Clutter elimination method and system based on graph neural network and space-time attention
By using graph neural networks and spatiotemporal attention-based methods, the confidence level of radar points is evaluated and clutter points are removed, which solves the problem of insufficient adaptability in existing technologies and achieves high-accuracy clutter removal and target detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-04-03
AI Technical Summary
Existing clutter removal methods lack adaptability and learning capabilities in complex and ever-changing real-world environments, making it difficult to meet the detection requirements for high adaptability and high reliability. This results in a high false alarm rate, affecting target detection and system efficiency.
We employ a graph neural network and spatiotemporal attention-based approach. By constructing a graph neural network and a Transformer module, we extract the dynamic features and correlations of the dots, evaluate the confidence of the dots, and remove clutter dots.
It improves the accuracy of clutter removal, reduces false alarms, enhances target detection capabilities and system efficiency, and reduces the workload of the signal processor.
Smart Images
Figure CN121784692A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar clutter removal technology, and particularly relates to a clutter removal method and system based on graph neural networks and spatiotemporal attention. Background Technology
[0002] Clutter removal is a crucial component of radar data processing, as clutter is a primary source of false alarms. A large number of false alarms can severely disrupt the system, potentially causing the automatic tracking system to malfunction and become desensitized to genuine threats. Effective clutter removal significantly reduces the false alarm rate, ensuring the reliability of alarms. In strong clutter backgrounds, weak target signals can be overwhelmed by clutter; advanced clutter suppression techniques can improve the ability to detect real targets. In radar data processing, excessive clutter generates numerous false, abrupt, and short-lived tracks, severely interfering with the initiation and stable tracking of real target tracks. Clutter removal can also greatly reduce the workload of signal and data processors, allowing valuable computing power to be focused on processing real targets, thus improving overall system efficiency and capacity.
[0003] Clutter removal methods typically employ moving target detection and constant false alarm rate (CFAR) detection. These methods rely on fixed prior models and single features, making them ill-suited for complex and ever-changing real-world environments. They lack adaptability and learning capabilities, and thus fail to meet the requirements for high adaptability and high reliability in detection. Summary of the Invention
[0004] To address the shortcomings of existing clutter removal techniques, the present invention aims to provide a clutter removal method and system based on graph neural networks and spatiotemporal attention. This method utilizes deep neural networks to mine the dynamic features of scanned clutter points, fully and effectively evaluates the confidence level of the clutter points, reduces the number of clutter points, and improves the accuracy of clutter removal.
[0005] The technical solution to achieve the purpose of this invention is as follows:
[0006] A clutter removal method based on graph neural networks and spatiotemporal attention includes:
[0007] Step 1: For all points scanned by the radar, extract all sector points scanned by sliding window, and construct 4D algorithm input data, which are three-dimensional coordinates X, Y, Z and time frame;
[0008] Step 2: Preprocess all scanned points by removing the absolute position information of the point sequence and normalizing the scale of all points;
[0009] Step 3: Construct a neural network model based on graph neural networks and spatiotemporal attention to obtain the confidence score for each point;
[0010] Step 4: Construct a training strategy and train the neural network model;
[0011] Step 5: In the initial stage of the trajectory, input all the current normalized points into the trained neural network model to obtain the classification confidence of each point. Points with a classification confidence higher than the threshold are the target points, and the trajectory is started for the target points.
[0012] Furthermore, the scale normalization is as follows:
[0013]
[0014] in, This represents a normalized video slice. This represents the i-th point after the movement. This represents a normalized video slice.
[0015] Furthermore, the neural network model includes a graph neural network constructed from multiple 3D convolutions and a Transformer module. The graph neural network calculates the neighborhood points of each point, extracts global neighborhood information using max pooling, maps the point trace features to a high-dimensional space, and concatenates the high-dimensional point trace features with the input point trace information, which is then input into the Transformer module. The Transformer module extracts the correlation between each point trace and outputs a confidence estimate for each point.
[0016] Furthermore, the graph neural network includes 3 convolutional layers, 3 batch Norm layers, and 3 Leaky ReLU layers.
[0017] Furthermore, the stride of each convolutional layer is set to a number greater than 1.
[0018] Furthermore, before inputting into the neural network model, a neighborhood graph is constructed with each point as the center, selecting k surrounding points to build its neighborhood graph, and the maximum feature within its neighborhood is calculated and spliced to the feature of that point. This step is repeated 3 times.
[0019] Furthermore, the neural network model is trained using the Adam optimizer.
[0020] Furthermore, the loss function for training the neural network model is:
[0021] ;
[0022] Where N is the number of waypoints. The label represents the current point; 1 indicates the target point, and 0 indicates clutter. This represents the probability that the current point is the target point.
[0023] Furthermore, the threshold value in step 5 is 0.5.
[0024] A clutter removal system based on graph neural networks and spatiotemporal attention includes:
[0025] The data scanning and acquisition unit extracts all sector points from the radar scan using a sliding window, and constructs 4D algorithm input data, which consists of three-dimensional coordinates X, Y, Z and time frame.
[0026] The data preprocessing unit preprocesses all scanned points, removing the absolute position information of the point sequence and normalizing the scale of all points;
[0027] The model building unit constructs a neural network model based on graph neural networks and spatiotemporal attention to obtain the confidence score for each point;
[0028] The training unit constructs the training strategy and trains the neural network model;
[0029] The output unit, during the initial stage of the trajectory, inputs all the current normalized points into the trained neural network model to obtain the classification confidence of each point. Points with a classification confidence higher than the threshold are considered target points, and the trajectory is initiated for the target points.
[0030] Step 1: For all points scanned by the radar, extract all sector points scanned by sliding window, and construct 4D algorithm input data, which are coordinates X, Y, Z and time frame;
[0031] Step 2: Preprocess all scanned points to remove the absolute position information of the point sequence, and use the Min-Max method to normalize the scale of all points to between 0 and 1.
[0032] Step 3: Construct a neural network based on graph neural networks and spatiotemporal attention. The model includes a graph neural network structure for encoding point features using 3D convolutional layers, a graph network structure for interconnecting points, and neighborhood feature extraction to obtain high-dimensional features for each point. Spatiotemporal attention encoding is performed on the points, and the obtained high-dimensional features are input into a Transformer structure to extract the correlations between points. The correlation features of each point are input into a fully connected layer to obtain a confidence score for each point. This score is used to classify the point. During training, skip connections are used to add initial data to the features extracted by the network to avoid catastrophic forgetting.
[0033] Step 4: Construct an algorithm training strategy, initialize model parameters, use a warm-up strategy to gradually increase the learning rate value to a preset peak value in the early stage of training, and then decrease it according to a preset strategy (cosine annealing) to achieve fine-tuning of parameters. To prevent gradient explosion when training long sequences, apply gradient clipping method.
[0034] Step 5: In the initial stage of the track, input all normalized points into the network to obtain the classification of each point. Subsequently, track initiation is only performed on the target point.
[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention performs clutter point confidence assessment at the beginning stage of the track, eliminates clutter points at the beginning stage of the track, and has good applicability; The present invention establishes a graph neural network and a Transformer architecture to jointly analyze the input point information and calculate the correlation between them, effectively extracting the features of real targets and false targets, and can correctly assess the point confidence and reduce clutter points. Attached Figure Description
[0036] Figure 1 This is a schematic diagram of the neural network model structure in the method of the present invention.
[0037] Figure 2 This is a structural diagram of the Transformer model in the method of this invention. Detailed Implementation
[0038] This invention proposes a clutter removal method based on graph neural networks and spatiotemporal attention, referencing... Figure 1 The method includes the following steps:
[0039] Step 1: For all points scanned by the radar, extract all sector points scanned by sliding window, construct 4D algorithm input data, namely coordinates X, Y, Z and time frame. Specifically, it includes: for sector-by-sector scan data, first extract the current scanned sector, then extract the data of the current sector and the two sectors to its left and right from the historical scan frame, construct fixed time window, and fixed sector window input data.
[0040] In this embodiment, for all points scanned by the radar, a sliding window is used to extract all sector points from the scan, constructing a 4D algorithm input point set. ,in Let i be the i-th point among N waypoints. These represent the x-coordinate, y-coordinate, z-coordinate, and time of the point, respectively.
[0041] Step 2: Preprocess all scanned points by removing the absolute position information of the point sequence and normalizing the scale of all points to between 0 and 1.
[0042] This embodiment, for the scanned points, subtracts the coordinates of the center point of all points from the coordinates of all points, and uses the Min-Max method to normalize all point coordinate values to between 0 and 1. This is applied to the acquired set of scanned points. Subtract the coordinates of the center point of the set of track points from the coordinates of each track point:
[0043]
[0044]
[0045]
[0046]
[0047] For the set of points after the movement For each point, the value is normalized to between 0 and 1 using the Min-Max method:
[0048]
[0049] Where min represents the minimum value and max represents the maximum value. This represents a normalized video slice.
[0050] Step 3: Construct a neural network based on graph neural networks and spatiotemporal attention. The model includes a graph neural network structure for encoding point features using 3D convolutional layers, a graph network structure for interconnecting points, and neighborhood feature extraction to obtain high-dimensional features for each point. Spatiotemporal attention encoding is applied to the points, and the obtained high-dimensional features are input into a Transformer structure to extract the correlations between points. The correlation features of each point are input into a fully connected layer to obtain a confidence score for each point. This score is used to classify the points. During training, skip connections are used to add initial data to the features extracted by the network to avoid catastrophic forgetting.
[0051] The construction of the neural network based on graph neural network and spatiotemporal attention in step 3 includes: the neural network model for the point feature encoding is constructed by the point neighborhood calculation module in the Dynamic Graph Convolutional Neural Network (DGCNN) to calculate the neighborhood points of each point, extracting global neighborhood information by max pooling, and constructing a high-dimensional feature encoding module with 3 layers of 3D convolutional layers. The input point information and the output high-dimensional features are concatenated and then input into a continuous 4-layer Transformer encoder. The input point information and the output point features of the Transformer are concatenated and then input into the point classification head to calculate the classification confidence of each point. Points with a confidence score higher than 0.5 are identified as target points, and the rest are clutter points.
[0052] Specifically, constructing neural networks based on graph neural networks and spatiotemporal attention, such as... Figure 1 As shown, the trajectory-video dual-modal neural network model includes: the neural network model for point feature encoding uses a point neighborhood calculation module in a Dynamic Graph Convolutional Neural Network (DGCNN) to calculate the neighborhood points of each point, extracts global neighborhood information using max pooling, and a high-dimensional feature encoding module constructed from three 3D convolutional layers. The input point information and the output high-dimensional features are concatenated and then input into a four-layer Transformer encoder. The input point information and the Transformer output point features are concatenated and then input into a point classification head to calculate the classification confidence of each point. Points with a confidence score higher than 0.5 are considered target points, while the rest are considered clutter points. Specifically, firstly, a neighborhood map is constructed centered on each point, selecting k surrounding points to build its neighborhood map, and calculating the maximum feature within its neighborhood, concatenating it with the feature of that point. This process is repeated three times. Then, a multi-layer 3D convolutional neural network is constructed, including three convolutional layers, three Batch Norm layers, and three Leaky ReLU layers. The stride of each convolutional layer is set to a number greater than 1 to reduce the feature size. This module maps point features to a high-dimensional space. The high-dimensional point features are concatenated with the input point information and fed into the Transformer module. After the Transformer's encoder module outputs the data, it is concatenated with the original input point information and the point information output from the graph neural network. Finally, the Transformer's decoder outputs a confidence estimate for each point. Points with a confidence score higher than 0.5 are considered target points; otherwise, they are considered clutter points. The Transformer module structure is technically and generally known, such as... Figure 2 As shown, it will not be elaborated on here.
[0053] Step 4: Construct the algorithm strategy and initialize the model parameters. The weights of the convolutional layers are randomly initialized using the Xavier method with a bias of 0. Set the number of steps in the initial linear growth phase of the warm-up, the peak learning rate (the highest learning rate after the warm-up ends), and set the cosine annealing parameters, including the initial learning rate and the minimum learning rate.
[0054] The algorithm training strategy in this embodiment specifically includes: initializing model parameters, randomly initializing convolutional layer weights using the Xavier method with a bias of 0, setting the number of steps in the initial linear growth phase of warm-up to 300, setting the peak learning rate (the highest learning rate after warm-up) to 1e-4, and setting cosine annealing parameters, including an initial learning rate of 1e-3 and a minimum learning rate of 1e-6, setting the optimizer to the Adam optimizer, and using the cross-entropy loss function.
[0055]
[0056] in, The label represents the current point; 1 indicates the target point, and 0 indicates clutter. This represents the probability that the current point is the target point.
[0057] Step 5: In the initial stage of the track, input all normalized points into the network to obtain the classification of each point. Subsequently, track initiation is only performed on the target point.
[0058] In the initial stage of the trajectory, all normalized points are input into the network to obtain the classification of each point. Subsequent trajectory initiation is only performed on the target point.
[0059]
[0060] In the initial stage of the track, the track point confidence of each point is calculated through the above network, a point mask with a confidence of less than 0.5 is constructed, and this mask is applied to all points to obtain the track points with a confidence of more than 0.5, which are then input into the subsequent track initiation.
[0061] For purposes of explanation, the foregoing description has been described with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Many modifications and variations are possible in light of the foregoing teachings. The embodiments were chosen and described in order to best explain the principles of the invention and its physical application, thereby enabling those skilled in the art to best utilize the invention and its various embodiments with various modifications as may be suitable for the particular application contemplated.
Claims
1. A clutter removal method based on graph neural networks and spatiotemporal attention, characterized in that, include: Step 1: For all points scanned by the radar, extract all sector points scanned by sliding window, and construct 4D algorithm input data, which are three-dimensional coordinates X, Y, Z and time frame; Step 2: Preprocess all scanned points by removing the absolute position information of the point sequence and normalizing the scale of all points; Step 3: Construct a neural network model based on graph neural networks and spatiotemporal attention to obtain the confidence score for each point; Step 4: Construct a training strategy and train the neural network model; Step 5: In the initial stage of the trajectory, input all the current normalized points into the trained neural network model to obtain the classification confidence of each point. Points with a classification confidence higher than the threshold are the target points, and the trajectory is started for the target points.
2. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 1, characterized in that, The scale normalization is as follows: ； in, This represents a normalized video slice. This represents the i-th point after the movement. This represents a normalized video slice.
3. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 1, characterized in that, The neural network model includes a graph neural network built with multiple 3D convolutions and a Transformer module. The graph neural network calculates the neighborhood points of each point, extracts global neighborhood information through max pooling, maps the point trace features to a high-dimensional space, and concatenates the high-dimensional point trace features with the input point trace information, which is then input into the Transformer module. The Transformer module extracts the correlation between each point trace and outputs a confidence estimate for each point.
4. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 2, characterized in that, The graph neural network includes 3 convolutional layers, 3 batch Norm layers, and 3 Leaky ReLU layers.
5. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 4, characterized in that, The stride of each convolutional layer is set to a number greater than 1.
6. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 5, characterized in that, Before inputting into the neural network model, construct a neighborhood graph centered on each point, selecting k surrounding points to build its neighborhood graph, and calculate the maximum feature within its neighborhood and stitch it into the feature of that point. Repeat this step 3 times.
7. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 3, characterized in that, The neural network model was trained using the Adam optimizer.
8. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 5, characterized in that, The loss function for training the neural network model is: ； Where N is the number of waypoints. The label represents the current point; 1 indicates the target point, and 0 indicates clutter. This represents the probability that the current point is the target point.
9. The clutter removal method based on graph neural networks and spatiotemporal attention according to claim 1, characterized in that, The threshold value in step 5 is 0.
5.
10. A clutter removal system implementing the clutter removal method according to any one of claims 1-9, characterized in that, include: The data scanning and acquisition unit extracts all sector points from the radar scan using a sliding window, and constructs 4D algorithm input data, which consists of three-dimensional coordinates X, Y, Z and time frame. The data preprocessing unit preprocesses all scanned points, removing the absolute position information of the point sequence and normalizing the scale of all points; The model building unit constructs a neural network model based on graph neural networks and spatiotemporal attention to obtain the confidence score for each point; The training unit constructs the training strategy and trains the neural network model; The output unit, during the initial stage of the trajectory, inputs all the current normalized points into the trained neural network model to obtain the classification confidence of each point. Points with a classification confidence higher than the threshold are considered target points, and the trajectory is initiated for the target points.