A deep learning-based anti-interference millimeter wave radar dynamic target detection system
By preprocessing and feature extraction of radar signals, combined with Cholesky decomposition and Transformer network, the problems of clutter interference and long time-series dependence in radar signal analysis are solved, and high-precision dynamic target detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-16
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies struggle to effectively separate weak dynamic signals from strong static clutter in radar signal analysis. They lack a physical understanding of the radar sensing process, suffer from improper multimodal feature fusion, and fail to capture long-term time-series dependencies, resulting in insufficient detection accuracy and reliability in complex environments.
A preprocessing module is used to remove static clutter and perform phase correction. A feature extraction module separates and processes amplitude and phase information. Feature fusion and long-term modeling are performed through Cholesky decomposition prediction network and Transformer network. A dual-stream Transformer fusion architecture is constructed for decision-making.
It significantly improves the accuracy and stability of dynamic target detection in complex environments, and can accurately detect target motion in strong clutter and low signal-to-noise ratio scenarios. It also has physical interpretability and efficient long-term dependency modeling capabilities.
Smart Images
Figure CN121522580B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar signal processing and deep learning fusion technology, and more specifically to a deep learning-based anti-jamming millimeter-wave radar dynamic target detection system. Background Technology
[0002] In recent years, non-contact sensing technologies, represented by millimeter-wave radar, have seen increasingly widespread application in fields such as smart homes, security monitoring, healthcare, and human-computer interaction due to their unique advantages, including privacy protection, resistance to ambient light interference, and the ability to penetrate non-metallic obstructions. One of the core tasks of these applications is to accurately detect, identify, and analyze the dynamic information of targets from the continuous data stream of radar echoes.
[0003] Despite the progress made in radar signal analysis, both traditional signal processing methods and general deep learning models reveal a series of deep-seated and interconnected problems when faced with complex, variable, and low signal-to-noise ratio real-world environments. These problems collectively constitute the core bottleneck of current technological development.
[0004] Question 1: The overwhelming interference of static clutter and the distortion of phase information
[0005] In real-world applications, the effective signals from a target (such as the faint rise and fall of a human chest cavity) are extremely weak, often having energy several orders of magnitude lower than the reflected signals from stationary objects like walls and furniture (i.e., "static clutter"). Existing methods struggle to effectively separate these two types of signals.
[0006] Question 2: The problem of "blind" fusion of multimodal features
[0007] Radar signals naturally contain information in two modes: amplitude and phase, which carry different physical meanings. Amplitude reflects the signal strength (and can serve as a proxy for the signal-to-noise ratio), while phase contains precise displacement information of the target. Existing deep learning models typically use a simple channel concatenation method when fusing these two types of information.
[0008] Question 3: The lack of prior physical knowledge and the "black box" problem of the model
[0009] General-purpose deep learning models (such as CNN and RNN) are purely data-driven "black boxes" that lack an intrinsic understanding of the physical laws behind radar perception processes.
[0010] Question 4: Insufficient ability to capture long-term global dependencies
[0011] The target's behavioral patterns, whether periodic physiological activities (breathing, heartbeat) or non-periodic complex movements (rolling over, sitting up), unfold on a relatively long timescale.
[0012] In summary, these problems are not isolated but intertwined, collectively forming a significant technological barrier that hinders the performance improvement and reliable application of radar sensing technology in complex real-world scenarios. Therefore, there is an urgent need in this field for a novel technological paradigm capable of intelligently fusing multimodal information, deeply embedding physical priors, efficiently capturing global temporal dependencies, and making reliable decisions. Summary of the Invention
[0013] In view of the above problems, the present invention is proposed to provide a deep learning-based anti-jamming millimeter-wave radar dynamic target detection system that overcomes or at least partially solves the above problems.
[0014] To achieve the above objectives, the present invention adopts the following technical solution:
[0015] In a first aspect, embodiments of the present invention provide a deep learning-based anti-jamming millimeter-wave radar dynamic target detection system, comprising:
[0016] Preprocessing module: Used to perform high-fidelity preprocessing on the raw millimeter-wave radar signal, remove static clutter, complete phase correction, and output a high signal-to-noise ratio dynamic signal after bilateral filtering and noise reduction;
[0017] Feature extraction module: used to perform front-end feature extraction on the high signal-to-noise ratio dynamic signal, by separating amplitude and phase information, generating in parallel the modulated phase stream and the feature stream jointly characterized by covariance and trajectory, and outputting two heterogeneous feature streams;
[0018] Long-term global dependency modeling module: used to perform feature fusion on the two heterogeneous feature streams, and fuse them into a unified fused feature map by adding them element by element; input the fused feature map into the Transformer network, and output dynamic target detection results.
[0019] Preferably, the preprocessing module includes:
[0020] Static clutter filtering unit: used to synthesize the amplitude and phase components of the original millimeter-wave radar signal into a complex matrix in the complex domain, calculate the average value of the complex matrix along the time dimension, and subtract the average value from the complex matrix to obtain a zero-mean complex signal containing dynamic target information;
[0021] Phase correction unit: used to extract the phase component from the zero-mean complex signal, and sequentially perform phase unwinding and linear detrending processing on the phase component to obtain a continuous and trend-free corrected phase signal;
[0022] Bilateral filtering unit: used to arrange the corrected phase signal into a two-dimensional spatiotemporal matrix, and to perform bilateral filtering and noise reduction on the two-dimensional spatiotemporal matrix through a two-dimensional bilateral filter to output the high signal-to-noise ratio dynamic signal.
[0023] Preferably, the bilateral filtering denoising includes:
[0024] S101: Traverse each central pixel in the two-dimensional spatiotemporal matrix;
[0025] S102: For each central pixel, select all neighboring pixels within its preset neighborhood range;
[0026] S103: Calculate the spatial distance weight and pixel value similarity weight between the neighboring pixels and the center pixel respectively;
[0027] S104: Multiply the spatial distance weight and the pixel value similarity weight element by element to obtain the combined weight corresponding to each neighboring pixel.
[0028] S105: Calculate the new denoised value of the center pixel by using the combined weights of all neighboring pixels to calculate the weighted average of the corresponding pixel values.
[0029] S106: Determine whether all the center pixels have been processed. If not, return to S102 to continue processing the next center pixel. If yes, output the high signal-to-noise ratio dynamic signal.
[0030] Preferably, the feature extraction module includes:
[0031] Signal normalization and separation unit: used to normalize the high signal-to-noise ratio dynamic signal and decompose it into amplitude information and phase information;
[0032] Modulation unit: used to perform amplitude gating processing on the amplitude information and phase modulation on the phase information to generate a modulated phase stream;
[0033] Covariance processing unit: performs inter-channel fusion and spatial average pooling on the phase signal, calculates the covariance matrix graph, extracts features from the covariance matrix graph to obtain a new feature map, upsamples the new feature map and performs TCN trajectory prediction to obtain the feature flow jointly represented by the covariance and trajectory.
[0034] Preferably, the specific processing procedure of the modulation unit is as follows:
[0035] The amplitude information is passed through three parallel convolutional layers with 3x3, 5x5, and 7x7 kernels to extract features at different scales;
[0036] After concatenating features of different scales along the channel dimension, batch normalization and GELU activation function are applied to obtain a multi-scale amplitude feature map.
[0037] The multi-scale amplitude feature map is fed into the amplitude gating generator, and 3x3 two-dimensional convolution, batch normalization, SiLU activation function, 1x1 two-dimensional convolution and Sigmoid activation function are performed respectively to obtain the gated signal map;
[0038] Phase modulation is performed on the phase information to obtain a phase feature map;
[0039] The modulated phase stream is obtained by performing element-wise matrix multiplication on the phase feature map and the gated signal map.
[0040] Preferably, the calculation of the covariance matrix specifically includes:
[0041] Construct a four-channel dynamic feature map;
[0042] The four-channel dynamic feature map is input into the Cholesky decomposition prediction network, specifically including:
[0043] The four-channel dynamic feature maps are concatenated along the channel dimension, and the concatenated feature maps are input into the prediction network. The prediction network includes, in sequence, 3x3 two-dimensional convolution, batch normalization, GELU activation function, 3x3 two-dimensional convolution, batch normalization, GELU activation function, 1x1 two-dimensional convolution and Tanh activation function;
[0044] The output of the prediction network is split into high-dimensional feature vectors by channel dimension.
[0045] Based on the high-dimensional eigenvectors and Cholesky components, a covariance matrix is constructed to obtain the covariance matrix diagram.
[0046] Preferably, feature extraction is performed on the covariance matrix graph to obtain a new feature map, including:
[0047] Perform eigenvalue decomposition on the covariance matrix of the covariance matrix graph;
[0048] Calculate the trace, logarithmic determinant, anisotropy, and principal eigenvectors based on the decomposed eigenvalues and eigenvectors;
[0049] The trace, logarithmic determinant, anisotropy, and principal eigenvector are concatenated and then subjected to layer normalization.
[0050] The normalized feature map is input into the refiner, which includes 1x1 two-dimensional convolution, normalization, GELU activation function and 1x1 two-dimensional convolution, to obtain the new feature map.
[0051] Preferably, TCN trajectory prediction includes:
[0052] The upsampled feature maps are concatenated along the channel dimension to form the input feature tensor to be processed;
[0053] Perform a dimension concatenation and flattening operation on the input feature tensor to convert it into a one-dimensional sequence suitable for time series modeling;
[0054] The one-dimensional sequence is processed through a linear layer and then input into a temporal convolutional network, which includes multiple TCN residual blocks and outputs a feature sequence processed by the TCN residual blocks.
[0055] The feature sequence is output through a linear layer as a feature flow jointly represented by the covariance and trajectory.
[0056] Preferably, the long-term global dependency modeling module includes:
[0057] Feature fusion unit: used to perform feature fusion on the two heterogeneous feature streams, and fuse them into a unified fused feature map by adding them element by element;
[0058] Quantization slicing unit: used to quantize the fused feature map and use a sliding window strategy to extract the effective time period containing the core dynamic signal on the time axis to obtain the sliced feature map;
[0059] Parallel stream processing unit: used to divide the sliced feature map into two independent parallel feature streams. Each feature stream is divided into multiple overlapping local blocks using a sliding window to obtain a sequence of local blocks;
[0060] Local block sequences are fed into a stream encoder to model the long-range temporal dependencies within each stream.
[0061] Hierarchical fusion and decision unit: After the output of the stream encoder, it is used to concatenate a learnable CLS token at the beginning of the local block sequence of each stream, and concatenate the sequences of the two streams in the feature dimension to form a joint sequence; the joint sequence is input to the cross-stream fusion Transformer module to perform bidirectional information interaction and CLS token extraction, and output dynamic target detection results through a fully connected layer.
[0062] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a deep learning-based anti-jamming millimeter-wave radar dynamic target detection system, which has the following effects:
[0063] (1) Effectively suppresses strong static clutter and fully preserves weak dynamic signals.
[0064] This invention directly subtracts the time mean in the complex domain to accurately remove the static background component, avoiding the non-selective attenuation of useful signals by frequency domain filtering. As a result, it can still recover micron-level target motion details with high fidelity in strong clutter environments, significantly improving the signal-to-noise ratio.
[0065] (2) Embed physical uncertainty modeling into a deep learning framework to improve feature interpretability and robustness.
[0066] Existing end-to-end networks treat covariance as a black box output, making it difficult to guarantee its physical legitimacy. This invention innovatively designs a Cholesky decomposition prediction network, forcing the output to satisfy a semi-positive definite covariance matrix through structural constraints, and further extracts statistical quantities with clear physical meaning, such as trace, logarithmic determinant, anisotropy, and principal eigenvectors. This mechanism enables the model not only to perceive "whether the target is moving," but also to understand "how, in which direction, and with what degree of uncertainty it is moving," significantly enhancing its discrimination ability in occluded, multi-target, or low signal-to-noise ratio scenarios.
[0067] (3) Efficiently model long-cycle dynamic dependencies and support real-time trajectory trend prediction.
[0068] Dynamic targets (such as pedestrians walking, vehicles starting and stopping, and robotic arms performing periodic movements) typically exhibit temporal correlations on the order of seconds or even longer. This invention employs a TCN network composed of dilated causal convolutions to achieve large receptive field modeling through a parallel structure, avoiding the gradient vanishing and serial bottlenecks of RNNs. While maintaining high inference efficiency, it accurately captures long-period motion patterns, providing a reliable basis for trajectory prediction and behavior prediction.
[0069] (4) Construct a dual-stream Transformer fusion architecture to achieve multi-view collaborative perception and global decision-making.
[0070] To address the issue that single-channel signals are susceptible to local obstruction or noise interference, this invention employs a dual-stream parallel processing + symmetrical cross-stream fusion + CLS Token aggregation mechanism. This enables information from different antenna or radar perspectives to fully interact at the high-level semantic level. Ultimately, a learnable classification token aggregates the global context, achieving end-to-end dynamic target detection with strong anti-interference capabilities and excellent generalization ability.
[0071] In summary, this invention significantly improves the accuracy, stability, and practicality of dynamic target detection systems in challenging scenarios such as complex environments, weak signals, and long-term dependencies without sacrificing the expressive power of deep learning, providing reliable technical support for applications such as intelligent monitoring, industrial automation, and security sensing. Attached Figure Description
[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0073] Figure 1 This is a schematic diagram of a deep learning-based anti-jamming millimeter-wave radar dynamic target detection system provided in an embodiment of the present invention;
[0074] Figure 2 This is a schematic diagram of the preprocessing module structure provided in an embodiment of the present invention;
[0075] Figure 3 This is a flowchart of bilateral filtering denoising provided in an embodiment of the present invention;
[0076] Figure 4 This is a schematic diagram of the feature extraction module structure provided in an embodiment of the present invention;
[0077] Figure 5 This is a flowchart of the feature extraction module provided in this embodiment of the invention;
[0078] Figure 6 This is a flowchart of the multi-core amplitude processor processing provided in an embodiment of the present invention;
[0079] Figure 7 This is a flowchart of the amplitude gating generator processing provided in an embodiment of the present invention;
[0080] Figure 8 This is the covariance matrix graph processing flow provided in the embodiments of the present invention;
[0081] Figure 9 This is a flowchart of covariance feature extraction provided in an embodiment of the present invention;
[0082] Figure 10 This is a flowchart of the TCN trajectory prediction process provided in an embodiment of the present invention;
[0083] Figure 11 This is a schematic diagram of the structure of the long-term global dependency modeling module provided in an embodiment of the present invention;
[0084] Figure 12 This is a flowchart of the long-term global dependency modeling module provided in this embodiment of the invention;
[0085] Figure 13 This is a flowchart of the stream encoder processing provided in an embodiment of the present invention;
[0086] Figure 14This is a flowchart of the feature fusion processing provided in an embodiment of the present invention;
[0087] Figure 15 This is a flowchart of the classification decision processing provided in an embodiment of the present invention. Detailed Implementation
[0088] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0089] This invention discloses a deep learning-based anti-jamming millimeter-wave radar dynamic target detection system, such as... Figure 1 As shown, it includes:
[0090] Preprocessing module: Used to perform high-fidelity preprocessing on the raw millimeter-wave radar signal, remove static clutter, complete phase correction, and output a high signal-to-noise ratio dynamic signal after bilateral filtering and noise reduction;
[0091] Feature extraction module: used for front-end feature extraction of high signal-to-noise ratio dynamic signals. By separating amplitude and phase information, it generates modulated phase stream and feature stream jointly characterized by covariance and trajectory in parallel, and outputs two heterogeneous feature streams.
[0092] Long-term global dependency modeling module: used to fuse two heterogeneous feature streams, and fuse them into a unified fused feature map by adding them element by element; input the fused feature map into the Transformer network, and output dynamic object detection results.
[0093] In this embodiment, as Figure 2 As shown, the preprocessing module includes:
[0094] Static clutter filtering unit: used to synthesize the amplitude and phase components of the original millimeter-wave radar signal into a complex matrix in the complex domain, calculate the average value of the complex matrix along the time dimension, and subtract the average value from the complex matrix to obtain a zero-mean complex signal containing dynamic target information;
[0095] Phase correction unit: used to extract phase components from zero-mean complex signals, and sequentially perform phase unwrapping and linear detrending processing on the phase components to obtain a continuous and trend-free corrected phase signal; among them, the phase unwrapping algorithm (unwrap_phase) is used to detect and correct abrupt changes exceeding a preset threshold (such as π).
[0096] Bilateral filtering unit: used to arrange the corrected phase signal into a two-dimensional spatiotemporal matrix, and to perform bilateral filtering and noise reduction on the two-dimensional spatiotemporal matrix through a two-dimensional bilateral filter to output a high signal-to-noise ratio dynamic signal.
[0097] This invention, by directly subtracting the average value in the complex domain, can more accurately strip away static background while preserving the integrity of dynamic signals to the maximum extent. Meanwhile, the original phase data is confined to the interval [-π, π]. When the continuous movement of the target exceeds this range, a 2π jump occurs, forming a discontinuous "sawtooth wave," which can seriously mislead subsequent processing. The phase unwrapping of this invention can eliminate this jump and restore its linear continuity. Furthermore, due to changes in ambient temperature or hardware drift, the phase signal often exhibits a slow linear trend; detrending operations can remove this interference unrelated to target movement. When calculating the filtered value of a point, the bilateral filter considers two weighting factors: spatial distance weight and pixel value similarity weight. It only averages points that are spatially adjacent and numerically similar, while retaining the original value for points with large numerical differences (i.e., "edges"). This mechanism allows the bilateral filter to effectively filter out random noise while perfectly preserving important edges and details of the signal, resolving the fundamental contradiction between noise reduction and fidelity preservation.
[0098] Furthermore, such as Figure 3 As shown, bilateral filtering denoising includes:
[0099] S101: Traverse each central pixel in the two-dimensional spatiotemporal matrix;
[0100] S102: For each center pixel, select all neighboring pixels within its preset neighborhood range;
[0101] S103: Calculate the two Gaussian weights between the neighboring pixels and the center pixel respectively:
[0102] The first weight is the spatial distance weight based on the spatial coordinate distance between neighboring pixels and the center pixel. Its value is determined by the spatial Gaussian function. The closer the spatial distance, the higher the weight.
[0103] The second weight is the pixel value similarity weight based on the difference in phase values between the two. Its value is determined by a numerical Gaussian function. The smaller the numerical difference, the higher the weight.
[0104] S104: Multiply the spatial distance weight and the pixel value similarity weight element by element to obtain the combined weight corresponding to each neighboring pixel.
[0105] S105: Calculate the new denoised value of the center pixel by using the combined weights of all neighboring pixels to calculate the weighted average of the corresponding pixel values.
[0106] S106: Determine whether all center pixels have been processed. If not, return to S102 to continue processing the next center pixel; if yes, output a high signal-to-noise ratio dynamic signal.
[0107] In this embodiment, as Figure 4 and Figure 5 As shown, the feature extraction module includes:
[0108] Signal normalization and separation unit: used to normalize and decompose high signal-to-noise ratio dynamic signals into amplitude and phase information;
[0109] Modulation unit: used to perform amplitude gating processing on amplitude information and phase modulation on phase information to generate modulated phase stream;
[0110] Covariance processing unit: The phase signal is fused between channels and spatially averaged (downsampled), the covariance matrix is calculated, the covariance matrix is used to extract features to obtain a new feature map, the new feature map is upsampled and the trajectory is predicted by TCN temporal convolutional network, and the feature flow jointly represented by covariance and trajectory is obtained.
[0111] Furthermore, the specific processing procedure of the modulation unit is as follows:
[0112] The amplitude information is passed through three parallel convolutional layers with 3x3, 5x5, and 7x7 convolutional kernels to extract features at different scales;
[0113] Features at different scales are concatenated along the channel dimension, then batch normalized and processed using the GELU activation function to obtain a multi-scale amplitude feature map, such as... Figure 6 As shown;
[0114] The multi-scale amplitude feature map is fed into an amplitude gating generator, where 3x3 2D convolution, batch normalization, SiLU activation, 1x1 2D convolution, and Sigmoid activation are performed respectively to obtain a gated signal map. This process is essentially a spatial attention generation network, consisting of a multi-kernel feature processor and a small convolutional network amplitude gating generator. The final output is a "gated signal" or "confidence map" with the same size as the input feature map and values between [0, 1]. Figure 7 As shown;
[0115] Phase modulation is performed on the phase information to obtain a phase feature map;
[0116] The modulated phase stream is obtained by performing element-wise matrix multiplication on the phase feature map and the gated signal map.
[0117] The amplitude signal strength of this invention directly reflects the quality and signal-to-noise ratio of the echo signal. A high-quality signal typically corresponds to a reliable phase value. This module aims to learn this correspondence, generating a "reliability" weight for each spatiotemporal location of the signal. By using convolutional kernels of various sizes such as 3x3, 5x5, and 7x7, the network can simultaneously capture amplitude variation features at different scales, thereby more accurately evaluating signal quality.
[0118] This invention significantly improves the quality of the phase signal input to the subsequent network by multiplying the separated phase signal element-by-element with the aforementioned gate signal.
[0119] Furthermore, such as Figure 8 As shown, the calculation of the covariance matrix diagram specifically includes:
[0120] A four-channel dynamic feature map is constructed, including: the current signal state map, the difference map between the current state and the initial state (reflecting long-term changes), the difference map between adjacent time points (reflecting short-term process noise), and the standard deviation map of the signal in the time dimension (reflecting overall volatility).
[0121] The four-channel dynamic feature map is input into the Cholesky decomposition prediction network, specifically including:
[0122] The four-channel dynamic feature maps are concatenated along the channel dimension. The concatenated feature maps are then input into the prediction network, which consists of 3x3 two-dimensional convolution, batch normalization, GELU activation function, 1x1 two-dimensional convolution and Tanh activation function, which can effectively handle spatiotemporal features.
[0123] The output of the prediction network is split into high-dimensional feature vectors by channel dimension;
[0124] A covariance matrix is constructed based on high-dimensional eigenvectors and Cholesky components, resulting in a covariance matrix graph. Specifically, the high-dimensional eigenvectors are transformed into a one-dimensional vector through dimension swapping and flattening, with the number of elements equal to the number of parameters required for the Cholesky factor. The reconstruction process, using a dedicated reconstruction function, is as follows: First, a lower triangular zero matrix L is initialized. Vector elements are then filled into the lower triangular portion of L. Diagonal elements and the strictly lower triangular portion are extracted from the filled matrix L. The Softplus activation function is applied to the extracted diagonal elements to ensure all diagonal values are positive. The regularized diagonal elements are reconstructed into a diagonal matrix, which is then added to the strictly lower triangular portion to obtain a regularized lower triangular matrix L1. Finally, by calculating P = L1 * L1 transpose, the final stable and physically correct covariance matrix graph is obtained.
[0125] This invention effectively ensures the physical validity and numerical stability of the covariance matrix by introducing a covariance modeling mechanism based on Cholesky factor prediction.
[0126] Since the covariance matrix theoretically must satisfy symmetry and positive semidefiniteness, directly outputting the complete matrix by a neural network can easily violate these physical constraints due to parameter perturbations, leading to training divergence or inference anomalies. This invention predicts the lower triangular Cholesky factor L and reconstructs the covariance matrix using P = L1 * L1. This naturally guarantees the symmetry and positive semidefiniteness of the output matrix at the network structure level, significantly improving the model's numerical robustness and convergence stability throughout the training and inference processes.
[0127] Significantly reduces parameter redundancy and improves learning efficiency:
[0128] For an n-dimensional state space, the complete covariance matrix contains n elements, which contain a large amount of symmetric redundancy. This invention only needs to predict n(n+1) independent parameters (i.e., the non-zero elements of the Cholesky factor), which not only reduces the output dimension by nearly half, but also avoids the additional burden of the network being forced to explicitly learn a symmetric structure, thereby accelerating the convergence speed, reducing the risk of overfitting, and improving parameter utilization efficiency.
[0129] Achieve deep fusion of multi-scale dynamic contexts to enhance the accuracy of uncertainty estimation:
[0130] To accurately characterize the time-varying uncertainties in the target's motion, the Cholesky predictor network designed in this invention does not rely solely on the current observation state. Instead, it fuses four types of dynamic features at the channel level: the current signal state, the long-term trend offset (the difference from the initial state), the short-term process noise (difference between adjacent frames), and the intensity of time-dimensional fluctuations (standard deviation map), using these as network inputs. This multi-source information fusion mechanism enables the model to simultaneously perceive the target's instantaneous changes, historical evolution, and overall fluctuation characteristics, thereby generating a more discriminative and physically consistent covariance representation, significantly improving the reliability of subsequent feature extraction and classification decisions.
[0131] Furthermore, such as Figure 9 As shown, feature extraction is performed on the covariance matrix graph to obtain a new feature map, including:
[0132] Perform eigenvalue decomposition on the covariance matrix of the covariance matrix graph;
[0133] Calculate the trace, logarithmic determinant, anisotropy, and principal eigenvectors based on the decomposed eigenvalues and eigenvectors;
[0134] The trace, logarithmic determinant, anisotropy, and principal eigenvector are concatenated and then subjected to layer normalization.
[0135] The normalized feature map is input into the refiner, which includes 1x1 2D convolution, normalization, GELU activation function and 1x1 2D convolution, to obtain a new feature map.
[0136] The complete covariance matrix of this invention contains a large amount of redundant information. Through eigenvalue decomposition, this information can be refined into several key scalars and vectors with clear physical meanings. For example, the "trace" represents the magnitude of the overall uncertainty, the "determinant" represents the volume of the uncertainty ellipsoid, and "anisotropy" describes the directionality of the uncertainty. These refined features are easier for subsequent networks to learn and utilize than the original matrix.
[0137] Furthermore, such as Figure 10 As shown, TCN trajectory prediction includes:
[0138] The upsampled feature maps are concatenated along the channel dimension to form the input feature tensor to be processed;
[0139] Perform dimensional concatenation and flattening operations on the input feature tensor to transform it into a one-dimensional sequence suitable for time series modeling;
[0140] A one-dimensional sequence is processed through a linear layer and then input into a temporal convolutional network (TCN). The TCN consists of N layers of TCN residual blocks, where i is a cyclic index variable, labeled from i=0 to i=N-1. The output is the feature sequence processed by the TCN residual blocks. Specifically, data flows sequentially through each layer, starting from the top TCN residual block. Each layer contains three core operations: first, a one-dimensional convolution (kernel size 3) is performed. The left padding length of the one-dimensional convolution operation is set to be equal to the dilation rate, constraining the receptive field of the convolution kernel to not exceed the left range of the current time step. This ensures that the output of the convolution operation at each time step depends only on the input values of the current and previous time steps. The dilation rate and padding value increase exponentially with the layer number, with the dilation coefficient of the i-th layer being 2. i Next, normalization is performed; finally, the GELU activation function is applied.
[0141] The feature sequence is output by a linear layer, which provides a feature flow jointly characterized by covariance and trajectory.
[0142] This invention feeds the upsampled feature map into a Temporal Convolutional Network (TCN) to predict the trajectory trend of a target. By using causal convolution and an exponentially increasing dilation rate, a large receptive field can be obtained with fewer layers, thus efficiently capturing this long-term temporal dependency, which is crucial for predicting motion trajectories. Compared with RNNs, TCNs can be computed in parallel, are more efficient, and do not suffer from the gradient vanishing problem.
[0143] In this embodiment, as Figure 11 and Figure 12 As shown, the long-term global dependency modeling module includes:
[0144] Feature fusion unit: used to fuse two heterogeneous feature streams, and fuse them into a unified fused feature map by adding them element by element;
[0145] Quantization slicing unit: used to quantize the fused feature map and use a sliding window strategy to extract the effective time period containing the core dynamic signal on the time axis to obtain the sliced feature map;
[0146] Parallel stream processing unit: used to divide the sliced feature map into two independent parallel feature streams. Each feature stream is divided into multiple overlapping local patches using a sliding window to obtain a sequence of local patches. The sliding window can be set to a window size of 25 and a step size of 10. Each local patch is then flattened into a one-dimensional vector.
[0147] The local block sequence is fed into the Stream Encoder. Each Stream Encoder consists of multiple layers of Transformer blocks, modeling the long-range temporal dependencies within its respective stream, such as... Figure 13 As shown;
[0148] The specific processing procedure of the Stream Encoder is as follows:
[0149] Parameter configuration: Receive the flag parameter Flag, and determine the hidden layer dimension and number of channels of the encoder based on the Flag value;
[0150] Encoder construction: Based on the determined configuration parameters, an encoder layer containing multiple Transformer blocks is constructed. Each Transformer block contains layer normalization, self-attention mechanism, residual connection, and multilayer perceptron processing. The number of Transformer blocks is a preset value, which can be selected as 3.
[0151] Forward propagation processing: The input hidden state is processed sequentially through the multiple Transformer blocks. Specifically, in the first stage: The original input is processed by the first layer normalization unit. The normalized features are then fed into the self-attention mechanism for computation. The self-attention mechanism produces two outputs: the main output flows to the first summing node, and the second output forms the output weight branch. At the first summing node, the main output of the self-attention mechanism is added to the original input through a residual connection. In the second stage: The features after the residual connection and addition enter the second layer normalization unit. The normalized features are then fed into a multilayer perceptron for nonlinear transformation. The output of the multilayer perceptron flows to the second summing node. At the second summing node, the output of the multilayer perceptron and the input features before the second layer normalization are combined and added through a residual connection, and the processing result is output.
[0152] The output of the last Transformer block is subjected to layer normalization to obtain the final encoder output.
[0153] Hierarchical fusion and decision unit: After the stream encoder output, it appends a learnable CLS Token (Classification Token) to the beginning of the local block sequence of each stream and concatenates the sequences of the two streams along the feature dimension to form a joint sequence. The joint sequence is input to the cross-stream fusion Transformer module. To avoid order bias, it typically performs two symmetrical fusions (stream 1 + stream 2 and stream 2 + stream 1), such as... Figure 14 As shown, the cross-stream fusion results from both directions are concatenated again and fed into a second final fusion Transformer module for higher-level global information aggregation. After processing, only the final state of the CLS Token corresponding to each Patch sequence is extracted, which encapsulates the essence of the entire sequence. All extracted CLS Token sequences are fed into a fully connected layer (Final FC) for final classification, and then dequantized to restore them to their original numerical range. The final output is the classification result Logits, as shown. Figure 15 As shown.
[0154] The specific processing procedure of Final FC is as follows:
[0155] All extracted CLS Token sequences are processed in parallel, and the features of the 6 channels are processed separately.
[0156] Six parallel classifiers process the separation features respectively, and then the original decision signals of the multiple classifiers are transformed into a unified, numerically stable comprehensive confidence vector through absolute value operation, stacking and summation.
[0157] The vector is normalized using a softmax-like process to generate an output that conforms to the probability distribution.
[0158] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0159] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A deep learning-based anti-jamming millimeter-wave radar dynamic target detection system, characterized in that, include: Preprocessing module: Used to perform high-fidelity preprocessing on the raw millimeter-wave radar signal, remove static clutter, complete phase correction, and output a high signal-to-noise ratio dynamic signal after bilateral filtering and noise reduction; Feature extraction module: used to perform front-end feature extraction on the high signal-to-noise ratio dynamic signal, by separating amplitude and phase information, generating in parallel the modulated phase stream and the feature stream jointly characterized by covariance and trajectory, and outputting two heterogeneous feature streams; Long-term global dependency modeling module: used to perform feature fusion on the two heterogeneous feature streams, and fuse them into a unified fusion feature map through element-by-element addition operation; The fused feature map is input into the Transformer network, which outputs dynamic target detection results. The feature extraction module includes: Signal normalization and separation unit: used to normalize the high signal-to-noise ratio dynamic signal and decompose it into amplitude information and phase information; Modulation unit: used to perform amplitude gating processing on the amplitude information and phase modulation on the phase information to generate a modulated phase stream; Covariance processing unit: performs inter-channel fusion and spatial average pooling on the phase flow, calculates the covariance matrix graph, extracts features from the covariance matrix graph to obtain a new feature map, upsamples the new feature map and performs TCN trajectory prediction to obtain the feature flow jointly represented by the covariance and trajectory. The calculation of the covariance matrix specifically includes: Construct a four-channel dynamic feature map; The four-channel dynamic feature map is input into the Cholesky decomposition prediction network, specifically including: The four-channel dynamic feature maps are concatenated along the channel dimension, and the concatenated feature maps are input into the prediction network. The prediction network includes, in sequence, 3x3 two-dimensional convolution, batch normalization, GELU activation function, 3x3 two-dimensional convolution, batch normalization, GELU activation function, 1x1 two-dimensional convolution and Tanh activation function; The output of the prediction network is split into high-dimensional feature vectors by channel dimension. Based on the high-dimensional eigenvectors and Cholesky components, a covariance matrix is constructed to obtain the covariance matrix diagram.
2. The system as described in claim 1, characterized in that, The preprocessing module includes: Static clutter filtering unit: used to synthesize the amplitude and phase components of the original millimeter-wave radar signal into a complex matrix in the complex domain, calculate the average value of the complex matrix along the time dimension, and subtract the average value from the complex matrix to obtain a zero-mean complex signal containing dynamic target information; Phase correction unit: used to extract the phase component from the zero-mean complex signal, and sequentially perform phase unwinding and linear detrending processing on the phase component to obtain a continuous and trend-free corrected phase signal; Bilateral filtering unit: used to arrange the corrected phase signal into a two-dimensional spatiotemporal matrix, and to perform bilateral filtering and noise reduction on the two-dimensional spatiotemporal matrix through a two-dimensional bilateral filter to output the high signal-to-noise ratio dynamic signal.
3. The system as described in claim 2, characterized in that, The bilateral filtering denoising includes: S101: Traverse each central pixel in the two-dimensional spatiotemporal matrix; S102: For each central pixel, select all neighboring pixels within its preset neighborhood range; S103: Calculate the spatial distance weight and pixel value similarity weight between the neighboring pixels and the center pixel respectively; S104: Multiply the spatial distance weight and the pixel value similarity weight element by element to obtain the combined weight corresponding to each neighboring pixel. S105: Calculate the new denoised value of the center pixel by using the combined weights of all neighboring pixels to calculate the weighted average of the corresponding pixel values. S106: Determine whether all the center pixels have been processed. If not, return to S102 to continue processing the next center pixel. If yes, output the high signal-to-noise ratio dynamic signal.
4. The system as described in claim 1, characterized in that, The specific processing procedure of the modulation unit is as follows: The amplitude information is passed through three parallel convolutional layers with 3x3, 5x5, and 7x7 kernels to extract features at different scales; After concatenating features of different scales along the channel dimension, batch normalization and GELU activation function are applied to obtain a multi-scale amplitude feature map. The multi-scale amplitude feature map is fed into the amplitude gating generator, and 3x3 two-dimensional convolution, batch normalization, SiLU activation function, 1x1 two-dimensional convolution and Sigmoid activation function are performed respectively to obtain the gated signal map; Phase modulation is performed on the phase information to obtain a phase feature map; The modulated phase stream is obtained by performing element-wise matrix multiplication on the phase feature map and the gated signal map.
5. The system as described in claim 4, characterized in that, Feature extraction is performed on the covariance matrix graph to obtain a new feature map, including: Perform eigenvalue decomposition on the covariance matrix of the covariance matrix graph; Calculate the trace, logarithmic determinant, anisotropy, and principal eigenvectors based on the decomposed eigenvalues and eigenvectors; The trace, logarithmic determinant, anisotropy, and principal eigenvector are concatenated and then subjected to layer normalization. The normalized feature map is input into the refiner, which includes 1x1 two-dimensional convolution, normalization, GELU activation function and 1x1 two-dimensional convolution, to obtain the new feature map.
6. The system as described in claim 5, characterized in that, TCN trajectory prediction includes: The upsampled feature maps are concatenated along the channel dimension to form the input feature tensor to be processed; Perform a dimension concatenation and flattening operation on the input feature tensor to convert it into a one-dimensional sequence suitable for time series modeling; The one-dimensional sequence is processed through a linear layer and then input into a temporal convolutional network, which includes multiple TCN residual blocks and outputs a feature sequence processed by the TCN residual blocks. The feature sequence is output through a linear layer as a feature flow jointly represented by the covariance and trajectory.
7. The system as described in claim 1, characterized in that, The long-term global dependency modeling module includes: Feature fusion unit: used to perform feature fusion on the two heterogeneous feature streams, and fuse them into a unified fused feature map by adding them element by element; Quantization slicing unit: used to quantize the fused feature map and use a sliding window strategy to extract the effective time period containing the core dynamic signal on the time axis to obtain the sliced feature map; Parallel stream processing unit: used to divide the sliced feature map into two independent parallel feature streams. Each feature stream is divided into multiple overlapping local blocks using a sliding window to obtain a sequence of local blocks; Local block sequences are fed into a stream encoder to model the long-range temporal dependencies within each stream. Hierarchical fusion and decision unit: After the output of the stream encoder, it is used to concatenate a learnable CLS token at the beginning of the local block sequence of each stream, and concatenate the sequences of the two streams in the feature dimension to form a joint sequence; the joint sequence is input to the cross-stream fusion Transformer module to perform bidirectional information interaction and CLS token extraction, and output dynamic target detection results through a fully connected layer.
Citation Information
Patent Citations
Behavior recognition method and device, electronic equipment and storage medium
CN116347374A
Electronic scanning radar device, reception wave direction estimation method, and reception wave direction estimation program
JP2011242288A