Millimeter wave radar cross-environment perception generalization method and system based on course learning

Abstract

The invention provides a millimeter-wave radar cross-environment perception generalization method and system based on course learning, and relates to the technical field of millimeter-wave radar perception, and the method comprises the steps: carrying out the constant false alarm rate detection of a distance velocity spectrogram obtained through preprocessing, and carrying out the interception to obtain a most easy sample data set; carrying out course learning stage progressive learning on the most likely sample data set to obtain a training stage total number and a training stage sample range of a millimeter wave radar cross-environment perception generalization model constructed by introducing a self-adaptive feature map down-sampling mechanism, and training the millimeter wave radar cross-environment perception generalization model; obtaining a trained millimeter wave radar cross-environment perception generalization model; and inputting to-be-identified millimeter-wave radar data into the millimeter-wave radar cross-environment perception generalization model, and outputting an environment perception identification result of the to-be-identified millimeter-wave radar data. According to the method, millimeter wave radar perception has cross-environment generalization capability, and the identification precision in an unknown scene is improved.

Description

Technical Field

[0001] This invention relates to the technical field of millimeter-wave radar sensing, and in particular to a method and system for generalizing millimeter-wave radar cross-environment sensing based on curriculum learning. Background Technology

[0002] Millimeter-wave radar has become a hot technology in wireless sensing research due to its stability and privacy protection advantages in scenarios such as medical health, smart homes and autonomous driving. Compared with traditional visual sensors, millimeter-wave radar can operate stably in low light, occlusion or bad weather conditions, while avoiding privacy leakage problems caused by high-resolution images. Therefore, it is particularly suitable for deployment in dynamic, complex and privacy-sensitive environments.

[0003] In recent years, most millimeter-wave radar domain generalization methods are based on traditional data augmentation or general strategies from the computer vision field. These methods have poor adaptability to the unique texture sparsity and physical meaning limitations of radar images, resulting in limited generalization effects. For example, existing technologies disclose millimeter-wave radar perception domain generalization methods based on data augmentation, but these methods can only expand data within a limited scope and are difficult to overcome the noise effects of diverse real-world environments, thus limiting their flexibility and adaptability. Existing technologies also disclose millimeter-wave radar perception conventional domain generalization methods adapted to the traditional computer image domain, with feature extraction based on the coordinated training between several classifiers and feature extractors, suitable for RGB images. However, given the sparsity and abstract nature of millimeter-wave radar images, these methods often fail to achieve good cross-environment generalization performance. Furthermore, when existing millimeter-wave radar images are deployed across different scenes, changes in the relative distance between the human body and the radar often cause significant vertical shifts in the spectral map. This shift not only causes spatial displacement of important semantic regions but also disrupts the feature extraction effect under a fixed downsampling window structure, resulting in severe semantic loss and consequently low target recognition accuracy. Summary of the Invention

[0004] To address the issues of lack of cross-environment generalization capability and low recognition accuracy in unknown scenarios in existing millimeter-wave radar sensing technologies, this invention proposes a millimeter-wave radar cross-environment sensing generalization method and system based on course learning. This method enables millimeter-wave radar sensing to have cross-environment generalization capability and improves recognition accuracy in unknown scenarios.

[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: A curriculum-based method for cross-environmental perception generalization of millimeter-wave radar includes the following steps: S1. Preprocess the raw millimeter-wave radar data to obtain the range-velocity spectrum; S2. Perform constant false alarm rate detection on the distance-velocity spectrum to obtain the easiest sample dataset; S3. Perform progressive learning from easy to difficult stages on the easiest sample dataset to obtain the total number of training stages and the range of training stage samples for the millimeter-wave radar cross-environment perception generalization model constructed by introducing an adaptive feature map downsampling mechanism. S4. Using the total number of training phases and the sample range of the training phases, train the millimeter-wave radar cross-environment perception generalization model to obtain the trained millimeter-wave radar cross-environment perception generalization model. S5. Input the millimeter-wave radar data to be identified into the millimeter-wave radar cross-environmental perception generalization model, and output the environmental perception identification result of the millimeter-wave radar data to be identified.

[0006] Preferably, the preprocessing of the raw millimeter-wave radar data to obtain the range-velocity spectrum includes: S101. Perform a discrete Fourier transform on the original millimeter-wave radar data in the slow time dimension to obtain velocity dimension information; S102. Perform a discrete Fourier transform on the original millimeter-wave radar data in the fast time dimension to obtain range dimension information; S103. Merge the velocity dimension information and the distance dimension information to obtain the distance-velocity map.

[0007] Preferably, the preprocessing of the raw millimeter-wave radar data further yields a range-angle spectrum, including: S111. Using an array structure of multiple receiving antennas, beamforming is performed on the initial millimeter-wave radar data at different angles to obtain angular dimension information; S112. Merge the angular dimension information and the distance dimension information to obtain the distance-angle spectrum.

[0008] Preferably, the distance dimension information in the distance-velocity spectrum includes several distance units, and the step of performing constant false alarm rate detection on the distance-velocity spectrum to obtain the easiest sample dataset includes: S201. Obtain the detection threshold for each distance cell; S202. Perform a unit-by-unit scan on several distance units, and mark the distance units whose energy exceeds the detection threshold corresponding to the current distance unit as active units, thereby obtaining several active segments that are interrupted due to noise interference. S203. Fill in the gaps in the activity segment to extract the longest continuous activity interval; S204. Record the start and end indices of the continuous activity interval, extract the corresponding sub-regions from the distance-velocity spectrum based on the start and end indices, and define the sub-regions as the easiest sample dataset. The calculation expression for the easiest sample dataset is as follows:

[0009] in, The distance-velocity spectrum is represented by the following coordinates: The pixel value at that time, This represents the width of the entire distance velocity spectrum. Therefore for lower bound and The upper bound is the range.

[0010] Preferably, the progressive learning process of the easiest sample dataset, from easy to difficult, includes: S301. Divide the training phases into multiple training phases based on the region size of the most easily sampled dataset to obtain the total number of training phases; S302. Based on the total number of training stages, introduce sample regions affected by multipath interference into the region of the easiest sample dataset in a progressive manner from easy to difficult, thereby obtaining the sample range of the training stages.

[0011] Preferably, the millimeter-wave radar cross-environment perception generalization model includes a feature extractor controlled by an adaptive feature map downsampling layer, a temporal recurrent neural network, and a multilayer perceptron. The output of the feature extractor is connected to the input of the temporal recurrent neural network, and the output of the temporal recurrent neural network is connected to the input of the multilayer perceptron.

[0012] Preferably, the feature extractor controls the downsampling strategy through the adaptive feature map downsampling layer to obtain an output sampled feature template, including: S411. The distance-velocity spectrum within the sample range of the training phase is used as the input feature map. The adaptive feature map downsampling layer performs dense convolution operations on the input feature map before downsampling to obtain an output feature map whose spatial size is not reduced and whose detailed information is fully preserved. S412. Two types of cross-subsampled grids are preset for the output feature map, and the values ​​of the two types of cross-subsampled grids are calculated respectively. Norm, and compare the two types of cross-sampling grids. The size of the norm, selection The cross-subsampled grid with a large norm is used as the output sampling feature template of the adaptive feature map downsampling layer.

[0013] Preferably, the comparison of the two types of cross-sampling grids The size of the norm is determined by the following formula:

[0014] in, Represents the two types of cross-sampling grids Norm comparison results This represents the sampling results on the first type of cross-sampling grid. This represents the sampling results on the second type of cross-sampling grid. Represents the first type of cross-sampling grid. Norm, Represents the second type of cross-sampling grid. Norm, Indicates taking and middle Functions with a large norm.

[0015] Preferably, the millimeter-wave radar cross-environment perception generalization model is trained until the loss function in the final training stage converges, at which point the training of the millimeter-wave radar cross-environment perception generalization model ends, and a trained millimeter-wave radar cross-environment perception generalization model is obtained.

[0016] This invention also proposes a millimeter-wave radar cross-environment perception generalization system based on curriculum learning, comprising: The preprocessing module preprocesses the raw millimeter-wave radar data to obtain the range-velocity spectrum. The constant false alarm rate detection module is used to perform constant false alarm rate detection on the distance-velocity spectrum and extract the easiest sample dataset; The progressive learning module for the course learning stage is used to perform progressive learning from easy to difficult on the easiest sample dataset, so as to obtain the total number of training stages and the range of training stage samples of the millimeter-wave radar cross-environment perception generalization model constructed by introducing an adaptive feature map downsampling mechanism. The training module is used to train the millimeter-wave radar cross-environment perception generalization model using the total number of training phases and the sample range of the training phases, so as to obtain the trained millimeter-wave radar cross-environment perception generalization model. The identification module is used to input the millimeter-wave radar data to be identified into the millimeter-wave radar cross-environmental perception generalization model and output the environmental perception identification result of the millimeter-wave radar data to be identified.

[0017] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention proposes a millimeter-wave radar cross-environment perception generalization method and system based on course learning. First, the original millimeter-wave radar data is preprocessed to obtain a range-velocity spectrum. A constant false alarm rate (CFAR) detection is performed on the range-velocity spectrum to extract the easiest sample dataset. Next, the easiest sample dataset undergoes progressive learning in stages, from easy to difficult. This progressive learning creates a gradual learning trajectory for the millimeter-wave radar cross-environment perception generalization model, alleviating overfitting to noise and enhancing the consistency and stability of the millimeter-wave radar perception system during cross-environment deployment. Then, an adaptive feature map downsampling mechanism is introduced into the millimeter-wave radar cross-environment perception generalization model. This mechanism significantly reduces the risk of information loss. It requires no additional parameter learning, is computationally simple, and can adaptively address target position translation issues caused by different viewpoints and distances. This effectively enhances the robustness of the millimeter-wave radar cross-environment perception generalization model to target spatial offsets and improves recognition consistency and accuracy under large-scale deployment conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a millimeter-wave radar cross-environment perception generalization method based on course learning proposed in an embodiment of the present invention. Figure 2 This is another flowchart illustrating a millimeter-wave radar cross-environment perception generalization method based on course learning proposed in this embodiment of the invention. Figure 3 This diagram illustrates the structure of a millimeter-wave radar cross-environment perception generalization system based on curriculum learning, as proposed in this embodiment of the invention. Detailed Implementation

[0019] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings; To facilitate understanding of this embodiment, the prior art information involved in this embodiment is first introduced as follows: With the widespread application of deep learning methods in millimeter-wave radar data processing, model performance is highly dependent on large-scale, high-quality datasets. However, millimeter-wave radar data is typically collected in specific experimental scenarios, introducing scene-specific data distribution biases that significantly reduce the model's generalization ability when facing new environments. This distribution difference caused by multipath effects severely impacts the performance of tasks such as human motion recognition. To mitigate this issue, domain adaptation techniques fine-tune the model by collecting data from the target environment to improve its environmental adaptability. However, this method is costly in practical deployment because data needs to be re-collected and labeled for each new environment, limiting the large-scale deployment of millimeter-wave sensing technology.

[0020] Against this backdrop, domain generalization, as a technique that enables model transfer without accessing target domain data, has become a research focus. Current millimeter-wave radar domain generalization methods are mostly based on traditional data augmentation or general strategies from computer vision. These methods have poor adaptability to the unique texture sparsity and physical meaning constraints of radar images, resulting in limited generalization effects. Therefore, in recent years, attempts have emerged to introduce learning into the field of domain generalization. Learning, by controlling the progressive increase in sample difficulty during training, helps guide the model to obtain more generalizable representations relying solely on the source domain, thereby improving its cross-scene recognition performance and reducing dependence on target domain data. This enhances the model's generalization ability in different environmental scenarios, overcoming the limitations caused by data scarcity and scene variations.

[0021] Chen X et al. proposed a method for directly generating millimeter-wave radar data using cue words based on a diffusion model. This method utilizes a set of diffusion models to generate radar sensing data that meets specific "cue word" requirements, generating massive amounts of data by adjusting simulation parameters (scene, material, and motion). The simulator incorporates the principles of electromagnetic wave propagation, ensuring the generated data has physical plausibility and achieving accurate, rich, and more realistic data distribution. The perception model is trained using only synthetic data and directly deployed to a real environment. However, while this diffusion model enhances data diversity, its generated results may deviate from physical laws, producing unreasonable samples. It requires significant computational power and is sensitive to hyperparameters. In high-dimensional data scenarios like millimeter waves, the model may struggle to converge stably or exhibit insufficient spatiotemporal consistency in the generated samples (e.g., discontinuous motion continuity).

[0022] Yao et al. proposed a method for millimeter-wave sensing domain generalization based on multi-source domain datasets. This method utilizes multiple source domain sub-classifiers and a shared feature extractor to coordinate and extract domain-invariant features. By using non-domain datasets and setting an optimized loss function on the fixed-weight sub-classifiers, the feature extractor is forced to extract invariant features, thus optimizing model generalization. However, this method requires multiple source datasets, making data construction complex and time-consuming. Furthermore, multi-source domain generalization itself can cause conflicts, increasing the instability of model training and potentially leading to model non-convergence when inter-domain biases are significant.

[0023] To enhance the cross-environment generalization capability of millimeter-wave human image classification models, this invention proposes a curriculum-based millimeter-wave radar cross-environment perception generalization method and system, combined with a non-learning adaptive feature map downsampling (APS) mechanism. Through training sample difficulty assessment and a progressive mechanism, along with anti-aliasing methods, the model training process is dynamically guided in an order of increasing difficulty, resisting information loss during feature extraction and enabling the model to learn more robust and generalizable feature representations in the source domain. This method does not rely on target domain samples, effectively improving the model's recognition accuracy in unknown scenarios. It exhibits good adaptability and deployment flexibility, providing an efficient, low-cost, and cross-scenario-capable training framework for millimeter-wave radar cross-environment perception generalization models.

[0024] The key terms used in this embodiment are explained as follows: LSTM: Long Short-Term Memory Network, an improved recurrent neural network structure that excels at handling long-term dependencies in time series data.

[0025] MLP: Multilayer Perceptron, which consists of multiple fully connected layers and is widely used for tasks such as classification and regression.

[0026] ResNet: Residual Network, which solves the gradient vanishing problem in deep network training by introducing residual connections, and performs well in tasks such as image recognition.

[0027] APS: Adaptive Feature Map Downsampling Layer, which replaces convolution and pooling downsampling operations in the feature extractor.

[0028] RGB image: A color image composed of three color channels: red, green, and blue. FFT: Fast Fourier Transform, an efficient and fast computation method that uses computers to calculate the Discrete Fourier Transform.

[0029] CFAR: Constant False Alarm Rate Detection, an adaptive detection algorithm designed to maintain a constant false alarm rate for target detection in the presence of noise and clutter.

[0030] Deep learning: A branch of machine learning that uses multi-layered neural networks to abstract data layer by layer and learn features. It is widely used in fields such as image recognition and natural language processing. Domain generalization: Domain adaptation: A transfer learning method that adapts a model to a target task by introducing data from the target domain when there are distributional differences between the source and target domains.

[0031] Domain generalization: During training, without relying on any target domain data, the model learns more generalizable representations, enabling it to be directly applied to unknown or changing target domain tasks.

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Example 1 like Figure 1 and Figure 2 As shown, this embodiment proposes a millimeter-wave radar cross-environment perception generalization method based on course learning, including the following steps: S1. Preprocess the raw millimeter-wave radar data to obtain the range-velocity spectrum; The preprocessing of the raw millimeter-wave radar data to obtain the range-velocity spectrum includes: S101. Perform a discrete Fourier transform on multiple chirp signal groups in each frame of the original millimeter-wave radar data in the slow time dimension (chirp dimension) to obtain velocity dimension information; S102. Perform a discrete Fourier transform on the original millimeter-wave radar data in the fast time dimension (sampling point dimension) to obtain the range dimension information of the reflected wave; S103. The results of the two frequency domain dimensions, the velocity dimension information and the range dimension information, are merged to obtain the range-velocity map; the final output range-velocity map has a size of 64×256, representing the energy intensity of the target at different distances and velocities.

[0034] The preprocessing of the raw millimeter-wave radar data also yields a range-angle spectrum, including: S111. Using an array structure of multiple receiving antennas, beamforming is performed on the initial millimeter-wave radar data at different angles to obtain angular dimension information; In S111, to obtain the directional information of the target reflected signal, after performing a fast time-dimension FFT, beamforming is performed at different angles on the third dimension (antenna dimension) of the initial millimeter-wave radar data using an array structure of multiple receiving antennas. Specifically, spatial domain beamforming is performed by constructing steering vectors at different angles to obtain angular information representing the energy concentration in the target reflection direction. Finally, the angular and range information are merged to obtain a range-angle spectrum with a size of 64×256, corresponding to the energy intensity at different distances and azimuth angles. Finally, the range-velocity spectrum and the range-angle spectrum are stitched together to form the input for network training, with a size of 128×256.

[0035] In S1, the raw millimeter-wave radar data is preprocessed to obtain the range-velocity spectrum and range-angle spectrum, which serve as inputs for subsequent course planning and the millimeter-wave radar cross-environment perception generalization model. The raw millimeter-wave radar data consists of raw data packets obtained by sampling intermediate frequency signals through an ADC. To facilitate subsequent processing by the millimeter-wave radar cross-environment perception generalization model, the raw millimeter-wave radar data needs to be transformed into image representations with spatial or motion information capabilities, namely the range-velocity spectrum and range-angle spectrum.

[0036] S112. Merge the angular dimension information and the distance dimension information to obtain the distance-angle spectrum.

[0037] To obtain the directional information of the target reflected signal, after performing a fast time-dimension FFT, beamforming processing is performed on the third dimension (antenna dimension) of the data using an array structure of multiple receiving antennas. Spatial domain beamforming is achieved by constructing steering vectors at different angles to obtain the energy concentration along the target reflection direction. Finally, this is merged with the range dimension to obtain a range-angle spectrum map with a size of 64×256, corresponding to the energy intensity at different distances and azimuth angles. Ultimately, the range-velocity map and the range-angle map are stitched together to form the input for network training, with a size of 128×256.

[0038] Here, the raw data acquired by millimeter-wave radar is preprocessed to obtain the millimeter-wave radar range-velocity spectrum and range-angle spectrum, which will serve as inputs for the subsequent course learning planning module and deep neural network. The raw data output by millimeter-wave radar is typically the raw data packet obtained by sampling the intermediate frequency signal through an ADC. To facilitate subsequent network recognition model processing, the raw radar signal needs to be converted into an image representation with spatial or motion information capabilities, namely the range-velocity spectrum and range-angle spectrum.

[0039] S2. Perform constant false alarm rate detection on the distance-velocity spectrum to obtain the easiest sample dataset; In S2, within the millimeter-wave radar scenario, the multipath propagation effect, target relative position, and signal coverage area vary significantly across different environments. Direct end-to-end training can easily lead to overfitting of the millimeter-wave radar cross-environment perception generalization model to the source domain scene-related features, making it difficult to generalize to new scenarios. This invention introduces a course learning mechanism, gradually inputting source domain data samples into the millimeter-wave radar cross-environment perception generalization model for training according to an "easy-to-difficult" strategy. This improves the convergence stability and cross-environment generalization capability of the millimeter-wave radar cross-environment perception generalization model. Specifically, the range dimension information in the range-velocity spectrum includes several range units. The range-velocity spectrum samples from the source domain are used as input. The energy distribution of the range-velocity spectrum is subjected to constant false alarm rate detection in the frequency domain to obtain the easiest sample dataset, including: S201. Obtain the detection threshold for each distance cell; wherein the step of obtaining the detection threshold for each distance cell is as follows: based on the peak energy and signal-to-noise ratio (SNR) of the spectrum, combined with the preset threshold coefficient and the number of reference cells, adaptively calculate the detection threshold for each distance cell.

[0040] S202. Scan each of the 256 distance cells one by one, and mark the distance cells whose energy exceeds the detection threshold corresponding to the current distance cell as active cells, thus obtaining several active segments that are interrupted due to noise interference. S203. Fill in the gaps in the activity segment to extract the longest continuous activity interval; wherein a "0 flip" algorithm with a tolerance of k is applied to fill in the gaps in the activity segment to compensate for temporary loss; S204. Record the start and end indices of the continuous activity interval, extract the corresponding sub-regions from the distance-velocity spectrum based on the start and end indices and output them, define the sub-regions as the easiest sample dataset, and apply this algorithm to the entire dataset. The calculation expression for the easiest sample dataset is as follows:

[0041] in, The distance-velocity spectrum is represented by the following coordinates: The pixel value at that time, This represents the width of the entire distance velocity spectrum. Therefore for lower bound and The upper bound is obtained by cropping from the distance-velocity spectrum.

[0042] S3. Perform progressive learning from easy to difficult stages on the easiest sample dataset to obtain the total number of training stages and the range of training stage samples for the millimeter-wave radar cross-environment perception generalization model constructed by introducing an adaptive feature map downsampling mechanism. S4. Using the total number of training phases and the sample range of the training phases, train the millimeter-wave radar cross-environment perception generalization model to obtain the trained millimeter-wave radar cross-environment perception generalization model. S5. Input the millimeter-wave radar data to be identified into the millimeter-wave radar cross-environmental perception generalization model, and output the environmental perception identification result of the millimeter-wave radar data to be identified.

[0043] This embodiment proposes a millimeter-wave radar cross-environment perception generalization method and system based on course learning. First, the original millimeter-wave radar data is preprocessed to obtain a range-velocity spectrum. A constant false alarm rate (CFAR) detection is performed on the range-velocity spectrum to extract the easiest sample dataset. Next, the easiest sample dataset undergoes progressive learning in stages, from easy to difficult. This progressive learning creates a gradual learning trajectory for the millimeter-wave radar cross-environment perception generalization model, alleviating overfitting to noise and enhancing the consistency and stability of the millimeter-wave radar perception system during cross-environment deployment. Then, an adaptive feature map downsampling mechanism is introduced into the millimeter-wave radar cross-environment perception generalization model. This mechanism significantly reduces the risk of information loss. It requires no additional parameter learning, is computationally simple, and can adaptively address target position translation issues caused by different viewpoints and distances. This effectively enhances the robustness of the millimeter-wave radar cross-environment perception generalization model to target spatial offset and improves recognition consistency and accuracy under large-scale deployment conditions. The present invention uses only source domain data during the training of the millimeter-wave radar cross-environment perception generalization model, without the need to introduce target scene samples, thereby reducing deployment costs and breaking through the bottleneck limitations of traditional methods in terms of sample dependence, downsampling distortion and generalization ability.

[0044] Example 2 See Figure 1 and Figure 2 This embodiment further illustrates a millimeter-wave radar cross-environment perception generalization method proposed in the above embodiments.

[0045] S3 describes a progressive learning process for the easiest sample dataset, progressing from easy to difficult in stages, including: S301. Divide the training phases into multiple training phases based on the region size of the most easily sampled dataset to obtain the total number of training phases; S302. Based on the total number of training stages, introduce sample regions affected by multipath interference into the region of the easiest sample dataset in a progressive manner from easy to difficult, thereby obtaining the sample range of the training stages.

[0046] First, training is conducted in the initial training phase. When the convergence of the millimeter-wave radar cross-environment perception generalization model on the training set of the current training phase reaches a preset standard—that is, the accuracy of the validation set stabilizes or the rate of decrease in the loss function slows down—it automatically enters the next training phase, expanding the sample area to increase the difficulty. Based on the total number of training phases, sample areas affected by multipath interference are progressively introduced from "easy" to "difficult," gradually transitioning the training samples from simple environmental noise to complex environmental noise. This phased training process aligns with the human learning process from easy to difficult, effectively guiding the millimeter-wave radar cross-environment perception generalization model to learn more generalized feature representations in the source domain and significantly improving its adaptability to complex echo features in unseen environments.

[0047] The millimeter-wave radar cross-environment perception generalization model is trained until the loss function of the final training stage converges, at which point the training of the millimeter-wave radar cross-environment perception generalization model ends, and a trained millimeter-wave radar cross-environment perception generalization model is obtained.

[0048] The millimeter-wave radar cross-environment perception generalization model includes a feature extractor controlled by an adaptive feature map downsampling layer, a temporal recurrent neural network, and a multilayer perceptron. The output of the feature extractor is connected to the input of the temporal recurrent neural network, and the output of the temporal recurrent neural network is connected to the input of the multilayer perceptron.

[0049] The feature extractor controls the downsampling strategy through the adaptive feature map downsampling layer to obtain an output sampled feature template, including: S411. The distance-velocity spectrum within the sample range of the training phase is used as the input feature map. The adaptive feature map downsampling layer is integrated after each convolutional layer of the feature extractor to replace the conventional pooling operation. The adaptive feature map downsampling layer performs a dense convolution operation with a stride of 1 on the input feature map before downsampling to obtain an output feature map whose spatial size is not reduced and whose detailed information is fully preserved. S412. Two types of cross-subsampled grids are preset for the output feature map, and the values ​​of the two types of cross-subsampled grids are calculated respectively. The norm is used to assess its information carrying capacity and to compare the two types of cross-sampling grids. The size of the norm, selection A cross-mode downsampling grid with a large norm is used as the output sampling feature template of the adaptive feature map downsampling layer; this not only preserves the ability of dense convolution to extract small features, but also achieves adaptive downsampling through an energy-optimal grid while ensuring information integrity.

[0050] The first type of cross-sampling grid consists of the intersections of all odd-numbered rows and odd-numbered columns, while the second type of cross-sampling grid consists of the intersections of all even-numbered rows and odd-numbered columns. The comparison of the two types of cross-sampling grids The size of the norm is determined by the following formula:

[0051] in, Represents the two types of cross-sampling grids Norm comparison results This represents the sampling results on the first type of cross-sampling grid. This represents the sampling results on the second type of cross-sampling grid. Represents the first type of cross-sampling grid. Norm, Represents the second type of cross-sampling grid. Norm, Indicates taking and middle Functions with a large norm.

[0052] Based on the sample range and total number of training stages defined in the progressive learning phases of the course, the millimeter-wave radar cross-environment perception generalization model is trained. Training stops when the loss function of the millimeter-wave radar cross-environment perception generalization model tends to stabilize in each stage, and then the next training stage begins. At the same time, an adaptive feature map downsampling (APS) layer is inserted to further alleviate the translation invariance of the feature extractor in the millimeter-wave radar cross-environment perception generalization model, so that it can maintain a stable output under the condition of input image offset, and further enhance the generalization.

[0053] This embodiment integrates course learning with the Adaptive Feature Map Downsampling (APS) layer into a unified deep neural network framework, forming a complete end-to-end training process. In the feature extractor, a lightweight residual network (ResNet18) is used to extract spatial features of the image. The downsampling strategy inside the lightweight residual network (ResNet18) is controlled by the Adaptive Feature Map Downsampling (APS) layer. Subsequently, a temporal recurrent neural network is used to encode temporal information and extract action sequence features. Finally, a multilayer perceptron (MLP) is used to output the environmental perception recognition result of the millimeter-wave radar data to be identified. Throughout the training process of the millimeter-wave radar cross-environmental perception generalization model, course learning controls the sample difficulty and training progress, while the Adaptive Feature Map Downsampling (APS) layer ensures that the offset of the input feature map does not affect the downsampling result, achieving the training goal of balancing structural generalization and local robustness.

[0054] Example 3 See Figure 3 A millimeter-wave radar cross-environment perception generalization system based on curriculum learning, comprising: The preprocessing module preprocesses the raw millimeter-wave radar data to obtain the range-velocity spectrum. The constant false alarm rate detection module is used to perform constant false alarm rate detection on the distance-velocity spectrum and extract the easiest sample dataset; The progressive learning module for the course learning stage is used to perform progressive learning from easy to difficult on the easiest sample dataset, so as to obtain the total number of training stages and the range of training stage samples of the millimeter-wave radar cross-environment perception generalization model constructed by introducing an adaptive feature map downsampling mechanism. The training module is used to train the millimeter-wave radar cross-environment perception generalization model using the total number of training phases and the sample range of the training phases, so as to obtain the trained millimeter-wave radar cross-environment perception generalization model. The identification module is used to input the millimeter-wave radar data to be identified into the millimeter-wave radar cross-environmental perception generalization model and output the environmental perception identification result of the millimeter-wave radar data to be identified.

[0055] In this embodiment, the original millimeter-wave radar data is first preprocessed to obtain a range-velocity spectrum. A constant false alarm rate (CFAR) detection is then performed on the range-velocity spectrum to extract the easiest sample dataset. Next, the easiest sample dataset undergoes a progressive learning process, from easy to difficult, creating a progressive learning trajectory for the millimeter-wave radar cross-environment perception generalization model. This alleviates the overfitting of the model to noise and enhances the consistency and stability of the millimeter-wave radar perception system during cross-environment deployment. Then, an adaptive feature map downsampling mechanism is introduced into the millimeter-wave radar cross-environment perception generalization model. This significantly reduces the risk of information loss. The adaptive feature map downsampling mechanism requires no additional parameter learning, is computationally simple, and can adaptively address target position translation issues caused by different viewpoints and distances. It effectively enhances the robustness of the millimeter-wave radar cross-environment perception generalization model to target spatial offset and improves recognition consistency and accuracy under large-scale deployment conditions.

[0056] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A millimeter-wave radar cross-environment perception generalization method based on curriculum learning, characterized in that, Includes the following steps: S1. Preprocess the raw millimeter-wave radar data to obtain the range-velocity spectrum; S2. Perform constant false alarm rate detection on the distance-velocity spectrum to obtain the easiest sample dataset; S3. Perform progressive learning from easy to difficult stages on the easiest sample dataset to obtain the total number of training stages and the range of training stage samples for the millimeter-wave radar cross-environment perception generalization model constructed by introducing an adaptive feature map downsampling mechanism. S4. Using the total number of training phases and the sample range of the training phases, train the millimeter-wave radar cross-environment perception generalization model to obtain the trained millimeter-wave radar cross-environment perception generalization model. S5. Input the millimeter-wave radar data to be identified into the millimeter-wave radar cross-environmental perception generalization model, and output the environmental perception identification result of the millimeter-wave radar data to be identified.

2. The millimeter-wave radar cross-environment perception generalization method according to claim 1, characterized in that, The preprocessing of the raw millimeter-wave radar data to obtain the range-velocity spectrum includes: S101. Perform a discrete Fourier transform on the original millimeter-wave radar data in the slow time dimension to obtain velocity dimension information; S102. Perform a discrete Fourier transform on the original millimeter-wave radar data in the fast time dimension to obtain range dimension information; S103. Merge the velocity dimension information and the distance dimension information to obtain the distance-velocity map.

3. The millimeter-wave radar cross-environment perception generalization method according to claim 2, characterized in that, The preprocessing of the raw millimeter-wave radar data also yields a range-angle spectrum, including: S111. Using an array structure of multiple receiving antennas, beamforming is performed on the initial millimeter-wave radar data at different angles to obtain angular dimension information; S112. Merge the angular dimension information and the distance dimension information to obtain the distance-angle spectrum.

4. The millimeter-wave radar cross-environment perception generalization method according to claim 3, characterized in that, The distance dimension information in the distance-velocity spectrum includes several distance units. The constant false alarm rate detection of the distance-velocity spectrum, obtaining the easiest sample dataset, includes: S201. Obtain the detection threshold for each distance cell; S202. Perform a unit-by-unit scan on several distance units, and mark the distance units whose energy exceeds the detection threshold corresponding to the current distance unit as active units, thereby obtaining several active segments that are interrupted due to noise interference. S203. Fill in the gaps in the activity segment to extract the longest continuous activity interval; S204. Record the start and end indices of the continuous activity interval, extract the corresponding sub-regions from the distance-velocity spectrum based on the start and end indices, and define the sub-regions as the easiest sample dataset. The calculation expression for the easiest sample dataset is as follows: in, The distance-velocity spectrum is represented by the following coordinates: The pixel value at that time, This represents the width of the entire distance velocity spectrum. Therefore for lower bound and The upper bound is the range.

5. The millimeter-wave radar cross-environment perception generalization method according to claim 1, characterized in that, The progressive learning process, from easy to difficult, on the easiest sample dataset includes: S301. Divide the training phases into multiple training phases based on the region size of the most easily sampled dataset to obtain the total number of training phases; S302. Based on the total number of training stages, introduce sample regions affected by multipath interference into the region of the easiest sample dataset in a progressive manner from easy to difficult, thereby obtaining the sample range of the training stages.

6. The millimeter-wave radar cross-environment perception generalization method according to claim 5, characterized in that, The millimeter-wave radar cross-environment perception generalization model includes a feature extractor controlled by an adaptive feature map downsampling layer, a temporal recurrent neural network, and a multilayer perceptron. The output of the feature extractor is connected to the input of the temporal recurrent neural network, and the output of the temporal recurrent neural network is connected to the input of the multilayer perceptron.

7. The millimeter-wave radar cross-environment perception generalization method according to claim 6, characterized in that, The feature extractor controls the downsampling strategy through the adaptive feature map downsampling layer to obtain an output sampled feature template, including: S411. The distance-velocity spectrum within the sample range of the training phase is used as the input feature map. The adaptive feature map downsampling layer performs dense convolution operations on the input feature map before downsampling to obtain an output feature map whose spatial size is not reduced and whose detailed information is fully preserved. S412. Two types of cross-subsampled grids are preset for the output feature map, and the values ​​of the two types of cross-subsampled grids are calculated respectively. Norm, and compare the two types of cross-sampling grids. The size of the norm, selection The cross-subsampled grid with a large norm is used as the output sampling feature template of the adaptive feature map downsampling layer.

8. The millimeter-wave radar cross-environment perception generalization method according to claim 7, characterized in that, The comparison of the two types of cross-sampling grids The size of the norm is determined by the following formula: in, Represents the two types of cross-sampling grids Norm comparison results This represents the sampling results on the first type of cross-sampling grid. This represents the sampling results on the second type of cross-sampling grid. Represents the first type of cross-sampling grid. Norm, Represents the second type of cross-sampling grid. Norm, Indicates taking and middle Functions with a large norm.

9. The millimeter-wave radar cross-environment perception generalization method according to claim 1, characterized in that, The millimeter-wave radar cross-environment perception generalization model is trained until the loss function of the final training stage converges, at which point the training of the millimeter-wave radar cross-environment perception generalization model ends, and a trained millimeter-wave radar cross-environment perception generalization model is obtained.

10. A millimeter-wave radar cross-environment perception generalization system based on curriculum learning, characterized in that, include: The preprocessing module preprocesses the raw millimeter-wave radar data to obtain the range-velocity spectrum. The constant false alarm rate detection module is used to perform constant false alarm rate detection on the distance-velocity spectrum and extract the easiest sample dataset; The progressive learning module for the course learning stage is used to perform progressive learning from easy to difficult on the easiest sample dataset, so as to obtain the total number of training stages and the range of training stage samples of the millimeter-wave radar cross-environment perception generalization model constructed by introducing an adaptive feature map downsampling mechanism. The training module is used to train the millimeter-wave radar cross-environment perception generalization model using the total number of training phases and the sample range of the training phases, so as to obtain the trained millimeter-wave radar cross-environment perception generalization model. The identification module is used to input the millimeter-wave radar data to be identified into the millimeter-wave radar cross-environmental perception generalization model and output the environmental perception identification result of the millimeter-wave radar data to be identified.

Images