Anti-interference millimeter wave radar dynamic target detection system based on deep learning

Abstract

The invention discloses an anti-interference millimeter wave radar dynamic target detection system based on deep learning, and relates to the technical field of radar signal processing and deep learning fusion. Static clutters are removed, phase correction is completed, and a high-signal-to-noise-ratio dynamic signal is output after bilateral filtering and denoising; the feature extraction module is used for performing front-end feature extraction on the high-signal-to-noise-ratio dynamic signal, generating a modulated phase flow and a feature flow jointly represented by covariance and a track in parallel by separating amplitude and phase information, and outputting two paths of heterogeneous feature flows; the long-time-sequence global dependency modeling module is used for performing feature fusion on the two paths of heterogeneous feature flows and fusing the two paths of heterogeneous feature flows into a unified fusion feature graph through element-by-element addition operation; and inputting the fusion feature map into a Transform network, and outputting a dynamic target detection result. And the dynamic target detection precision and robustness in a complex environment are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar signal processing and deep learning fusion, more particularly to a kind of anti-jamming millimeter wave radar dynamic target detection system based on deep learning. BACKGROUND

[0002] In recent years, non-contact sensing technology represented by millimeter wave radar is widely used in smart home, security monitoring, medical health, human-computer interaction and other fields due to its unique advantages such as privacy protection, anti-environmental light interference and penetration of non-metallic barriers. One of the core tasks of these applications is to accurately detect, identify and analyze the dynamic information of the target from the continuous data stream of radar echoes.

[0003] Although the existing technology has made some progress in radar signal analysis, both traditional signal processing methods and general deep learning models have exposed a series of deep and interrelated problems when facing complex, variable and low signal-to-noise ratio environments, which have become the core bottleneck of current technology development: Problem one: overwhelming interference of static clutter and distortion of phase information In real application scenarios, the effective signal from the target (such as the weak fluctuation of the human chest) is extremely weak, and its energy is often several orders of magnitude lower than the reflected signal from stationary objects (i.e. "static clutter") such as walls and furniture. Existing methods cannot effectively separate the two signals.

[0004] Problem two: "blind" fusion of multi-modal features Radar signals naturally contain amplitude and phase information of two modalities, which carry different physical meanings. Amplitude reflects the strength of the signal (which can be used as a proxy for signal-to-noise ratio), while phase contains accurate displacement information of the target. Existing deep learning models usually use simple channel concatenation (Concatenation) to fuse the two types of information.

[0005] Problem three: lack of physical prior knowledge and "black box" of model General deep learning models (such as CNN, RNN) are purely data-driven "black boxes" that lack an intrinsic understanding of the physical laws behind the radar sensing process.

[0006] Problem four: insufficient ability to capture long-term global dependencies The behavior patterns of the target, whether periodic physiological activities (breathing, heartbeat) or non-periodic complex actions (turning over, sitting up), are developed on a long time scale.

[0007] In summary, these problems are not isolated, but are intertwined with each other, and together form a huge technical barrier, hindering the performance improvement and reliable application of radar perception technology in complex real-world scenarios. Therefore, there is an urgent need in the field for a new technical paradigm that can intelligently fuse multi-modal information, deeply embed physical priors, efficiently capture global temporal dependencies, and make reliable decisions. SUMMARY

[0008] In view of the above problems, the present application is proposed in order to provide a deep learning-based anti-interference millimeter wave radar dynamic target detection system that overcomes the above problems or at least partially solves the above problems.

[0009] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a deep learning-based anti-interference millimeter wave radar dynamic target detection system, comprising: A preprocessing module for high-fidelity preprocessing of original millimeter wave radar signals, removing static clutter, completing phase correction, and outputting high signal-to-noise ratio dynamic signals after bilateral filter denoising; A feature extraction module for front-end feature extraction of the high signal-to-noise ratio dynamic signals, separating amplitude and phase information, generating modulated phase streams in parallel, and outputting two heterogeneous feature streams represented by covariance and trajectory; A long-time global dependency modeling module for feature fusion of the two heterogeneous feature streams, fused into a unified fusion feature map through element-by-element addition operation; inputting the fusion feature map into a Transformer network, and outputting a dynamic target detection result.

[0010] Preferably, the preprocessing module comprises: A static clutter filtering unit for combining the amplitude and phase components of the original millimeter wave radar signals into a complex matrix in the complex domain, calculating the average value of the complex matrix along the time dimension, and subtracting the average value from the complex matrix to obtain a zero-mean complex signal containing dynamic target information; A phase correction unit for extracting the phase component from the zero-mean complex signal, and sequentially performing phase unwrapping processing and linear detrending processing on the phase component to obtain a corrected phase signal that is continuous and free of trend interference; A bilateral filtering unit for arranging the corrected phase signal into a two-dimensional spatiotemporal matrix, performing bilateral filter denoising on the two-dimensional spatiotemporal matrix through a two-dimensional bilateral filter, and outputting the high signal-to-noise ratio dynamic signal.

[0011] Preferably, the bilateral filter denoising comprises: S101: Traverse each center pixel point in the two-dimensional spatiotemporal matrix; S102: For each of the center pixel points, select all neighborhood pixel points within its preset neighborhood range; S103: Calculate the spatial distance weight and the pixel value similarity weight between the neighborhood pixel points and the center pixel point respectively; S104: Multiply the spatial distance weight and the pixel value similarity weight element by element to obtain the combined weight corresponding to each neighborhood pixel point; S105: Weighted average of the pixel value corresponding to all neighborhood pixel points using the combined weight, calculate the new value of the center pixel point after denoising; S106: Determine whether all the center pixel points have been processed, if not, return to S102 to continue processing the next center pixel point; if yes, output the high signal-to-noise ratio dynamic signal.

[0012] Preferably, the feature extraction module comprises: Signal normalization and separation unit: for normalizing and decomposing the high signal-to-noise ratio dynamic signal into amplitude information and phase information; Modulation unit: for amplitude gating processing of the amplitude information, and phase modulation of the phase information, generating a modulated phase stream; Covariance processing unit: inter-channel fusion and spatial average pooling of the phase signal, calculation of the covariance matrix graph, feature extraction of the covariance matrix graph to obtain a new feature graph, upsampling of the new feature graph and TCN trajectory prediction, to obtain the feature stream of the covariance and trajectory joint representation.

[0013] Preferably, the modulation unit specifically processes as follows: The amplitude information is passed through three parallel 3x3, 5x5, and 7x7 convolution kernels to extract features of different scales; The features of different scales are concatenated in the channel dimension, then batch normalized and processed by GELU activation function to obtain a multi-scale amplitude feature map; The multi-scale amplitude feature map is input into an 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 a gating signal map; The phase information is phase-modulated to obtain a phase feature map; The phase feature map and the gating signal map are multiplied element by element to obtain the modulated phase stream.

[0014] Preferably, the calculation of the covariance matrix graph specifically comprises: Construct a four-channel dynamic feature map; The four-channel dynamic feature map is input into a Cholesky decomposition prediction network, specifically including: The four-channel dynamic feature map is spliced in the channel dimension, and the spliced feature map is input into a prediction network, which sequentially includes a 3x3 two-dimensional convolution, batch normalization, a GELU activation function, a 3x3 two-dimensional convolution, batch normalization, a GELU activation function, a 1x1 two-dimensional convolution, and a Tanh activation function. The output of the prediction network is output as a high-dimensional feature vector through channel dimension splitting; A covariance matrix is constructed based on the high-dimensional feature vector and Cholesky components, and the covariance matrix graph is obtained.

[0015] Preferably, feature extraction is performed on the covariance matrix graph to obtain a new feature graph, including: Eigenvalue decomposition is performed on the covariance matrix of the covariance matrix graph; The trace, logarithmic determinant, anisotropy, and principal eigenvector are calculated according to the decomposed eigenvalues and eigenvectors; The trace, logarithmic determinant, anisotropy, and principal eigenvector are spliced and then subjected to layer normalization; The layer-normalized feature graph is input into a refiner including a 1x1 two-dimensional convolution, normalization, a GELU activation function, and a 1x1 two-dimensional convolution, to obtain the new feature graph.

[0016] Preferably, TCN trajectory prediction includes: The up-sampled feature graph is spliced in the channel dimension to form an input feature tensor to be processed; The input feature tensor is subjected to dimension splicing and flattening operations to convert it into a one-dimensional sequence suitable for time series modeling; The one-dimensional sequence is input into a time series convolution network after being processed by a linear layer, and the time series convolution network includes a plurality of TCN residual blocks, and outputs a feature sequence processed by the TCN residual blocks; The feature sequence outputs a feature stream of the covariance and trajectory joint representation through a linear layer.

[0017] Preferably, the long-time series global dependency modeling module includes: A feature fusion unit is configured to fuse the two heterogeneous feature streams into a unified fusion feature graph through an element-by-element addition operation; A quantization slicing unit is configured to quantize the fusion feature graph and adopt a sliding window strategy to intercept an effective time period containing a core dynamic signal on a time axis to obtain a sliced feature graph; 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.

[0018] 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: (1) Effectively suppresses strong static clutter and fully preserves weak dynamic signals. 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.

[0019] (2) Embed physical uncertainty modeling into a deep learning framework to improve feature interpretability and robustness. 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.

[0020] (3) Efficiently model long-term dynamic dependencies and support real-time trajectory trend prediction. 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.

[0021] (4) Construct a double-flow Transformer fusion architecture to realize multi-view collaborative perception and global decision-making In view of the problem that a single signal is easy to be interfered by local occlusion or noise, the present application realizes full interaction of information of different antennas or radar views at a high-level semantic level through double-flow parallel processing + symmetric cross-flow fusion + CLS Token aggregation mechanism, and finally condenses global context by a learnable classification Token, so that an end-to-end dynamic target detection with strong anti-interference and excellent generalization ability is realized.

[0022] In summary, the present application significantly improves the accuracy, stability and practicability of the dynamic target detection system in complex environments, weak signals, long period dependencies and other typical challenge scenarios without sacrificing the expression ability of deep learning, and provides reliable technical support for intelligent monitoring, industrial automation, security perception and other applications. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creating any creative labor.

[0024] Figure 1 A schematic diagram of an anti-interference millimeter wave radar dynamic target detection system based on deep learning provided in an embodiment of the present application; Figure 2 A structure schematic diagram of a preprocessing module provided in an embodiment of the present application; Figure 3 A bilateral filtering denoising flowchart provided in an embodiment of the present application; Figure 4 A structure schematic diagram of a feature extraction module provided in an embodiment of the present application; Figure 5 A feature extraction module processing flowchart provided in an embodiment of the present application; Figure 6 A multi-core amplitude processor processing flowchart provided in an embodiment of the present application; Figure 7 An amplitude gate generator processing flowchart provided in an embodiment of the present application; Figure 8 A covariance matrix processing flowchart provided in an embodiment of the present application; Figure 9 A covariance feature extraction flowchart provided in an embodiment of the present application; Figure 10 A TCN trajectory prediction flowchart provided in an embodiment of the present application; Figure 11 A long-time sequence global dependence modeling module structure diagram provided in an embodiment of the present application is shown in the figure; Figure 12 A long-time sequence global dependence modeling module processing flowchart provided in an embodiment of the present application is shown in the figure; Figure 13 A stream encoder processing flowchart provided in an embodiment of the present application is shown in the figure; Figure 14 A feature fusion processing flowchart provided in an embodiment of the present application is shown in the figure; Figure 15 A classification decision processing flowchart provided in an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0026] An anti-interference millimeter wave radar dynamic target detection system based on deep learning is disclosed in an embodiment of the present application, as shown in the figure, comprising: Figure 1 A preprocessing module: used for high-fidelity preprocessing of original millimeter wave radar signals, removing static clutter, completing phase correction, and outputting high signal-to-noise ratio dynamic signals after double-sided filter denoising; A feature extraction module: used for front-end feature extraction of high signal-to-noise ratio dynamic signals, generating modulation phase streams and feature streams jointly represented by covariance and trajectory through separation of amplitude and phase information in parallel, and outputting two-way heterogeneous feature streams; A long-time sequence global dependence modeling module: used for feature fusion of two-way heterogeneous feature streams, fused into a unified fusion feature map through element-by-element addition operation; the fusion feature map is input into a Transformer network to output a dynamic target detection result. In the embodiment, as shown in the figure, the preprocessing module comprises:

[0027] A static clutter filtering unit: used for combining amplitude and phase components of original millimeter wave radar signals into a complex matrix in a complex domain, calculating an average value of the complex matrix along a time dimension, and subtracting the average value from the complex matrix to obtain a zero-mean complex signal containing dynamic target information; Figure 2 ​​The phase correction unit is used for extracting a phase component from the zero-mean complex signal, and sequentially performing phase unwrapping processing and linear detrending processing on the phase component to obtain a corrected phase signal which is continuous and free of trend interference.

[0028] The bilateral filtering unit is used for arranging the corrected phase signal into a two-dimensional space-time matrix, performing bilateral filtering denoising on the two-dimensional space-time matrix through a two-dimensional bilateral filter (Bilateral filter), and outputting a high signal-to-noise ratio dynamic signal.

[0029] The present application can more accurately strip the static background while maximizing the integrity of the dynamic signal by directly subtracting the average value in the complex domain. At the same time, the original phase data is wrapped in the interval [-π, π], when the continuous motion of the target exceeds this range, a 2π jump will occur, forming a discontinuous "sawtooth wave", which will seriously mislead the subsequent processing. The phase unwrapping of the present application can eliminate such jumps and restore its linear continuity. In addition, due to environmental temperature changes or hardware drift, the phase signal often superimposes a slow linear trend, and the detrending operation can remove this interference unrelated to the target motion. The bilateral filter considers two weight factors: spatial distance weight and pixel value similarity weight when calculating the filtered value of a point, and only averages those points that are spatially adjacent and similar in value, while retaining the original value for points with large value differences (i.e. "edges"). This mechanism enables the bilateral filter to effectively filter out random noise while perfectly preserving important edges and details of the signal, solving the fundamental contradiction between denoising and fidelity.

[0030] Further, as shown in Figure 3 the bilateral filtering denoising includes: S101: traversing each center pixel point in the two-dimensional space-time matrix; S102: for each center pixel point, selecting all neighborhood pixel points within a preset neighborhood range thereof; S103: calculating two Gaussian weights between the neighborhood pixel points and the center pixel point, respectively: The first weight is a spatial distance weight based on the spatial coordinate distance between the neighborhood pixel points and the center pixel point, and the value thereof is determined by a spatial Gaussian function, the closer the spatial distance, the higher the weight; The second weight is a pixel value similarity weight based on the phase value difference between the two, and the value thereof is determined by a numerical Gaussian function, the smaller the numerical difference, the higher the weight; S104: multiplying the spatial distance weight and the pixel value similarity weight element by element to obtain a combined weight corresponding to each neighborhood pixel point; S105: Weighted average of the corresponding pixel value of all neighborhood pixel points using the combined weight of the neighborhood pixel points, and calculate the new value of the center pixel point after denoising; S106: Determine whether all center pixel points have been processed, if not, return to S102 to continue processing the next center pixel point; if yes, output the high signal-to-noise ratio dynamic signal.

[0031] In this embodiment, as shown in Figure 4 and Figure 5 The feature extraction module comprises: Signal normalization and separation unit: for normalizing and decomposing the high signal-to-noise ratio dynamic signal into amplitude information and phase information; Modulation unit: for amplitude gating processing of amplitude information, and phase modulation of phase information, generating modulated phase stream; Covariance processing unit: inter-channel fusion and spatial average pooling (down-sampling) of phase signal, calculation of covariance matrix graph, feature extraction of covariance matrix graph to obtain new feature graph, up-sampling of new feature graph and TCN temporal convolutional network (Temporal Convolutional Networks) trajectory prediction, to obtain feature stream of covariance and trajectory joint representation.

[0032] Further, the specific processing process of the modulation unit is: The amplitude information is passed through three parallel 3x3, 5x5, and 7x7 convolution kernels to extract features of different scales; The features of different scales are concatenated in the channel dimension, then batch normalized and processed by GELU activation function to obtain multi-scale amplitude feature maps, as shown in Figure 6 ; The multi-scale amplitude feature maps are sent to 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 gating signal map; the above process is essentially a spatial attention generation network, which is composed of a multi-kernel (Multi-Kernel) feature processor and a small convolution network amplitude gating generator, and finally outputs a "gating signal" or "confidence map" with the same size as the input feature map and a value between [0, 1], as shown in Figure 7 ; The phase information is phase-modulated to obtain a phase feature map; The phase feature map and the gating signal map are multiplied element by element to obtain the modulated phase stream.

[0033] The strength of the amplitude signal of the present application directly reflects the quality of the echo signal and the signal-to-noise ratio. A high-quality signal usually corresponds to a reliable phase value. This module aims to learn this correspondence and generate a "credibility" weight for each spatiotemporal position of the signal. By using convolution kernels of various sizes such as 3x3, 5x5, 7x7, etc., the network can capture amplitude change features of different scales at the same time, thus more accurately evaluating the signal quality.

[0034] The present application greatly improves the quality of the phase signal input to the subsequent network by element-wise multiplication of the separated phase signal and the above-mentioned gating signal.

[0035] Further, as shown in Figure 8 The calculation of the covariance matrix map specifically includes: A four-channel dynamic feature map is constructed, including: a current signal state map, a difference map between the current state and the initial state (reflecting long-term changes), a difference map of adjacent time points (reflecting short-term process noise), and a standard deviation map of the signal in the time dimension (reflecting overall volatility); The four-channel dynamic feature map is input into the Cholesky decomposition prediction network, specifically including: The four-channel dynamic feature map is spliced in the channel dimension, and the spliced feature map is input into the prediction network, which includes 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 in turn, which can effectively process spatiotemporal features; The output of the prediction network is output as a high-dimensional feature vector through channel dimension splitting; Based on the high-dimensional feature vector and the Cholesky component, a covariance matrix is constructed to obtain a covariance matrix map. Specifically, the high-dimensional feature vector is processed through dimension exchange and flattening to convert it into a one-dimensional vector, and the number of elements of the one-dimensional vector is equal to the number of parameters required by the Cholesky factor. Through a special reconstruction function, the reconstruction process is as follows: first, initialize a lower triangular zero matrix L, fill the vector elements into the lower triangular part of L, extract the diagonal elements and the strictly lower triangular part from the filled matrix L, apply the Softplus activation function to the extracted diagonal elements to ensure that all diagonal values are positive, reconstruct the regularized diagonal elements into a diagonal matrix, and then add the strictly lower triangular part to obtain the regularized lower triangular matrix L1. Finally, by calculating P = L1 * L1 transpose, the final stable and physically correct covariance matrix map is obtained.

[0036] The present application effectively guarantees the physical legality and numerical stability of the covariance matrix by introducing a covariance modeling mechanism based on Cholesky decomposition factor prediction: Since the covariance matrix must satisfy symmetry and semi-definiteness in theory, if the complete matrix is directly output by the neural network, it is easy to violate the above physical constraints due to parameter perturbation, and then cause training divergence or reasoning abnormality. The present application predicts the lower triangular Cholesky factor L, and reconstructs the covariance matrix in the form of P = L1 * L1, which naturally guarantees the symmetry and semi-definiteness of the output matrix from the network structure level, significantly improving the numerical robustness and convergence stability of the model in the whole training and reasoning process.

[0037] Greatly reduce parameter redundancy and improve learning efficiency: For an n-dimensional state space, the complete covariance matrix contains n elements, of which there is a lot of symmetric redundancy. The present application only needs to predict n(n+1) independent parameters (i.e. non-zero elements of Cholesky factor), which not only reduces the output dimension by nearly half, but also avoids the additional burden of forcing the network to learn the symmetric structure explicitly, thereby speeding up the convergence, reducing the risk of overfitting, and improving the parameter utilization efficiency.

[0038] Realize deep fusion of multi-scale dynamic context, and enhance the accuracy of uncertainty estimation: In order to accurately depict the time-varying uncertainty in the target motion process, the Cholesky predictor network designed by the present application not only depends on the current observation state, but also fuses four types of dynamic features, i.e. current signal state, long-term trend offset (difference from initial state), short-term process noise (adjacent frame difference) and time dimension fluctuation intensity (standard deviation map), as network input. This multi-source information fusion mechanism enables the model to simultaneously perceive the instantaneous change, historical evolution and overall fluctuation characteristics of the target, thereby generating a covariance representation that is more discriminative and physically consistent, significantly improving the reliability of subsequent feature extraction and classification decision.

[0039] Further, as shown in Figure 9 , feature extraction is performed on the covariance matrix graph to obtain a new feature map, including: Eigenvalue decomposition is performed on the covariance matrix of the covariance matrix graph; Trace, log determinant, anisotropy and principal eigenvector are calculated according to the decomposed eigenvalues and eigenvectors; The trace, log determinant, anisotropy and principal eigenvector are spliced and then layer normalized; The layer normalized feature map is input into a refiner, including 1x1 two-dimensional convolution, normalization, GELU activation function and 1x1 two-dimensional convolution, to obtain a new feature map.

[0040] The complete covariance matrix of the application contains a large amount of redundant information. Through eigenvalue decomposition, the information can be refined into several key, clearly physically meaningful scalars and vectors. For example, the "trace" represents the size of the overall uncertainty, the "determinant" represents the volume of the uncertainty ellipsoid, and the "anisotropy" describes the directionality of the uncertainty. These refined features are easier for the subsequent network to learn and utilize than the original matrix.

[0041] Further, as shown in Figure 10 TCN trajectory prediction includes: The up-sampled feature map is spliced in the channel dimension to form an input feature tensor to be processed; The input feature tensor is subjected to dimension splicing and flattening operation to convert it into a one-dimensional sequence suitable for time series modeling; The one-dimensional sequence is input into the time series convolutional network after linear layer processing, and the time series convolutional network includes N layers of TCN residual blocks, i is a loop index variable, marked as i=0 to i=N-1, and the feature sequence processed by the TCN residual block is output. Specifically, the data starts from the top TCN residual block and flows through each layer for processing. Each layer includes three core operations: first, one-dimensional convolution (convolution kernel size 3), the left padding length of the one-dimensional convolution operation is set to a value equal to the expansion rate, and the receiving domain of the convolution kernel is constrained not to exceed the left range of the current time step, so that the output of the convolution operation at each time step only depends on the input values of the current time and historical time. The expansion rate and padding value increase exponentially by level number, and the expansion coefficient of the i-th layer is 2 i ; then normalized; and finally through the GELU activation function; The feature sequence is output through the linear layer as a feature stream of the covariance and trajectory joint representation.

[0042] The application inputs the up-sampled feature map into the time series convolutional network (TCN) to predict the motion trajectory trend of the target, and through the use of causal convolution and exponentially increasing expansion rate, a huge receptive field can be obtained with fewer layers, thereby efficiently capturing the long-term sequential dependence, which is crucial for predicting the motion trajectory. Compared with RNN, TCN can be calculated in parallel, is more efficient, and does not have the problem of gradient disappearance.

[0043] In this embodiment, as shown in Figure 11 and Figure 12 The long-term global dependence modeling module includes: The feature fusion unit is used for feature fusion of two heterogeneous feature streams, and the unified fusion feature map is fused through the element-by-element addition operation; 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 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. 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; The specific processing procedure of the Stream Encoder is as follows: 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; 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. 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.

[0044] The output of the last Transformer block is subjected to layer normalization to obtain the final encoder output.

[0045] Hierarchical fusion and decision unit: a learnable CLS Token (Classification Token) is spliced at the beginning of each local patch sequence of the stream encoder output, and the sequences of the two streams are spliced in the feature dimension to form a joint sequence; the joint sequence is input into a cross-stream fusion Transformer module, in order to avoid sequence bias, two fusion operations (stream 1 + stream 2 and stream 2 + stream 1) are usually performed symmetrically, as shown in Figure 14 The cross-stream fusion results of the two directions are spliced again and sent to 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 condenses the essence information of the entire sequence. All extracted CLS Token sequences are sent to a fully connected layer (Final FC) for final classification decision, and after the dequantization operation, the original numerical range is restored, and finally the classification result Logits is output, as shown in Figure 15 .

[0046] The specific processing process of the Final FC is as follows: All extracted CLS Token sequences are processed in parallel, and the features of the six channels are processed separately; The six parallel classifiers process the separated features respectively, and then the original decision signals of multiple classifiers are converted into a unified and numerically stable comprehensive confidence vector through absolute value operation, stacking and summation operation.

[0047] The vector is normalized by class softmax to generate an output conforming to the probability distribution.

[0048] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between various embodiments can be referred to each other. For the device disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0049] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based anti-interference millimeter wave radar dynamic target detection system, characterized in that, The method comprises the following steps: a preprocessing module: for high-fidelity preprocessing of the original millimeter wave radar signal, removing static clutter, completing phase correction and outputting high signal-to-noise ratio dynamic signal after double-sided filter denoising; a feature extraction module: for front-end feature extraction of the high signal-to-noise ratio dynamic signal, separating amplitude and phase information, generating modulated phase flow and feature flow of covariance and trajectory joint representation in parallel, and outputting two heterogeneous feature flows; a long-time sequence global dependence modeling module: for feature fusion of the two heterogeneous feature flows, fused into a unified fusion feature map through element-by-element addition operation; input the fusion feature map into the Transformer network, and output the dynamic target detection result.

2. The system of claim 1, wherein, The preprocessing module comprises: a static clutter filtering unit: for combining the amplitude and phase components of the original millimeter wave radar signal into a complex matrix in the complex domain, calculating the average value of the complex matrix along the time dimension, and subtracting the average value from the complex matrix to obtain a zero-mean complex signal containing dynamic target information; a phase correction unit: for extracting the phase component from the zero-mean complex signal, and sequentially performing phase unwrapping processing and linear detrending processing on the phase component to obtain a corrected phase signal that is continuous and free of trend interference; a double-sided filtering unit: for arranging the corrected phase signal into a two-dimensional space-time matrix, performing double-sided filtering denoising on the two-dimensional space-time matrix through a two-dimensional bilateral filter, and outputting the high signal-to-noise ratio dynamic signal.

3. The system of claim 2, wherein, The double-sided filtering denoising comprises: S101: traversing each center pixel point in the two-dimensional space-time matrix; S102: for each center pixel point, selecting all neighborhood pixel points within a preset neighborhood range thereof; S103: respectively calculating the spatial distance weight and the pixel value similarity weight between the neighborhood pixel points and the center pixel point; S104: multiplying the spatial distance weight and the pixel value similarity weight element by element to obtain a combined weight corresponding to each neighborhood pixel point; S105: calculating the new value of the center pixel point after denoising by weighted averaging the pixel values of all neighborhood pixel points using the combined weights thereof; S106: determining whether all the center pixel points have been processed, if not, returning to S102 to process the next center pixel point; if yes, outputting the high signal-to-noise ratio dynamic signal.

4. The system of claim 1, wherein, The feature extraction module comprises: a signal normalization and separation unit: for normalizing and decomposing the high signal-to-noise ratio dynamic signal into amplitude information and phase information; a modulation unit: for performing amplitude gating processing on the amplitude information, and performing phase modulation on the phase information to generate a modulated phase flow; a covariance processing unit: for inter-channel fusion and spatial average pooling of the phase signal, calculating a covariance matrix graph, extracting a new feature graph from the covariance matrix graph, and upsampling and TCN trajectory prediction of the new feature graph to obtain the feature flow of covariance and trajectory joint representation.

5. The system of claim 4, wherein, The specific processing process of the modulation unit is: The amplitude information is extracted through three parallel convolution layers of 3x3, 5x5, and 7x7 convolution kernels to obtain features of different scales; The features of different scales are spliced in the channel dimension, and then batch normalization and GELU activation function processing are performed to obtain a multi-scale amplitude feature map; The multi-scale amplitude feature map is input into an 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 a gating signal map; The phase information is phase-modulated to obtain a phase feature map; The phase feature map and the gating signal map are subjected to element-by-element matrix multiplication to obtain the modulated phase flow.

6. The system of claim 4 or 5, wherein, The calculation of the covariance matrix map specifically includes: Constructing a four-channel dynamic feature map; The four-channel dynamic feature map is input into a Cholesky decomposition prediction network, which specifically includes: The four-channel dynamic feature map is spliced in the channel dimension, and the spliced feature map is input into a prediction network. The prediction network includes 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 in sequence; The output of the prediction network is output as a high-dimensional feature vector through channel dimension splitting; Based on the high-dimensional feature vector and the Cholesky component, a covariance matrix is constructed to obtain the covariance matrix map.

7. The system of claim 6, wherein, Feature extraction is performed on the covariance matrix map to obtain a new feature map, including: Eigenvalue decomposition is performed on the covariance matrix of the covariance matrix map; The trace, logarithmic determinant, anisotropy, and principal eigenvector are calculated according to the decomposed eigenvalues and eigenvectors; The trace, logarithmic determinant, anisotropy, and principal eigenvector are spliced and subjected to layer normalization; The layer-normalized feature map is input into a refiner, including 1x1 two-dimensional convolution, normalization, GELU activation function, and 1x1 two-dimensional convolution, to obtain the new feature map.

8. The system of claim 7, wherein, TCN trajectory prediction includes: The up-sampled feature map is spliced in the channel dimension to form an input feature tensor to be processed; Dimension splicing and flattening operations are performed on the input feature tensor to convert it into a one-dimensional sequence suitable for time series modeling; The one-dimensional sequence is input into a time series convolution network after linear layer processing. The time series convolution network includes multiple TCN residual blocks, and outputs a feature sequence processed by the TCN residual blocks; The feature sequence is output as a feature flow of the covariance and trajectory joint representation through a linear layer.

9. The system of claim 1, wherein, The long-time series global dependency modeling module includes: A feature fusion unit is used to fuse the two heterogeneous feature flows into a unified fusion feature map through element-by-element addition operation; A quantization slicing unit is used to quantize the fusion feature map and adopt a sliding window strategy to intercept an effective time period containing core dynamic signals on the time axis to obtain a sliced feature map; A parallel stream processing unit is 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 slicing method to obtain a local block sequence; The local block sequence is sent into a stream encoder to model long-range temporal dependencies within each stream; A hierarchical fusion and decision unit is used to splice a learnable CLS Token at the beginning of the local block sequence of each stream after the stream encoder outputs, and the sequences of the two streams are spliced in the feature dimension to form a joint sequence; the joint sequence is input into a cross-stream fusion Transformer module to perform bidirectional information interaction, CLS Token extraction, and output a dynamic target detection result through a fully connected layer.

Images