A physical constraint based anti-multipath millimeter wave radar DOA estimation method

Abstract

This invention belongs to the field of radar signal processing technology and relates to a physically constrained DOA estimation method for anti-multipath millimeter-wave radar. First, it obtains the sampling covariance matrix of the received signal from the millimeter-wave radar array and reconstructs it into a two-dimensional data matrix based on its real and imaginary parts. Next, a hierarchical training process is designed, first performing unsupervised training on a deep neural network model, followed by transfer learning fine-tuning. Then, the two-dimensional data matrix is ​​input into the trained deep neural network model, which outputs a spatial spectrum estimation map corresponding to the angular domain. Finally, peak search is performed on the spatial spectrum estimation map to determine the target's DOA estimate. This invention addresses the heavy reliance of existing deep learning methods on labeled datasets and, by utilizing a deep learning network model, achieves efficient and low-cost adaptation from general simulation environments to specific real-world scenarios, improving the feasibility of anti-multipath millimeter-wave radar DOA estimation.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology and relates to a physical constraint-based method for DOA estimation of anti-multipath millimeter-wave radar. Background Technology

[0002] Direction of arrival (DOA) estimation for millimeter-wave radar aims to determine the location of a signal source through array signal processing and is widely used in fields such as autonomous driving and intelligent sensing. However, in enclosed environments such as indoor spaces and tunnels, the multipath effect caused by signal reflection and superposition along multiple paths can severely degrade the algorithm's performance.

[0003] To cope with multipath interference, existing technologies mostly adopt deep learning-based DOA estimation methods. For example, Chinese Patent Publication No. CN113466845A describes a local search angle measurement method based on dual steering vectors to resist multipath interference. This method constructs a dual steering vector matrix that matches the terrain and estimates the angle of low elevation targets through local spectral peak search and interpolation optimization, thereby improving the angle measurement accuracy under multipath conditions.

[0004] Another Chinese patent publication, CN119270223A, describes a high-precision angle measurement method for low elevation angle multipath removal in meter-wave radar based on feature game theory. This method constructs a generative adversarial network and utilizes feature game theory optimization to enhance target echo characteristics. It then combines this with a classical positioning estimation algorithm for angle measurement, aiming to separate the target from multipath signals.

[0005] However, the existing technologies still have the following inherent drawbacks: 1. Existing DOA estimation relies on large-scale, high-quality labeled datasets for supervised training. In multipath scenarios, obtaining accurate DOA label data corresponding to complex environments is extremely difficult and costly. Without labeled data, supervised learning paradigms cannot work effectively.

[0006] 2. Existing deep learning models are usually black-box fitting models, which fail to effectively integrate the inherent physical constraints in array signal processing, such as data fidelity, spatial spectrum sparsity and structure. Therefore, it is difficult to distinguish between direct path and multipath interference in essence, resulting in the feature representations learned by them being not very effective against multipath interference, and having limited generalization ability and estimation accuracy. Summary of the Invention

[0007] In view of this, in order to solve the problems mentioned in the background technology, a physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar is proposed.

[0008] The objective of this invention can be achieved through the following technical solution: This invention provides a physical constraint-based method for estimating the DOA of millimeter-wave radar against multipath, comprising: obtaining the sampling covariance matrix of the received signal of the millimeter-wave radar array, and reconstructing the real and imaginary parts of the sampling covariance matrix into a two-dimensional data matrix.

[0009] Using a simulated multipath scene dataset generated by ray tracing, a deep neural network model is pre-trained in an unsupervised manner with multiple physical constraints. Then, using on-site calibration data collected by a millimeter-wave radar array for the current detection scene, the pre-trained deep neural network model is fine-tuned using transfer learning to obtain a fully trained deep neural network model.

[0010] The two-dimensional data matrix is ​​input into the trained deep neural network model, and the model outputs a spatial spectrum estimation map corresponding to the angular domain.

[0011] Peak search is performed based on the spatial spectrum estimation map to determine the target's direction of arrival (DOA) estimate.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention utilizes the simulated multipath scene dataset generated by ray tracing to construct an unsupervised loss function that integrates data fidelity constraints, sparsity constraints and structural constraints to pre-train the network. It does not require DOA angle labels of the real environment, but only forces the network to reconstruct the spatial spectrum estimation of the input sampling covariance matrix through the steering vector matrix to complete the learning, thus solving the problem of the serious dependence of existing deep learning methods on labeled datasets.

[0013] (2) This invention designs a hierarchical training process. After pre-training using a simulated multipath scene dataset, a transfer learning fine-tuning stage is introduced. The pre-trained model parameters are loaded and the encoder is frozen. Only a small amount of on-site calibration data collected in the current detection scene is used to train the subsequent parts of the network. This enables the deep learning network model to quickly learn and compensate for the clutter characteristics of a specific environment, achieving efficient and low-cost adaptation from a general simulation environment to specific real-world scenarios, and improving feasibility. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the execution flow of the method of the present invention.

[0016] Figure 2 This is a schematic diagram of the spatial spectrum estimation diagram of the present invention.

[0017] Figure 3 This is a technical effect diagram of the present invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 As shown, this invention provides a physical constraint-based anti-multipath millimeter-wave radar DOA estimation method, comprising: the millimeter-wave radar array receives a complex signal containing target range, velocity, and azimuth information, and its sampling covariance matrix is ​​composed of the complex correlation coefficients between each receiving channel, containing the spatial characteristics of the signal wave direction of arrival, which is the analytical data for DOA estimation. However, since the deep learning network models used for subsequent feature extraction or direction estimation are usually based on real-number domain operations, they cannot directly process the complex form of the sampling covariance matrix. Therefore, the first step of the method needs to obtain the sampling covariance matrix and convert the complex correlation information therein into structured data that can be effectively processed in the real-number domain. Specifically: S11. Obtain the sampling covariance matrix of the millimeter-wave radar array received signal, including: performing a fast Fourier transform of the range dimension and Doppler dimension on the raw ADC data collected by each receiving channel of the millimeter-wave radar array, wherein the range dimension fast Fourier transform is performed on the sampling point sequence within a single transmission pulse period, converting the time-domain sampling signal into a frequency-domain representation, and corresponding the frequency points to different range units. The Doppler Fast Fourier Transform is performed on a sequence of complex range images of the same range cell within multiple emission pulse periods, and is used to analyze target velocity information.

[0020] After the above two fast Fourier transforms, range-Doppler two-dimensional data corresponding to each receiving channel are generated. This data represents the amplitude and phase of the corresponding position signal under different range units and different combinations of Doppler units.

[0021] The real power value is obtained by taking the square of the modulus of the distance-Doppler two-dimensional data for each receiving channel. The real power values ​​of all receiving channels at the same distance cell index and the same Doppler cell index are summed to achieve incoherent accumulation processing of the distance-Doppler two-dimensional data of each receiving channel.

[0022] The constant false alarm rate (CFAR) detection algorithm is used to detect targets in the accumulated power spectrum. As a specific embodiment of the CFAR detection algorithm, the cell-average CFAR detection algorithm is used to perform the following operations: a two-dimensional sliding detection window is defined and the center of this window is set as the cell to be detected. It is worth noting that the cell to be detected corresponds to a specific distance cell index and Doppler cell index.

[0023] To avoid the impact of target energy leakage on noise estimation, a rectangular ring-shaped area formed by several rows and columns of units adjacent to the unit under test is used as a protection area, and units outside the protection area are used as reference units. The average power value of all reference units is calculated, and the average value is used as the average noise power estimate of the local background where the unit under test is located.

[0024] Based on the constant false alarm probability preset in the system design, for example Based on the number of reference units Q, the difference between the negative Q root of the constant false alarm probability and 1 is calculated. Then, the difference is multiplied by the number of reference units to determine the threshold factor.

[0025] The average noise power estimate is multiplied by a threshold factor to obtain a dynamic preset noise power threshold.

[0026] If the power value of the unit to be detected is greater than or equal to the preset noise power threshold, then a target is determined to exist within the unit to be detected, and it is marked as a target unit. The corresponding distance unit index and Doppler unit index are recorded. Otherwise, it is determined that there is no target.

[0027] For each detected target cell, a complex signal snapshot in the corresponding range-Doppler cell coordinates is extracted from all receiving channels and arranged into a column vector according to the channel order, which serves as the array observation vector.

[0028] Calculate the Euclidean norm of the array observation vector, and divide each element in the vector by the Euclidean norm to complete the amplitude normalization process.

[0029] based on The sampling covariance matrix consists of array observation vectors from the same target cell, which have undergone amplitude normalization. Calculated using the following formula: .

[0030] In the formula, Indicates the first The conjugate transpose of the array observation vectors Indicates the first array observation vectors, The sample covariance matrix represents the total number of snapshots counted. The dimension is , The matrix represents the number of receiving channels, with its rows and columns corresponding to the indices of the receiving channels. The matrix elements correspond to the complex correlation estimates between the signals of each receiving channel, reflecting the spatial correlation structure of the signals.

[0031] The matrix multiplication (outer product) of a vector and its conjugate transpose is the instantaneous correlation matrix, which reflects the instantaneous correlation coefficient between the amplitude and phase relationships of the signals in each receiving channel of the array at a single snapshot moment.

[0032] This means summing the instantaneous correlation matrices corresponding to all snapshots to smooth out the effects of instantaneous noise, interference, or random fluctuations that may be included in a single snapshot, thereby obtaining a stable inter-channel correlation estimate.

[0033] This means dividing the summation result by the total number of snapshots to obtain the sampling covariance matrix under the final unbiased estimate.

[0034] The above formula calculates the covariance matrix, which represents the spatial correlation structure of the signal, by collecting spatial signal snapshots of the target at multiple times and using statistical averaging. This matrix contains the target's direction of arrival (DOA) information, providing a mathematical basis for subsequent DOA estimation.

[0035] S12. Reconstruct the real and imaginary parts of the sampling covariance matrix into a two-dimensional data matrix, including: arranging the real parts of all elements in the sampling covariance matrix according to their row and column positions in the original matrix to form the real part matrix; arranging the imaginary parts of all elements in the same position to form the imaginary part matrix. Both the real and imaginary parts matrices are real matrices of the same dimension.

[0036] The real and imaginary matrices are concatenated along the column direction to form a two-dimensional data matrix with the number of rows equal to the number of receiving channels and the number of columns equal to twice the number of receiving channels.

[0037] The trace index of the sampling covariance matrix is ​​obtained by summing the main diagonal elements. All elements in the two-dimensional data matrix are then divided by the trace index to obtain the standardized two-dimensional data matrix.

[0038] It should be noted that the first step involves splitting the complex covariance matrix into real and imaginary parts and concatenating them, completely preserving all real and imaginary information of the original matrix, thus providing a complete real-domain input for subsequent processing. By standardizing the concatenated matrix based on the matrix trace, the input data is made scale-invariant, thereby focusing the analysis on the spatial correlation structure of the signal represented by the relative relationships of matrix elements, improving the ability of subsequent processing steps to identify the signal itself.

[0039] To estimate the signal spatial spectrum from the standardized two-dimensional data matrix obtained in the first step, this invention employs a deep neural network model for end-to-end nonlinear mapping. The deep neural network model includes an encoder, a bottleneck layer, a decoder, and an output layer connected in sequence. The model processing flow is as follows: Encoder: Performs multi-level convolution and downsampling processing on the input two-dimensional data matrix, extracts and outputs spatial feature maps at different levels, where the horizontal and vertical axes of the spatial feature maps still correspond to the row and column dimensions of the two-dimensional data matrix.

[0040] Bottleneck layer: Receives the spatial feature map output from the final layer of the encoder and performs feature integration operations using dilated convolution at the final layer scale. Specifically, by setting a dilation rate greater than 1, the convolution kernel can cover a larger receptive field without increasing the number of parameters or reducing the resolution, thus integrating the potentially scattered correlation contextual information from multipath signals in the spatial domain and outputting an integrated spatial feature map. This aims to enhance the model's ability to model long-range dependencies in complex multipath structures, providing a more aggregated feature representation for subsequent decoding and angle estimation.

[0041] The decoder upsamples the spatial feature map output from the bottleneck layer, gradually restoring the spatial dimension. It then performs channel-wise concatenation and fusion of the upsampled result with the corresponding layer's spatial feature map output from the encoder. This involves connecting the upsampled features from the decoder with features of the same scale from the encoder along the feature channel dimension, forming a composite feature map with an increased number of channels. The aim is to combine the details captured in the encoding stage with the semantic information restored in the decoding stage, thereby utilizing both low-level local cues and high-level global context when reconstructing the spatial spectrum, thus improving estimation accuracy.

[0042] Output layer: The spatial feature map output by the decoder is processed by an activation function (such as Sigmoid or Softmax), which maps and reshapes it into a spatial spectrum estimation vector. The length of the spatial spectrum estimation vector is equal to the total number of discrete angle samples within a predefined detection angle range. The value at each position in the vector represents the relative intensity of the signal energy on the corresponding discrete angle sample, and is used to represent the confidence or probabilistic estimate of the target signal energy from the corresponding discrete angle direction.

[0043] The predefined detection angle range is set according to the specific application scenario (such as ±60° forward detection of vehicle radar) to cover the required azimuth angle range. In addition, the spatial spectrum estimation vector can be converted into a spatial spectrum estimation map in the corresponding angle domain to facilitate subsequent peak finding and DOA estimation.

[0044] However, deep learning-based estimation methods rely on training with data bearing precise angle labels, while obtaining such real-world data in practical applications is extremely costly and difficult to cover complex and variable multipath scenarios. Therefore, in the second step of this invention, a stage strategy combining unsupervised pre-training and on-site fine-tuning is adopted during the model training phase to improve the model's accuracy and generalization ability under limited real-world data conditions. Specifically: S21. Using a simulated multipath scenario dataset generated by ray tracing, the implementation process is as follows: A virtual three-dimensional scene model consistent with the current detection scene environment is established, including the fixed installation position of the millimeter-wave radar, the target's trajectory, and information on typical reflectors in the environment, such as walls, ground, and metal obstacles.

[0045] An electromagnetic ray tracing algorithm is applied to emit several rays from the phase center of the transmitting antenna into the scene space. The propagation path of each ray is tracked after being reflected or scattered by the target scatterer or environmental reflector until it returns to the receiving array.

[0046] The electromagnetic waves arriving at the target from the transmitting antenna via different paths, and the signal parameters returning to the receiving array via each reflection path, are calculated. These parameters include propagation delay, Doppler frequency shift, complex amplitude attenuation, and angle of arrival. The propagation delay is calculated by dividing the total length of the signal propagation path determined by ray tracing by the speed of light. The Doppler frequency shift is calculated by substituting the instantaneous radial velocity of the target or reflector relative to the radar into the Doppler effect formula. The complex amplitude attenuation is the sum of the path free-space propagation loss and the calibration reflection coefficient loss when reflected by the reflector. The angle of arrival is calculated analytically based on the final direction of the ray reaching the receiving array, combined with the array's geometric coordinate system, determining its azimuth and elevation angles relative to the array's normal direction.

[0047] Based on the radiation pattern function and geometric distribution of the millimeter-wave radar array antenna, for each signal path, the complex radiation pattern function values ​​of the transmitting and receiving antennas are queried according to its arrival angle. The phase delay is calculated in combination with the geometric position of the array elements. The original array data containing the direct path signal, the first reflection path signal, and the multiple reflection path signals are synthesized. Gaussian white noise with corresponding power is superimposed according to the preset signal-to-noise ratio to simulate the real receiving environment, forming a simulated multipath scenario dataset.

[0048] It should be noted that the complete process of ray tracing, parameter calculation, and acquisition of raw array data described above can be achieved using professional electromagnetic simulation or wireless channel simulation software (such as Altair Feko). Those skilled in the art only need to configure the aforementioned 3D scene model, material properties, and radar parameters; the simulation software can then directly output the results.

[0049] S22. Pre-training the deep neural network model in an unsupervised manner with multiple physical constraints, including: S221. For the simulated multipath scene dataset, calculate the sampling covariance matrix and reconstruct it into a two-dimensional data matrix as unlabeled pre-training samples.

[0050] S222. Construct an unsupervised loss function that integrates multiple physical constraints, wherein the unsupervised loss function is obtained by linearly weighting data fidelity constraints, sparsity constraints, and structural constraints.

[0051] The data fidelity constraint term is used to ensure that the covariance of the output spectrum reconstruction is consistent with the input covariance. Its execution logic is as follows: a steering vector matrix of the millimeter-wave radar array is predefined. The columns of the steering vector matrix are composed of steering vectors corresponding to discrete angle samples covering a predefined detection angle range, and the rows are composed of the receiving channels of the millimeter-wave radar array.

[0052] The spatial spectrum estimation vector output by the current iteration of the deep neural network model is converted into a diagonal matrix. Specifically, each value in the spatial spectrum estimation vector is used as the value in the corresponding position on the diagonal, and the remaining values ​​on the main line are set to 0.

[0053] Multiply the diagonal matrix by the conjugate transpose of the guiding vector matrix to obtain the reconstructed covariance matrix under the current iteration output of the deep neural network model.

[0054] The two-dimensional data matrix of the current pre-trained sample is restored to the original covariance matrix, and the Frobenius norm between it and the reconstructed covariance matrix is ​​calculated as the data fidelity loss term.

[0055] The sparsity constraint term is the L1 norm of the spatial spectrum estimation vector output by the deep neural network model, which is used to make the spatial spectrum exhibit a sharp main lobe and suppress spurious peaks.

[0056] The structural constraint term is the result of calculating the second-order difference norm of the spatial spectrum estimation vector output by the deep neural network model. It is used to smooth the envelope of the spatial spectrum estimation vector, thereby suppressing spurious fluctuations and spikes caused by multipath interference or noise.

[0057] Furthermore, the above linear weight allocation follows the following logic: the data fidelity constraint has the highest weight, followed by the sparsity constraint, and the structure constraint has the lowest weight. For example, the weight of the data fidelity constraint is set to 1.0, the weight of the sparsity constraint is 0.2, and the weight of the structure constraint is 0.05.

[0058] S223. Input the pre-trained samples into the deep neural network model to obtain the output spatial spectrum estimation vector and use it to calculate the unsupervised loss value.

[0059] S224. Based on the unsupervised loss value, update the parameters of the deep neural network model using the gradient backpropagation algorithm. Specifically, starting from the final unsupervised loss value, calculate the partial derivative of the unsupervised loss value with respect to each updatable parameter along the deep neural network model. Subtract the product of the partial derivative and the pre-calibrated learning rate (e.g., 0.001) from the current value of each updatable parameter.

[0060] Repeat training until any of the following stopping conditions are met: (1) Loss convergence: The absolute value of the difference between the average unsupervised loss value calculated in the current training round and the average loss value of the previous 10 training rounds is continuously less than a preset threshold, for example .

[0061] (2) Reach the predetermined maximum number of training rounds, such as 500 rounds.

[0062] The pre-training of the deep neural network model is completed when any one of the conditions is met.

[0063] It should be explained that the Frobenius norm, L1 norm, and second-order difference norm are mathematical tools and are common knowledge in this field; their general calculation process will not be elaborated here.

[0064] Specific examples of updatable parameters within a deep neural network model include, but are not limited to: convolutional layers in each level of the encoder, dilated convolutional layers in the bottleneck layer, transposed convolutional layers of the sampling operation, and weight parameters and bias vectors of the output layer.

[0065] S23. Using the field calibration data collected by the millimeter-wave radar array for the current detection scenario, the pre-trained deep neural network model is fine-tuned using transfer learning to obtain a trained deep neural network model. The fine-tuning process includes: calculating the sampling covariance matrix of the field calibration data collected in the current detection scenario, reconstructing it into a two-dimensional data matrix to obtain fine-tuning training samples. The field calibration data refers to finite-scale, mechanistically constrained measured data collected using a millimeter-wave radar array deployed in an actual detection scenario. A typical acquisition method is to place standard reflectors or cooperative targets at several known fixed locations in the scenario and control the radar to detect them. This process only requires knowing the approximate azimuth range of the target (e.g., within a ±30° sector directly in front of the radar) and does not require high-precision angle measurements. The purpose is to obtain data samples of the actual channel response and multipath characteristics in the real scenario.

[0066] Load the parameters of the pre-trained deep neural network model and freeze the weight parameters of the first preset number of convolutional layers responsible for feature extraction in the encoder, such as the first 3 layers.

[0067] Using the two-dimensional data matrix in the fine-tuning dataset as input, the corresponding rough target orientation interval as the supervision signal, and the sum of the energy estimates on all angle samples outside the rough orientation interval as the regional supervision loss value.

[0068] The sum of the unsupervised loss value used in pre-training and the supervised loss value in this region is used as the overall loss value in the fine-tuning stage. The parameters of the bottleneck layer, decoder, and output layer of the deep neural network model are updated only through the gradient backpropagation algorithm.

[0069] When the decrease in the unsupervised loss function during consecutive preset training rounds is consistently below a set threshold, the magnitude of the set threshold can be exemplarily defined as follows: Stop fine-tuning and save the model parameters.

[0070] The first and second steps of this invention have completed the standardization processing of the array signal and trained a deep neural network model capable of learning complex mapping relationships from the data. Therefore, the third step inputs the two-dimensional data matrix into the trained deep neural network model, and utilizes the feature extraction and nonlinear transformation capabilities learned by the model to generate a spatial spectrum estimation map in a predefined angular domain, which intuitively reflects the distribution of signal energy in different directions.

[0071] To extract the spectral peak positions corresponding to the real signal source from the spatial spectrum estimation map, the fourth step of this invention performs peak search based on the spatial spectrum estimation map to determine the target's direction of arrival (DOA) estimate. Specifically, this includes: performing a local maximum search on the spatial spectrum estimation vector output by the deep neural network model to identify all candidate peaks whose energy intensities are greater than a preset energy intensity detection threshold. The preset energy intensity detection threshold can be set as the sum of the mean of all current energy intensities and three times the standard deviation.

[0072] If multiple candidate peaks exist, a judgment is made based on the spatial smoothness prior covered by the structural constraints: if the absolute value of the angular interval between two peaks and the expected mirror angle offset generated by the dominant reflective structure of the scene (such as the ground) is less than 1 degree, and their energy ratio is within the range of 0.1-0.7, then multipath mirror peaks belonging to the same signal source are merged. The expected radial angle offset can be obtained through a single measurement or calibration during the scene deployment phase and input as a fixed parameter into the judgment logic.

[0073] For candidate peak pairs that are judged to be multipath mirrors, only the peak with the largest energy value is retained.

[0074] For the remaining candidate peaks that are not judged as multipath mirrors, a preset angular resolution interval is used as a window, for example, it can be set to 1.5 to 2 times the theoretical resolution specified in the millimeter-wave radar array technical specifications. The candidate peaks that are spatially adjacent are suppressed by non-maximum, and the peak with the highest energy is retained in each window.

[0075] For each retained peak, within its corresponding angular neighborhood, the centroid method is used to calculate the weighted average of the angles, with the energy values ​​of each angular sample in the neighborhood as weights, to obtain the target's direction of arrival (DOA) estimate.

[0076] It should be noted that the above multipath image discrimination method is mainly for single-reflection image scenarios generated by a dominant strong reflective surface, such as the ground or a large wall. This scenario is the most common and significant type of multipath in practical applications, and this method has excellent discrimination performance under this condition.

[0077] For environments with multiple reflections or more complex propagation paths, as a scalable implementation, a typical multipath angle-energy pattern library for this scenario can be pre-constructed through electromagnetic simulation or on-site measurements. During actual discrimination, candidate peak combinations are matched with the pattern library to achieve more accurate multipath identification and merging. The expansion method can be selectively implemented according to actual needs.

[0078] Furthermore, the exemplary values ​​(such as angle tolerance and energy ratio range) involved in the above multipath mirror discrimination method are typical values ​​provided for illustrating the principle. In practical applications, those skilled in the art can adaptively adjust the example values ​​according to specific scenario characteristics and accuracy requirements, all of which fall within the scope of conventional custom operations within the scope of the method of this invention.

[0079] After determining the DOA estimation result of the target, the present invention further includes the step of generating a three-dimensional point cloud of the target: using the determined DOA estimation value as the azimuth or pitch angle information of the target.

[0080] The constant false alarm rate (CFAR) detection algorithm uses distance and Doppler information obtained during the detection process. The distance information is used to determine the radial length of the target in the spatial vector, and the Doppler information is used to determine the radial velocity of the target.

[0081] Distance and angle data in polar coordinates are converted into X, Y, and Z coordinates in Cartesian coordinates. Combined with timestamp information, a target point cloud containing position and velocity features is rendered in three-dimensional space and output to the perception backend for target tracking and recognition.

[0082] Figure 2The invention demonstrates a spatial spectrum estimation map generated after processing a multipath scene containing a direct-path target and its ground reflection image using the method of the present invention. The horizontal axis of the spatial spectrum estimation map represents the azimuth angle, the vertical axis represents the normalized energy intensity, and the energy distribution is represented by a heatmap color map.

[0083] The image shows a sharp, highly concentrated main peak, corresponding to a direct-path target; secondary peaks with lower energy and symmetrical angles may appear on either side, corresponding to multipath mirror reflection signals. This visually demonstrates the ability of deep neural network models to effectively extract target orientation information even under multipath interference.

[0084] Meanwhile, to verify the effectiveness of the proposed method, experiments were conducted on a set of simulated multipath scene datasets generated by ray tracing. The experiments used an array of observation vectors containing 20 snapshots, and model training and testing were performed according to the first to fourth steps described in this invention.

[0085] Figure 3 This diagram illustrates the performance convergence process of a deep neural network model during the pre-training phase. The horizontal axis represents the number of training epochs, and the vertical axis represents the values ​​of various evaluation metrics. The diagram includes three curves: the precision curve, which measures the proportion of samples judged as positive by the model that are actually positive, reflecting the accuracy of the estimation results.

[0086] Recall curve: measures the proportion of all true positive samples that are correctly identified by the model, reflecting the ability to detect the target signal.

[0087] The F1 score curve is the harmonic mean of precision and recall, and is a comprehensive indicator of model performance.

[0088] The figure shows that all three curves rise rapidly with increasing training epochs and then gradually stabilize. The red dot in the figure indicates the peak value of the F1 score curve during the entire training process, which is 96.52%, corresponding to the 134th training epoch. This result demonstrates that the unsupervised pre-training method described in this invention enables the model to efficiently learn the ability to accurately estimate signal direction from multipath array data without the need for real angle labels, ultimately achieving excellent overall performance on the test set.

[0089] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications or additions should fall within the protection scope of the present invention.

Claims

1. A physically constrained method for estimating the DOA of anti-multipath millimeter-wave radar, characterized in that, include: Obtain the sampling covariance matrix of the signal received by the millimeter-wave radar array, and reconstruct the real and imaginary parts of the sampling covariance matrix into a two-dimensional data matrix; Using a simulated multipath scene dataset generated by ray tracing, a deep neural network model is pre-trained in an unsupervised manner with multiple physical constraints. This includes: calculating the sampling covariance matrix and reconstructing it into a two-dimensional data matrix for the simulated multipath scene dataset, which serves as unlabeled pre-training samples. An unsupervised loss function integrating multiple physical constraints is constructed, which is obtained by linear weighting of data fidelity constraint, sparsity constraint and structural constraint. Input the pre-trained samples into the deep neural network model to obtain the output spatial spectrum estimation vector and use it to calculate the unsupervised loss value. Based on the unsupervised loss value, the parameters of the deep neural network model are updated using the gradient backpropagation algorithm. The training is repeated until the unsupervised loss value converges to a preset range or reaches a predetermined number of iterations, thus completing the pre-training of the deep neural network model. The pre-trained deep neural network model was fine-tuned using transfer learning by utilizing the on-site calibration data collected by the millimeter-wave radar array for the current detection scenario, resulting in a fully trained deep neural network model. The deep neural network model comprises an encoder, a bottleneck layer, a decoder, and an output layer connected in sequence. The model processing flow is as follows: Encoder: Performs multi-level convolution and downsampling on the input two-dimensional data matrix to extract and output spatial feature maps at different levels; Bottleneck layer: Receives the spatial feature map output from the final layer of the encoder, and performs feature integration operation using dilated convolution at the final layer scale to output the integrated spatial feature map; Decoder: Upsamples the spatial feature map output by the bottleneck layer to gradually restore the spatial dimension, and then performs channel splicing and fusion of the upsampled result with the spatial feature map output by the corresponding layer in the encoder; Output layer: The activation function is applied to process the spatial feature map output by the decoder, generating a spatial spectrum estimation vector that corresponds one-to-one with discrete angle samples within a predefined detection angle range, which is then labeled as the spatial spectrum estimation map of the predefined angle domain; The two-dimensional data matrix is ​​input into the trained deep neural network model, and the model outputs a spatial spectrum estimation map corresponding to the angle domain. Peak search is performed based on the spatial spectrum estimation map to determine the target's direction of arrival (DOA) estimate.

2. The physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar according to claim 1, characterized in that, The acquisition of the sampling covariance matrix of the millimeter-wave radar array received signal includes: Fast Fourier transforms of the range dimension and Doppler dimension are performed on the raw ADC data acquired by each receiving channel of the millimeter-wave radar array to generate range-Doppler two-dimensional data corresponding to each receiving channel. The range-Doppler two-dimensional data from each receiving channel are incoherently accumulated, and a constant false alarm rate (CFAR) detection algorithm is used to detect the target in the accumulated power spectrum. The average noise power of the reference cell is calculated by sliding the detection window, and the distance-Doppler cell coordinates of the target are determined by combining the preset noise power threshold. For each detected target cell, a complex signal snapshot in the corresponding range-Doppler cell coordinates is extracted from all receiving channels to form an array observation vector; The array observation vector is normalized in magnitude, and the sampling covariance matrix is ​​calculated based on multiple snapshots. The rows and columns of the sampling covariance matrix correspond to the indices of the receiving channels, and the matrix elements correspond to the complex correlation estimates between the signals of each receiving channel.

3. The physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar according to claim 1, characterized in that, The step of reconstructing the real and imaginary parts of the sampling covariance matrix into a two-dimensional data matrix includes: The sampling covariance matrix is ​​split into a real part matrix and an imaginary part matrix; The real part matrix and the imaginary part matrix are concatenated along the column direction to form a two-dimensional data matrix with the number of rows equal to the number of receiving channels and the number of columns equal to twice the number of receiving channels; The trace index of the sampling covariance matrix is ​​obtained by summing the main diagonal elements. All elements in the two-dimensional data matrix are then divided by the trace index to obtain the standardized two-dimensional data matrix.

4. The physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar according to claim 1, characterized in that, The simulated multipath scene dataset generated using ray tracing includes: A virtual three-dimensional scene model is established, including the radar installation location, target movement trajectory, and information on typical reflectors in the environment, including walls, ground, and metal obstacles. An electromagnetic ray tracing algorithm is applied to calculate the electromagnetic waves that travel from the transmitting antenna to the target via different paths, as well as the signal parameters that return to the receiving array via each reflection path. These parameters include propagation delay, Doppler frequency shift, complex amplitude attenuation, and angle of arrival. Based on the radiation pattern function and geometric distribution of the millimeter-wave radar array antenna, the original array data containing direct path signals, primary reflection path signals, and multiple reflection path signals is synthesized, and Gaussian white noise is superimposed to simulate the real receiving environment, forming a simulated multipath scenario dataset.

5. The physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar according to claim 1, characterized in that, The data fidelity constraints include: A predefined steering vector matrix for a millimeter-wave radar array is defined. The columns of the steering vector matrix are composed of steering vectors corresponding to discrete angle samples covering a predefined detection angle range, and the rows are composed of receiving channels of the millimeter-wave radar array. The spatial spectrum estimation vector output by the current iteration of the deep neural network model is converted into a diagonal matrix. The diagonal matrix is ​​then multiplied by the conjugate transpose of the leading vector matrix to obtain the reconstruction covariance matrix under the current iteration output of the deep neural network model. The two-dimensional data matrix of the current pre-trained sample is restored to the original covariance matrix, and the Frobenius norm between it and the reconstructed covariance matrix is ​​calculated as the data fidelity loss term.

6. The physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar according to claim 1, characterized in that, The sparsity constraints and structural constraints include: Calculate the L1 norm and second-order difference norm of the spatial spectrum estimation vector output by the deep neural network model, respectively. Use the L1 norm result as a sparsity constraint term and the second-order difference norm result as a structural constraint term.

7. The physical constraint-based DOA estimation method for anti-multipath millimeter-wave radar according to claim 1, characterized in that, The fine-tuning of the pre-trained deep neural network model using transfer learning includes: Calculate the sampling covariance matrix of the field calibration data collected in the current detection scenario, and reconstruct it into a two-dimensional data matrix to obtain fine-tuning training samples; Load the parameters of the pre-trained deep neural network model and freeze the weight parameters of the pre-set number of convolutional layers in the encoder responsible for feature extraction. Using the two-dimensional data matrix in the fine-tuning dataset as input, and the corresponding rough target orientation interval as the supervision signal, the regional supervision loss value is calculated. The sum of the unsupervised loss value used in pre-training and the regional supervision loss value is used as the overall loss value in the fine-tuning stage. The parameters of the bottleneck layer, decoder and output layer of the deep neural network model are updated only through the gradient backpropagation algorithm. When the decrease in the unsupervised loss function is lower than the set threshold in consecutive preset training rounds, fine-tuning is stopped and the model parameters are saved.

8. The DOA estimation method for anti-multipath millimeter-wave radar based on physical constraints according to claim 1, characterized in that, The determined direction-of-arrival (DOA) estimate of the target includes: Local maxima search is performed on the spatial spectrum estimation vector output by the deep neural network model to identify all candidate peaks whose energy intensities are greater than the preset energy intensity detection threshold; If there are multiple candidate peaks, then based on the spatial smoothing prior covered by the structural term constraints, multipath mirror peaks belonging to the same signal source are identified and merged. For candidate peak pairs that are judged to be multipath mirror images, only the peak with the largest energy value is retained; For the remaining candidate peaks that are not judged as multipath mirrors, non-maximum suppression is performed on the candidate peaks that are spatially adjacent, using a preset angular resolution interval as a window, and the peak with the highest energy is retained in each window. For each retained peak, the angle estimate is calculated in its corresponding angular neighborhood using the centroid method, which serves as the target's direction of arrival (DOA) estimate.

Images