An intelligent point-trace aggregation method for radar target detection

Abstract

This invention provides an intelligent point trace aggregation method for radar target detection. The method generates training samples, comprising multiple sample images, each of which is a labeled point trace image. A model is trained using the training samples to generate a training model. Radar data is processed to generate an image to be detected. The image to be detected is input into the training model, and the target position is detected using the trained model to obtain a point trace aggregation result. The method employs a neural network to perform secondary detection on the new "point trace image," eliminating newly emerging clutter and false alarms, and achieving target detection under low-threshold CFAR conditions.

Description

Technical Field

[0001] The invention belongs to the technical field of radar signal processing and relates to an intelligent point trace aggregation method in radar target detection. Background Art

[0002] Track aggregation in radar target detection is a key technology before radar tracking can be implemented. It aggregates adjacent track points detected by CFAR to form a single target track. This effectively reduces the number of duplicate targets, thereby lowering the false alarm rate. Common methods include the centroid method, the geometric center method, and the least squares method, depending on application requirements. Other methods include clustering points based on the closest distance or the furthest distance, depending on the scenario.

[0003] A ground reconnaissance radar target trace aggregation method (Patent Publication No.: CN201911310086.7) proposes a ground reconnaissance radar target trace aggregation method. First, all raw target information within a single coherent integration cycle is extracted, including the range gate number, Doppler signal, and target power. Then, based on the target's distance, point trace aggregation is performed for both close-range and long-range targets, ultimately summarizing the trace aggregation results for the current accumulation cycle. This method aggregates traces based on the range gate number, Doppler signal, and target power, using pre-designed aggregation logic.

[0004] A method for clustering millimeter-wave radar target traces for unmanned vessels (Patent Publication No. CN111175730A) proposes a method for clustering millimeter-wave radar target traces for unmanned vessels. Based on the principle that traces belonging to the same target have similar positions and velocities, the Euclidean distance between traces is calculated, used to measure their similarity. This method clusters target traces in the dimensions of spatial position and velocity, and also uses a pre-designed clustering logic.

[0005] A target aggregation method based on a ground surveillance radar system (Patent Publication No.: CN109100697A) proposes a target aggregation method based on a ground surveillance radar system. This method aggregates target point information in three dimensions (speed, distance, and azimuth) separately. The results of the aggregation in the previous dimension serve as the input for the aggregation in the next dimension. This method aggregates point information in the dimensions of speed, distance, and azimuth, and still uses pre-designed aggregation logic.

[0006] The above patents all use pre-designed models and cannot suppress false alarms. Summary of the Invention

[0007] To solve the problems existing in the prior art, the present invention provides an intelligent point trace aggregation method for radar target detection, comprising the following steps:

[0008] Generate training samples: the training samples include multiple sample images, each sample image is a marked point trace image;

[0009] Use training samples to train the model and generate a training model;

[0010] Process the radar data to generate images to be detected;

[0011] Input the image to be detected into the training model, detect the target position through the training model, and obtain the point trace condensation result;

[0012] The formation process of the dot trace diagram is specifically as follows:

[0013] Obtaining raw radar echo data, and forming a first dot trace diagram after data processing. Each point in the first dot trace diagram is a three-dimensional matrix. Let the matrix be A[M,N,L], where M is the range dimension, representing the range gate; N is the azimuth, which is the radar beam pointing at a certain moment; and L is the Doppler channel.

[0014] Merge N and L into one dimension, that is, expand each column in N dimension into L columns, thus obtaining a quasi-point trace diagram;

[0015] The quasi-dot trace diagrams of the continuously scanned R circles are superimposed to form a dot trace diagram.

[0016] Furthermore, the change process of a certain point when the first dot trace diagram is merged into the quasi-dot trace diagram is specifically as follows:

[0017] Assume that a point in the first dot trace is represented by P(m,n,l), and its mapping to the corresponding point in the quasi-dot trace is represented by Q(x,y), then:

[0018]

[0019] Where ￡ is the size of the L dimension, i.e. the number of Doppler channels;

[0020] Assume that the pixel values of this point are R, G, and B respectively, then R=G=B=(P–MIN)*255 / (MAX-MIN), where MAX=max(P), MIN=min(P), and P represents the signal amplitude at point Q.

[0021] Furthermore, the specific method of generating the image to be detected is the same as the method of generating the dot trace image in the sample training stage.

[0022] Furthermore, R is an integer greater than or equal to 3.

[0023] Furthermore, the detected target position is represented by a rectangular frame, and the midpoint of the rectangular frame is selected as the point trace aggregation result.

[0024] Furthermore, the target condensation detection results are aggregated to form target traces, thereby completing target detection.

[0025] Furthermore, the models trained include YoloV4, RetinaNet, CenterNet and Cascade RCNN, and the data processing includes channel correction, pulse compression, beamforming, moving target display, coherent / incoherent accumulation and constant false alarm processing.

[0026] Furthermore, the marking method is: marking the target according to the collected original radar echo and GPS data of the target, marking the target according to the CFAR detection result, or manually marking the target.

[0027] Compared with the prior art, the present invention has the following technical effects:

[0028] 1. Achieve target detection with low-threshold CFAR: The present invention can achieve target detection with low-threshold CFAR. When the CFAR threshold becomes low, it also means that a large amount of clutter and false alarms appear. However, the present invention uses a neural network to perform secondary detection on the new "dot trace pattern", which can eliminate the newly appeared clutter and false alarms and achieve target detection with low-threshold CFAR.

[0029] 2. Detection of small targets in strong clutter: Small targets are often obscured by strong clutter. This invention utilizes multi-turn radar echo data to enhance the characteristics of small targets and achieve detection of small targets in strong clutter.

[0030] 3. Solve the problems of difficulty in clutter suppression and high false alarm rate in complex scenes: In complex scenes, the distribution of clutter is uncertain. The present invention uses a neural network to learn a large number of effective samples to obtain clutter suppression capabilities, thereby solving the problems of difficulty in clutter suppression and high false alarm rate in complex scenes.

[0031] 4. Realizes multi-dimensional point trace condensation; the present invention detects effective targets at one time through a neural network, and realizes point trace condensation in five dimensions: "distance, Doppler, azimuth, time, and signal amplitude".

[0032] 5. While concentrating on the real target points, it suppresses clutter and false alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of embodiment 1 of the present invention.

[0034] Figure 2 This is a dot trace diagram of the first embodiment of the present invention.

[0035] Figure 3 This is the quasi-point trace diagram of the first embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram of the points that need to be condensed in Example 1 of the present invention.

[0037] Figure 5 This is the quasi-point trace diagram of the small target in Example 1 of the present invention.

[0038] Figure 6 This is a dot trace diagram of a small target according to the first embodiment of the present invention.

[0039] Figure 7 This is a dot plot of the clutter in Example 1 of the present invention.

[0040] Figure 8 This is a schematic diagram of the principle of forming a dot trace diagram according to the first embodiment of the present invention.

[0041] Figure 9 This is a schematic diagram of the dot plot annotation according to the first embodiment of the present invention.

[0042] Figure 10 This is a detection result diagram of the dot trace diagram of the radar scanning ten circles according to the first embodiment of the present invention.

[0043] Figure 11 Schematic diagram of target point traces formed by summarizing the detection results of Example 1 of the present invention.

[0044] Figure 12 This is a detection result diagram of the dot trace diagram of three rounds of radar scanning in Example 1 of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be described in further detail below with reference to the embodiments and accompanying drawings.

[0046] Example 1:

[0047] like Figure 1 As shown, the present invention provides an intelligent point trace aggregation method for radar target detection, and the specific implementation process is divided into two stages: a model training stage and a model use stage.

[0048] 1. Training phase

[0049] Object detection based on deep neural networks is a form of supervised learning. It automatically extracts target features through convolutional neural networks. The larger the data sample, the stronger its generalization ability. Therefore, before using a model, it must be trained with training samples based on detection requirements. After training, the model can be deployed.

[0050] 1) Model selection

[0051] Radar target detection is performed using a deep neural network, either a self-designed network or a publicly available one. This invention differs from the prior art in that it transforms the point trace aggregation problem after constant false alarm processing into a target detection problem using a "point trace map." This invention uses the publicly available YoloV4, RetinaNet, CenterNet, or Cascade RCNN for target detection.

[0052] 2) Training sample generation

[0053] Sample generation includes channel calibration, pulse compression, beamforming, moving target display, coherent / incoherent integration, constant false alarm processing, and dot trace plot construction. These steps are all standard radar signal processing procedures. This embodiment focuses on dot trace plot construction and dot trace aggregation.

[0054] Then, all the traces after constant false alarm processing are integrated with information in five dimensions: "distance, Doppler, azimuth, time, and signal amplitude" to form a trace map.

[0055] Training sample generation, that is, how to plot the point traces after radar constant false alarm processing into an azimuth Doppler / range amplitude diagram, namely a point trace diagram, is the key and core of the present invention.

[0056] Assume that the radar scans one circle. The first trace after constant false alarm processing is defined as A, which is a three-dimensional matrix. Let this matrix be A[M,N,L], where M is the range dimension, representing the range gate; N is the azimuth, which is the direction of the radar beam at a certain moment; and L is the Doppler channel, which is the Doppler channel information obtained through MTD calculation under a beam at a certain azimuth.

[0057] Merge N and L into one dimension. Specifically, expand each column in N dimension into L columns, such as Figure 2 The strips shown are the Doppler channel information at a certain direction. The principle diagram of the formation of the "dot trace diagram" can be seen Figure 8 , direction 0, direction 1... all contain 32 Doppler channels.

[0058] Assume that a point P(m,n,l) in the dot matrix is mapped to a point Q(x,y) on the two-dimensional image. The calculation method is as follows:

[0059] y=m

[0060] x=n*￡+l

[0061] Where ￡ is the size of the L dimension, that is, the number of Doppler channels, and the specific value is 32.

[0062] Let the pixel value of this point be R, G, B,

[0063] MAX=max(P)

[0064] MIN=min(P)

[0065] R=G=B=(P–MIN)*255 / (MAX-MIN), where P represents the signal amplitude at point Q.

[0066] A quasi-point trace diagram containing only the four dimensions of "range, Doppler, azimuth, and signal amplitude" can be seen. Figure 3 This image shows the distribution of dots as the radar scans a circle. The brightness of each dot reflects the strength of the dot signal.

[0067] When the quasi-dot traces of ten consecutive scans are superimposed together, the time dimension is added, thus forming a dot trace to be inspected, such as Figure 2 A typical dot plot can be seen. Figure 2 The horizontal axis represents the azimuth and Doppler channel of the trace, and the vertical axis represents the range gate. The brightness of the trace represents the signal strength. We call this diagram the Azimuth Doppler / Range Amplitude Diagram. Each stripe in the diagram represents a Doppler channel at a different azimuth. In this example, a stripe contains 32 Doppler channels.

[0068] Figure 4 The box in the figure brings together the points to be condensed. As can be seen from the figure, the target's points occupy multiple locations, Doppler channels, and range gates in the point trace diagram. Therefore, the point trace condensation problem becomes a target detection problem in the image domain. The system simply detects the target within the box and uses the center point of the box as the point trace condensation result to complete the target point trace condensation.

[0069] In order to ensure the implementation effect, the purpose of superimposing the information of ten circles of dots in this embodiment is to enhance the detection capability of small targets and increase the distinction between small targets and clutter. Figure 5 For the quasi-point trace map of small targets, Figure 6 It is a "dot trace map" of small targets. Figure 7 This is a clutter dot trace diagram. From the figure, we can see that the small targets in the quasi-dot trace diagram are difficult to see with the naked eye, and are very similar to the clutter dot trace diagram. However, the small targets in the dot trace diagram have more points clustered together, and their characteristics become very obvious.

[0070] However, ten circles are the optimal number of circles under the current conditions. In the actual practice, when the number of circles is greater than or equal to 3, the method described in this embodiment can be applied for detection. The actual application effect when the number of scanning circles is three is as follows: Figure 12 shown.

[0071] After the image is formed, it needs to be annotated. Annotation is to draw the position of the target in the dot trace diagram. The annotation content includes two points: the target position (minx, miny, maxx, maxy) and the target category. To improve the accuracy of annotation, there are usually two methods: A) Annotate the target with reference to GPS data; B) A skilled annotator annotates the target based on long-term experience. The schematic diagram of annotation can be seen Figure 9 .

[0072] 3) Server training

[0073] This example uses the YoloV4 model to train the detection network. The identification category is one (valid target). After the training samples are labeled, the model is trained on an Nvidia V100 for four days. Other common object detection AI models can be used for training, such as RetinaNet, CenterNet, TwoStage's CascadeRCNN, and so on.

[0074] Apply the trained model to detect the target, and the detection results are shown in Figure 10 The box represents the target position, and the center point is the target aggregation result. Figure 11 The target point trace is formed after all the detection results are summarized. The horizontal axis is the direction and the vertical axis is the range gate.

[0075] As can be seen from the figure, this method can reduce the constant false alarm threshold for target detection. After the model training is completed, it can be used for inference in the next stage.

[0076] 2. Application stage

[0077] After the model training and deployment are completed, the model can be applied to implement the detection function. The radar target detection of the present invention is divided into three parts: data preprocessing, inference image generation and point trace formation.

[0078] 1) Data preprocessing

[0079] Data preprocessing primarily involves channel correction, pulse compression, beamforming, moving target display, coherent / incoherent integration, and constant false alarm (CFAR) processing. These operations are consistent with standard radar signal processing. In actual products, this is often accomplished using FPGAs or DSPs. After completion, the data is typically transmitted to subsequent processing modules via RapidIO.

[0080] 2) Reasoning image generation

[0081] The radar trace data after CFAR detection is a three-dimensional matrix. According to the "trace map" generation method in the training phase, an inference image, i.e., the image to be detected, is formed.

[0082] 3) Point condensation

[0083] The system calls the YoloV4 model, and the input data is the "dot trace map" formed by accumulating 10 circles. After the model test, the location of the target concentration is inferred. The "dot trace map" after the inference is completed can be seen Figure 10 .

[0084] This paper proposes an intelligent point trace aggregation method for radar target detection. This method uses artificial intelligence technology to learn from a large number of samples. Under the condition of low constant false alarm processing threshold, it can achieve synchronous point trace aggregation in five dimensions: range, Doppler, azimuth, time, and signal amplitude, while suppressing clutter and false alarms. The present invention has three characteristics:

[0085] 1. The present invention is intelligent condensation. This method can condense effective target points, and does not condense and discard clutter false alarms. It is suitable for low-threshold CFAR detection.

[0086] 2. The agglomerative model of the present invention is not a pre-designed agglomerative logic, but a model obtained by learning a large number of samples.

[0087] 3. The present invention can simultaneously condense radar traces containing five dimensional information: distance, Doppler, azimuth, signal amplitude, and time.

[0088] The present invention can solve the following problems:

[0089] 1. Target detection under low-threshold CFAR: The present invention enables target detection under low-threshold CFAR. When the CFAR threshold decreases, it also means that a large amount of clutter and false alarms appear. However, the present invention uses a neural network to perform secondary detection on the new "dot trace pattern", which can eliminate the newly appeared clutter and false alarms and achieve target detection under low-threshold CFAR.

[0090] 2. Detection of small targets in strong clutter: Small targets are often obscured by strong clutter. This invention uses ten-circle radar echo data to enhance the characteristics of small targets and achieve detection of small targets in strong clutter.

[0091] 3. Solve the problems of difficulty in clutter suppression and high false alarm rate in complex scenes: In complex scenes, the distribution of clutter is uncertain. The present invention uses a neural network to learn a large number of effective samples to obtain clutter suppression capabilities, thereby solving the problems of difficulty in clutter suppression and high false alarm rate in complex scenes.

[0092] 4. Realizes multi-dimensional point trace condensation; the present invention detects effective targets at one time through a neural network, and realizes point trace condensation in five dimensions: "distance, Doppler, azimuth, time, and signal amplitude".

[0093] 5. While concentrating on the real target points, it suppresses clutter and false alarms.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent point trace aggregation method for radar target detection, characterized in that: The following steps are involved: Generate training samples: The training samples include multiple sample images, each sample image is a marked point trace image; the marked content includes the target position; Use the training samples to train the model to obtain the target position of the training samples and generate a training model; The radar data is processed to generate an image to be detected. The specific method for generating the image to be detected is the same as the method for generating the dot trace map in the sample training stage. The image to be detected is input into the training model, and the target position of the image to be detected is detected by the training model to obtain the dot trace condensation result. The formation process of the dot trace diagram is specifically as follows: Obtaining raw radar echo data, and forming a first dot trace diagram after data processing. Each point in the first dot trace diagram is a three-dimensional matrix. Let the matrix be A[M,N,L], where M is the range dimension, representing the range gate; N is the azimuth, which is the radar beam pointing at a certain moment; and L is the Doppler channel. Merge N and L into one dimension, that is, expand each column in N dimension into L columns, thus obtaining a quasi-point trace diagram; The quasi-dot trace diagrams of the continuously scanned R circles are superimposed to form a dot trace diagram.

2. The intelligent point trace aggregation method for radar target detection according to claim 1, characterized in that: The specific change process of a certain point when the first dot trace diagram is merged into the quasi-dot trace diagram is: Assume that a point in the first dot trace is represented by P(m,n,l), and its mapping to the corresponding point in the quasi-dot trace is represented by Q(x,y), then: Where £ is the size of the L dimension, i.e. the number of Doppler channels; Assume that the pixel values of this point are R, G, and B respectively, then R=G=B = (P–MIN)*255 / (MAX-MIN), where MAX = max(P), MIN = min(P), and P represents the signal amplitude at point Q.

3. The intelligent point trace aggregation method for radar target detection according to claim 2, characterized in that: R is an integer greater than or equal to 3.

4. The intelligent point trace aggregation method for radar target detection according to claim 3, characterized in that: The detected target position is represented by a rectangular box, and the midpoint of the rectangular box is selected as the point condensation result.

5. The intelligent point trace aggregation method for radar target detection according to claim 4, characterized in that: The target detection results are aggregated to form target traces, thus completing the target detection.

6. The intelligent point trace aggregation method for radar target detection according to claim 5, characterized in that: The model to be trained is one of YoloV4, RetinaNet, CenterNet and Cascade RCNN, and the data processing includes channel correction, pulse compression, beamforming, moving target display, coherent / incoherent accumulation and constant false alarm processing.

7. The intelligent point trace aggregation method for radar target detection according to claim 6, characterized in that: The marking method is: marking the target according to the collected original radar echo and GPS data of the target, marking the target according to the CFAR detection result, or marking the target manually.

Images