target classification task is performed to obtain unique information about the target distribution in the real environment distance-Doppler signal map.

The target detection model, which is a segmentation model, uses a target classification network as an encoder to obtain target specific information. The target classification network can classify whether a patch has a target or not. [Table 1] is

It represents the configuration of the target classification network, i.e. the encoder, of the real-environment distance-Doppler signal map.

The encoder takes a 32 × 32 RD map patch as input and consists of 3 convolution layers, a max pooling layer and a pulley connect layer. In the encoder, batch normalization and Rectified Linear Unit (ReLU) are applied after every convolutional layer. In addition, only ReLU is applied after the pulley connect layer.

The target detection model according to this embodiment means a pixel-level semantic segmentation network for a real environment range-Doppler signal map.

Referring to FIG. 4, the basic structure of the target detection model is an encoder-decoder network model. In the encoder 242 part, high-level feature components can be efficiently obtained using a classification network.

The orange blocks in the classification module in encoder 242 represent downsampling due to the max pooling layer. Encoders can be pre-trained by classification tasks.

In the decoder 244 part, green blocks represent upsampling due to the deconvolution layer. At each deconvolution layer, the output size doubles to the initial input size. ReLU and batch normalization are applied after the deconvolution layer.

Spatial information of low-level features in the processing of the encoder and decoder may be preserved by the target feature compression unit 246 . In order to preserve the spatial information of the features, the target feature compression unit 246 first and second outputs of a decoder such as U-Net where the outputs of the first and second convolutional layers of the encoder 242 are represented by purple blocks. to be added element by element.

Finally, the target detection model segments the target in patches of the 32 × 32 real-environment distance-Doppler signal map.

The target detection model uses a ground truth map to segment the target. The ground truth map of the target detection model contains labels for targets of size 3 × 3 that contain the distribution of target signals.

Since there is more background data than target data in the patches of the real-environment distance-Doppler signal map, it is necessary to balance the positive and negative labels. In order to mitigate this imbalanced data distribution of positive and negative labels, the weighted binary cross entropy loss function may be used. This loss function can be defined as [Equation 1].

In the target detection model according to the present embodiment, the weighted binary cross-entropy loss function (Loss) assigns more weight to a loss due to false negatives (FN) using a hyper parameter λ. σ(x) means the sigmoid function of the output x of the network (neural network model), and y means the label of the ground truth map.

5 is a flowchart illustrating a learning model processing operation for AI-based radar target detection according to an embodiment of the present invention.

The radar target detection apparatus 200 performs an operation of generating a target detection model (neural network model) by learning a neural network model for clutter removal.

The radar target detection apparatus 200 obtains a radar received signal through a field test (S510), and generates a range-Doppler signal map based on the acquired radar received signal as learning input data (S520).

The radar target detection apparatus 200 inputs learning input data into a preset learning model, labels target data and clutter data (S530), and repeatedly performs an operation of labeling the target data and clutter data (S540). .

The radar target detection apparatus 200 repeatedly performs the step of labeling target data and clutter data to generate a target detection model (neural network model) based on the determined weight matrix (S550). Here, the target detection model is preferably a learning model based on a deep neural network (DNN).

6 is a flowchart illustrating a method for detecting a radar target based on artificial intelligence according to an embodiment of the present invention.

The radar target detection apparatus 200 transmits (radiates) a radar waveform signal for target detection and receives a reflection signal obtained by reflecting the radar waveform signal from targets (S610).

The radar target detection apparatus 200 generates a sum channel signal based on the reflected signal received based on the radar waveform (S620).

The radar target detection apparatus 200 generates a first signal map based on distance-Doppler by performing Doppler processing based on the sum channel signal (S630 and S632). The radar target detection apparatus 200 generates a first signal map, which is a range-Doppler signal map for a real environment, by performing a Doppler direction fast Fourier transform (FFT) on a sum-channel signal obtained by collecting a specified number of pulse signals. Here, the first signal map may include a plurality of clutter and background noise of the real environment.

The radar target detection apparatus 200 removes clutter from the first signal map by applying a pre-learned neural network model, and generates a second signal map from which the clutter is removed.

The radar target detection apparatus 200 receives the first signal map, performs downsampling, classifies objects related to the target within the first signal map, and generates first feature data (S640). The radar target detection apparatus 200 may generate compressed target feature data by filtering at least one feature data generated in the process of generating the first feature data (S642). The radar target detection apparatus 200 receives the first feature data and the compressed target feature data, performs upsampling, restores the size of the first signal map including objects related to the target, and generates a second signal map (S644). , S650).

The radar target detection apparatus 200 applies the CA-CFAR scheme to display target candidates from which clutter is removed on the radar map based on the second signal map (S660), and finally detects the target (S670). Here, the finally detected target may be a moving target, but is not necessarily limited thereto.

In each of FIGS. 5 and 6, each step is described as sequentially executed, but is not necessarily limited thereto. In other words, each of FIGS. 5 and 6 is not limited to a time-series sequence because it will be applicable to changing and executing the steps described in each of FIGS. 5 and 6 or executing one or more steps in parallel.

The radar target detection method according to the present embodiment described in FIGS. 5 and 6 may be implemented as an application (or program) and recorded on a recording medium readable by a terminal device (or computer). All kinds of recording devices in which an application (or program) for implementing the radar target detection method according to the present embodiment is recorded and a recording medium readable by a terminal device (or computer) stores data that can be read by a computing system. or medium.

7 to 9 are exemplary diagrams illustrating target detection experiments and test results of a radar target detection apparatus according to an embodiment of the present invention.

< Target detection experiment: data set >

A real environment range-Doppler signal map was generated by radar on a hill with views of both the coast and the sea, with targets such as ships and moving objects mixed with noise and clutter from the surrounding environment.

The classification data set includes both targeted and untargeted patches. The training set is about 60,000 patches, the validation set is about 12,000 patches, and the test set is about 12,000 patches.

The split data set contains only patches with targets. In the ground truth map for training the target detection model, 0 labels are included in pixels without target information, and 1 labels are included in pixels with target information. The training set is about 30,000 patches, the validation set is about 6,000 patches, and the test set is about 6,000 patches including all ground truth maps.

<Target Detection Experiment: Training>

In order to train the target detection model, the classification network is first trained.

A binary cross-entropy loss function using L2 regularization is applied to the classification network. The pre-trained model of the classification network is used for the encoder of the target detection model.

Since ReLU activation is non-linear activation, the decoder of the target detection model uses Kaiming He initialization. Only the last layer, which does not contain ReLU activations, uses Xavier Uniform initialization for weight initialization and is optimized with stochastic gradient descent. The training time of the target detection model was about 7 hours with a unit equipped with a GTX TITAN X with about 12 GB of memory.

<Target Detection Experiment: Experimental Results>

Ambiguity resolution is commonly used to resolve the similarity between two signals. In radar signal processing, radar detects reflectors in periodic pulse modulation. Since pulse modulation is a periodic signal, it creates a periodic reflection signal at every period. Thus, there is an ambiguity of not knowing how many pulses have been transmitted before the received signal. A common solution to resolve this ambiguity is to use multiple Pulse Repetition Frequencies (PRFs).

7 is a diagram for explaining how ambiguity is resolved. Each of the two pulse modulations has PRF f1 and PRF f2. The reflected signal appears periodically by pulse modulation. Referring to FIG. 7 , the two reflector signals overlap at point A. Point A can be deduced from the actual distance from the radar. Using this principle, four types of PRF range-Doppler signal maps are used to resolve the ambiguity of the detected target. The actual distance of the target obtained from the ambiguity resolution can be displayed on the radar map according to the azimuth taken from the radar.

8 shows a target with actual distance displayed and a radar map.

The blue dot in (a) of FIG. 8 represents the result (target candidate) detected by CFAR, and the red dot in (b) of FIG. 8 represents the result (target candidate) detected through the target detection model according to the present embodiment. indicate The green circle represents the actual target position. Several filter and cluster methods can additionally be used to obtain the actual unambiguous target location. Therefore, it is better if the target candidates of the CFAR and target detection models are close to the actual targets.

[Table 2] shows the precision and recall of the target detection model and CFAR based on real targets.

Referring to [Table 2], it can be confirmed that the target detection model outperforms the CFAR in terms of recall and precision. When λ is 8/3, the target detection model yields about 4 times better detection results than CFAR in terms of precision. Since the target detection model according to the present embodiment has learned the distribution of targets in a real environment, it can better exclude clutter and background noise from the target.

Meanwhile, a gradient-weighted CAM (GradCam) may store a gradient of a weight in a specific layer and indicate an effect of the gradient on an actual input image as a heat map. This allows us to see how the network learns the distribution of real targets.

9 shows a heat map for the last convolutional layer of the target classification network. 9(a) shows a grayscale image of a distance-Doppler map patch. In the range-Doppler map patch, the point with the greatest power is shown in blue, and the point where the actual target is located is shown in red. Figure 9(b) shows that the network does not simply focus on the maximum value of the range-Doppler map patch, but focuses on the distribution of the target.

The above description is only illustrative of the technical idea of the embodiment of the present invention, and those skilled in the art to which the embodiment of the present invention pertains may make various modifications and modifications within the scope not departing from the essential characteristics of the embodiment of the present invention. transformation will be possible. Therefore, the embodiments of the present invention are not intended to limit the technical idea of the embodiment of the present invention, but to explain, and the scope of the technical idea of the embodiment of the present invention is not limited by these examples. The protection scope of the embodiments of the present invention should be construed according to the claims below, and all technical ideas within the equivalent range should be construed as being included in the scope of the embodiments of the present invention.

200: target detection device
210: radar antenna 220: receiver
230: Doppler processing unit 240: Clutter removal unit
250: target detection unit 260: learning model processing unit

Claims (12)

a radar antenna for transmitting a radar waveform signal and receiving a reflection signal obtained by reflecting the radar waveform signal on a target;
a receiver configured to generate a sum channel signal based on the reflected signal;
a Doppler processor generating a distance-Doppler based first signal map by performing Doppler processing based on the sum channel signal;
a clutter removal unit to remove clutter from the first signal map by applying a pre-learned neural network model and to generate a second signal map from which the clutter is removed; and
A target detection unit for detecting a target using the second signal map,
The clutter removal unit may include an encoder configured to generate first feature data by receiving the first signal map and performing down-sampling to classify an object related to a target within the first signal map; a target feature compression unit generating target feature compressed data by filtering at least one feature data generated in the process of generating the first feature data; and a decoder configured to receive the first feature data and compressed target feature data, perform upsampling, restore a second signal map including an object related to the target, and generate the restored second signal map. Characteristic radar target detection device. According to claim 1,
A radar target detection device, characterized in that it further comprises a learning model processing unit for learning a neural network model for clutter removal and generating the neural network model. According to claim 2,
The learning model processing unit,
Create a distance-Doppler signal map generated based on the radar received signal obtained through the field test as learning input data,
The step of labeling target data and clutter data by inputting the learning input data into a preset learning model is repeatedly performed, and the neural network model is generated based on the weight matrix determined through the step. detection device. delete According to claim 1,
The target detection unit,
Based on the second signal map, a target candidate from which clutter is removed is displayed on a radar map, and the target is finally detected,
Wherein the target detection unit detects the target using a CA-CFAR (Cell Average-Constant False Alarm Rate) method. According to claim 1,
The encoder,
A radar target detection device characterized in that it uses a first signal map of 32 × 32 size as an input and is composed of three convolution layers, a maximum pooling layer, and a fully connected layer. According to claim 6,
The target feature compression unit,
Generating first target feature compressed data and second target feature compressed data by filtering the output of each of the first and second convolution layers of the encoder;
The first target feature compression data is added to a second deconvolution layer of the decoder, and the second target feature compression data is added to a first deconvolution layer of the decoder. . A method for detecting a target in a radar target detection device,
a radar transmission/reception step of transmitting a radar waveform signal and receiving a reflection signal obtained by reflecting the radar waveform signal on a target;
a receiving step of generating a sum channel signal based on the reflected signal;
a Doppler processing step of generating a distance-Doppler based first signal map by performing Doppler processing based on the sum channel signal;
a clutter removal step of removing clutter from the first signal map by applying a pre-learned neural network model and generating a second signal map from which the clutter is removed; and
A target detection step of detecting a target using the second signal map,
The clutter removal step may include an encoding step of generating first feature data by receiving the first signal map and performing down-sampling to classify an object related to a target within the first signal map; a target feature compression step of filtering at least one feature data generated in the process of generating the first feature data to generate target feature compressed data; and a decoding step of receiving the first feature data and the compressed target feature data, performing upsampling, restoring a second signal map including an object related to the target, and generating a restored second signal map. A radar target detection method, characterized in that. delete According to claim 8,
The target detection step,
Based on the second signal map, a target candidate from which clutter is removed is displayed on a radar map, and the target is finally detected,
Wherein the target detection step detects the target using a CA-CFAR (Cell Average-Constant False Alarm Rate) method. According to claim 8,
The encoding step is
A radar target detection method characterized by using a first signal map having a size of 32 × 32 as an input and consisting of three convolution layers, a maximum pooling layer, and a fully connected layer. According to claim 11,
The target feature compression step,
Filtering the output of each of the first and second convolution layers in the encoding step to generate first target feature compressed data and second target feature compressed data;
The first target feature compression data is added to a second deconvolution layer in the decoding step, and the second target feature compression data is added to a first deconvolution layer in the decoding step. detection method.

Images