A millimeter wave radar target detection method based on a self-training framework and a two-stage data augmentation strategy

Abstract

The present application belongs to the technical field of millimeter wave radar target detection, and particularly relates to a millimeter wave radar target detection method based on a self-training framework and a two-stage data augmentation strategy. The method comprises the following steps: step one, pre-training and model initialization; step two, target domain pseudo label generation; step three, one-stage structure augmentation; step four, two-stage signal augmentation; step five, joint training and model iterative updating; step six, dynamic pseudo label screening and self-training cycle; and step seven, final model evaluation and deployment. The present application is based on millimeter wave radar perception data, and fuses source domain supervised training, a self-training iterative mechanism and a two-stage augmentation strategy to enhance the robustness and generalization ability of the model from two dimensions of structure and signal.

Description

Technical Field

[0001] This invention belongs to the field of millimeter-wave radar target detection technology, specifically relating to a millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy. Background Technology

[0002] Millimeter-wave radar, with its excellent penetration capability, all-weather operation, and low power consumption, has been widely used in various fields such as autonomous driving, intelligent transportation, and security monitoring. In these application scenarios, target detection is a key task in the perception system and is of great significance for improving the overall system performance. However, compared with traditional visual images or LiDAR data, millimeter-wave radar data presents the following significant challenges:

[0003] 1. Data sparsity and unstructured data: Millimeter-wave radar data often manifests as point clouds or two-dimensional reflectance maps, with a density far lower than that of images or lidar, and an irregular spatial structure. This sparsity makes it difficult for existing deep learning-based object detection models to fully extract effective features.

[0004] 2. Strong noise and obstruction interference: Because millimeter-wave signals are easily affected by factors such as multipath reflection from metal surfaces, human body obstruction, and environmental noise, the raw data often contains a large number of false targets and noise, which places higher demands on the generalization ability of the model.

[0005] 3. Weak cross-domain generalization ability: In reality, millimeter-wave radar applications operate in complex and varied environments. Data distribution differs significantly across different scenarios (such as urban roads, underground parking garages, and rainy / foggy weather), often resulting in a significant performance drop in trained models in new target domains. Traditional supervised learning methods are particularly difficult to transfer to other domains when labeled data is lacking.

[0006] In recent years, self-supervised learning and self-training methods have been widely applied to cross-domain image recognition and semantic segmentation tasks. Self-training frameworks, by generating pseudo-labels on the target domain, guide the model to adaptively optimize under unsupervised conditions, demonstrating strong practicality. However, directly applying them to millimeter-wave radar data still faces several limitations: First, the quality of pseudo-labels is unstable and is greatly affected by data noise and model bias. Second, the lack of augmentation strategies applicable to the characteristics of millimeter-wave radar data makes it difficult for pseudo-labels to fully realize their effects; most current augmentation methods focus on images or point clouds and lack modeling of the characteristics of millimeter-wave radar signal layers.

[0007] Therefore, there is an urgent need for a novel approach for millimeter-wave radar target detection tasks that combines a self-training mechanism with an enhancement strategy that is more adapted to signal characteristics, in order to alleviate the problem of cross-domain performance degradation and improve the application effect of the model in real-world scenarios. Summary of the Invention

[0008] The purpose of this invention is to propose a millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy, which effectively improves the adaptability and detection accuracy of the millimeter-wave radar target detection model in the target domain without requiring target domain labeled data.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: A millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy is proposed, with the following specific steps: Step 1, Pre-training and Model Initialization: Train the basic target detection model using existing source domain millimeter-wave radar labeled data; Step 2, target domain pseudo-label generation: Apply the pre-trained model to unlabeled millimeter-wave radar data in the target domain, generate candidate boxes (used to label target locations) and their confidence scores as pseudo-labels through model prediction; and set a confidence score threshold to retain prediction results with confidence scores higher than the threshold. Step 3, First-level structural augmentation: Perform geometric augmentation operations on the spatial structure hierarchy for the target domain data and its pseudo-labels; Step 4, Secondary Signal Augmentation: Based on the primary structural augmentation, a millimeter-wave radar signal layer modeling mechanism is introduced to simulate the effects of signal disturbances and noise in the real environment; Step 5, Joint Training and Model Iterative Update: By fusing source domain labeled data, target domain pseudo-label data, and two-level augmented data, a joint training set is constructed to retrain the object detection model; Step 6, Dynamic pseudo-label selection and self-training loop: Repeat steps 2 to 6 by dynamically adjusting the confidence threshold until the model converges or reaches the preset number of iterations; Step 7, Final Model Evaluation and Deployment: The trained model is evaluated and tested in the target domain, and the model that passes the evaluation test is deployed for millimeter-wave radar target detection.

[0010] Optionally, step two of the present invention involves masking areas below the confidence threshold.

[0011] Optionally, the primary structure augmentation of the present invention includes at least one of the following three items: randomly scaling the target size, randomly rotating the target direction, and randomly translating the target position.

[0012] Optionally, the secondary signal augmentation of the present invention includes at least one of the following three aspects: signal disturbance simulation, background noise enhancement, and occlusion and missing information simulation.

[0013] Optionally, the signal disturbance simulation described in this invention involves applying a disturbance term to the radar reflection intensity.

[0014] in, These represent the radar signal strength and the radar signal strength after adding disturbance, respectively. This is the weather attenuation coefficient. Indicates whether to simulate severe weather.

[0015] Optionally, the background noise enhancement described in this invention is as follows:

[0016] in, This represents the original background noise. This indicates the background noise after adding interference. The time-dependent noise figure, This represents the day / night parameters.

[0017] Optionally, the occlusion and missingness simulation described in this invention is: randomly occluding the target area, randomly injecting false alarm targets, or simulating sparse or partially disappeared point clouds.

[0018] Optionally, the loss function for retraining the detection model according to the present invention is:

[0019] in, Loss due to source domain supervision; For the target domain pseudo-label supervision loss; To augment the supervised loss of the sample; These are the weighting coefficients.

[0020] Optionally, the dynamically adjusted confidence threshold described in this invention is,

[0021] in, The confidence threshold for the k-th iteration. To increment the step size, This is the maximum threshold.

[0022] Optionally, step seven of the present invention evaluates and tests the model by means of average detection precision, recall rate and / or false alarm rate; If the model does not perform as expected, further optimization can be achieved by adjusting the pseudo-label threshold, increasing the augmentation intensity, or improving the target domain data modeling method.

[0023] Beneficial effects: This invention is based on millimeter-wave radar perception data, and integrates source domain supervised training, self-training iterative mechanism and two-level augmentation strategy to enhance the robustness and generalization ability of the model from two dimensions: structural layer and signal layer. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in 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.

[0025] Figure 1 The flowchart of a millimeter-wave radar target detection method based on refined data modeling and augmentation provided by the present invention is shown. Detailed Implementation

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0028] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0029] like Figure 1 As shown in the figure, this embodiment presents a millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy. The specific steps are as follows: Step 1: Pre-training and Model Initialization First, a basic target detection model is trained using existing source-domain millimeter-wave radar labeled data, such as a target detection model based on a deep convolutional network or Transformer structure (e.g., YOLO, Faster R-CNN), to obtain initial model parameters. During the training process using the target detection model, multi-scale feature maps are extracted, including range-Doppler maps, angular spectral maps, and reflectance intensity maps, thus ensuring that the model has strong feature representation capabilities.

[0030] This embodiment uses millimeter-wave radar to collect 100 frames of initial millimeter-wave radar raw data, and uses the original 100 frames of radar raw data to train an initial target detection model. YOLOv8 is used as the basic network structure, and the training loss function includes bounding box regression loss, class cross-entropy loss and confidence loss.

[0031] Step 2: Generating pseudo-tags for the target domain A pre-trained model is applied to unlabeled millimeter-wave radar data in the target domain. Candidate bounding boxes and their confidence scores are generated by the model predictions and used as labels. A confidence threshold is set, and predictions with higher confidence scores are retained as pseudo-labels. The threshold is set to... At that time, the set of pseudo-tags is: in, This represents the target bounding box predicted by the model. This represents the corresponding confidence level.

[0032] In this embodiment, the pseudo-label retains the detection results with a confidence level greater than 0.6, and covers the low-confidence areas to avoid the pseudo-label noise misleading the training.

[0033] Step 3, First-level structural augmentation (structural data augmentation) Geometric augmentation operations at the spatial structure level are performed on the target domain data and its pseudo-labels (target location and confidence level), including: Randomly scale the target size:

[0034] Randomly rotate the target direction:

[0035] Randomly translate the target position:

[0036] in, The original center point of the target The image width and height are shown above. The above augmentation operation can effectively simulate the perception characteristics of a target under different observation angles and changes in the field of view. Indicate target Indicates the target angle. Indicates the scaling factor. Indicates the rotation factor. This represents the translation factor for the width and height of the center point.

[0037] Step four, secondary signal augmentation (signal-level data enhancement): While preserving structural features, a signal layer modeling mechanism unique to millimeter-wave radar is introduced to simulate the effects of signal disturbances and noise in the real environment. 1. Signal disturbance simulation: Applying a disturbance term to the radar reflection intensity: ,in, These represent the radar signal strength and the radar signal strength after adding disturbance, respectively. This is the weather attenuation factor, typically set at 0.15. Indicates whether to simulate severe weather (such as rain, snow, fog).

[0038] 2. Background noise enhancement: ,in This represents the original background noise. This indicates the background noise after adding interference. The time-dependent noise figure approaches 1 during the day and approaches 0 at night. This represents the day / night parameters.

[0039] 3. Occlusion and missing data simulation: Randomly occlude the target area or introduce background false alarm points to enhance the model's adaptability to abnormal situations.

[0040] (1) Occlude part of the target area (e.g., insert a block of all zeros or background values ​​in the target area); (2) Randomly inject false alarm targets (generate false target labels in non-target areas); (3) Defects such as sparse or partially disappeared point clouds in simulation.

[0041] 4. The fused augmented data is defined as follows: ,in The original radar signal data, For environmental parameter vectors, Indicates the complexity of the background.

[0042] This signal-level augmentation process significantly enhances the realism and diversity of pseudo-labeled training samples, contributing to the model's adaptability to different learning environments and its robustness against signal interference.

[0043] Step 5: Joint Training and Iterative Model Updates The source domain labeled data, target domain pseudo-label data, and two-level augmented data are fused to construct a joint training set, which is then used to retrain the detection model.

[0044] in: Loss due to source domain supervision; For the target domain pseudo-label supervision loss; To augment the supervised loss of the sample; These are the weighting coefficients. In this embodiment, the weighting coefficients can be selected as 1.0, 0.7, and 0.5. A hybrid training strategy is adopted, which integrates multi-domain data in a weighted manner to guide model optimization and effectively mitigates the interference of pseudo-label errors.

[0045] Step Six: Dynamic Pseudo-Label Filtering and Self-Training Loop A dynamic confidence adjustment strategy is adopted, gradually increasing the pseudo-label threshold as model accuracy improves, thereby guiding the model to generate higher-quality pseudo-labels and achieving closed-loop optimization of self-training. The specific strategy is as follows:

[0046] in, The threshold for the k-th iteration. To increment the step size, Set the maximum threshold. Repeat steps two through six until the model converges or reaches the preset number of iterations.

[0047] The initial value for the false label confidence threshold is set to [value]. , The threshold increment step size, This is the maximum confidence threshold. After each iteration, based on the pseudo-labels generated by the new model for the target domain data, target detection boxes with confidence levels below the current threshold are removed, and high-confidence pseudo-labels are retained for the next round of joint training. Steps two through six are repeated to gradually improve the quality of pseudo-labels and promote the convergence of the model's transfer learning process towards higher accuracy. Stability is typically achieved after 2-3 iterations.

[0048] Step 7: Final Model Evaluation and Deployment After training convergence, the final detection model is systematically evaluated in the target domain using standard radar target detection metrics, including: ● Average detection accuracy (mAP): Reflects overall detection performance; ●Recall: Reflects the completeness of target detection; ● False Alarm Rate (FAR): Assess the likelihood of false alarms.

[0049] The model's generalization ability is evaluated using the above metrics. If the model's performance does not meet expectations, it can be further optimized by adjusting the pseudo-label threshold, increasing the augmentation intensity, or improving the target domain data modeling method. Finally, the model is deployed to a millimeter-wave radar platform to adapt to real traffic environments or security monitoring systems, achieving highly reliable radar target detection tasks.

[0050] In summary, this embodiment combines a self-training strategy with a structure + signal two-layer augmentation method, which not only improves the usability of pseudo-labels but also enhances the robustness and accuracy of the detection model in different target domains.

[0051] It should be noted that the above are merely preferred embodiments of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy, characterized in that, The specific steps are as follows: Step 1, Pre-training and Model Initialization: Train the basic target detection model using existing source domain millimeter-wave radar labeled data; Step 2, target domain pseudo-label generation: The pre-trained model is applied to unlabeled millimeter-wave radar data in the target domain. Candidate boxes and their confidence scores are generated by the model prediction as pseudo-labels. A confidence threshold is set, and prediction results with confidence scores higher than the threshold are retained. Step 3, First-level structural augmentation: Perform geometric augmentation operations on the spatial structure hierarchy for the target domain data and its pseudo-labels; Step 4, Secondary Signal Augmentation: Based on the primary structural augmentation, a millimeter-wave radar signal layer modeling mechanism is introduced to simulate the effects of signal disturbances and noise in the real environment; Step 5, Joint Training and Model Iterative Update: By fusing source domain labeled data, target domain pseudo-label data, and two-level augmented data, a joint training set is constructed to retrain the object detection model; Step 6, Dynamic pseudo-label selection and self-training loop: Repeat steps 2 to 6 by dynamically adjusting the confidence threshold until the model converges or reaches the preset number of iterations; Step 7, Final Model Evaluation and Deployment: The trained model is evaluated and tested in the target domain, and the model that passes the evaluation test is deployed for millimeter-wave radar target detection. The secondary signal augmentation includes three aspects: signal disturbance simulation, background noise enhancement, and occlusion and missing signal simulation. The signal disturbance simulation is as follows: a disturbance term is applied to the radar reflection intensity. in, These represent the radar signal strength and the radar signal strength after adding disturbance, respectively. This is the weather attenuation coefficient. Indicates whether to simulate severe weather; The background noise enhancement is as follows: in, This represents the original background noise. This indicates the background noise after adding interference. The time-dependent noise figure, Indicates day-night parameters; The occlusion and missing simulation is: randomly occluding the target area, randomly injecting false alarm targets, or simulating sparse or partially disappearing point clouds. The augmented data after fusion is defined as: ,in The original radar signal data, For environmental parameter vectors, Indicates the complexity of the background.

2. The millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy according to claim 1, characterized in that, Step two involves masking areas below the confidence threshold.

3. The millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy according to claim 1, characterized in that, The primary structure augmentation includes at least one of the following three items: randomly scaling the target size, randomly rotating the target direction, and randomly translating the target position.

4. The millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy according to claim 1, characterized in that, The loss function for retraining the detection model is: in, Loss due to source domain supervision; For the target domain pseudo-label supervision loss; To augment the supervised loss of the sample; These are the weighting coefficients.

5. The millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy according to claim 1, characterized in that, The dynamically adjusted confidence threshold is: in, The confidence threshold for the k-th iteration. To increment the step size, This is the maximum threshold.

6. The millimeter-wave radar target detection method based on a self-training framework and a two-level data augmentation strategy according to claim 1, characterized in that, Step seven evaluates and tests the model by measuring average detection precision, recall, and / or false alarm rate. If the model does not perform as expected, further optimization can be achieved by adjusting the pseudo-label threshold, increasing the augmentation intensity, or improving the target domain data modeling method.

Images