A millimeter wave radar elevation angle ambiguity discrimination method based on directional diagram characteristics and machine learning

Abstract

The present application relates to the technical field of millimeter wave radar signal processing, in particular to a millimeter wave radar pitch angle ambiguity discrimination method based on directional diagram features and machine learning, and the specific steps are as follows: S1, channel response directional diagram data collection; S2, training data generation; S3, data enhancement processing; S4, feature data preprocessing; S5, model training and optimization; S6, model verification and evaluation; S7, ambiguity discrimination and confidence output; In the complex environment and multi-target scene, unlike the traditional method, the shortcomings of the prior art are overcome, the directional diagram features are introduced, the radar hardware channel characteristics are fully utilized, the channel response directional diagram features and the machine learning model are combined, the performance of the pitch angle ambiguity discrimination is improved, the feature extraction and model inference calculation amount are small, and the real-time processing requirements of the radar are met.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar signal processing technology, specifically a method for fuzzy discrimination of elevation angle in millimeter-wave radar based on radiation pattern features and machine learning. Background Technology

[0002] Millimeter-wave radar is widely used in autonomous driving, intelligent transportation, and drone navigation due to its high precision and strong anti-interference capabilities. However, in practical applications, vehicle-mounted millimeter-wave radar has a narrow detection range in the pitch direction, and often faces phase ambiguity problems during measurement, leading to incorrect detection of the pitch angle of targets above or on the ground, reducing the accuracy of pitch angle measurement, and thus affecting related driving safety functions.

[0003] Experiments have revealed that the radiation pattern characteristics of millimeter-wave radar contain rich angular information, but existing technologies have failed to fully utilize this hardware characteristic. Especially in complex environments and multi-target scenarios, the accuracy of traditional methods often falls short of practical application requirements. Therefore, there is an urgent need for a pitch angle ambiguity discrimination method that can fully utilize the radar hardware characteristics while possessing strong anti-jamming capabilities. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a method for ambiguity discrimination of elevation angle in millimeter-wave radar based on radiation pattern features and machine learning.

[0005] A method for ambiguity resolution of elevation angle in millimeter-wave radar based on radiation pattern features and machine learning, the specific steps of which are as follows: S1. Channel response pattern data acquisition: Acquire N channel pattern data corresponding to the angle reflector target within the entire angular range, the target's horizontal angle and elevation angle values ​​output by the radar, and the actual horizontal and elevation angles of the target. S2. Training data generation: Extract corresponding feature data from the collected channel radiation pattern data, and mark the data outside the radar detection elevation angle range as ambiguous and the data within the range as unambiguous. S3. Data Augmentation Processing: To address the issues of insufficient training data sample size or unbalanced distribution, the labeled feature data is augmented, including adding random noise and small angle shift perturbations. S4. Preprocessing of feature data: Normalize the original feature data and the enhanced data to eliminate the influence of dimensional differences; S5. Model Training and Optimization: Based on the preprocessed dataset, the training set, validation set, and test set are divided. The random forest model is used for training, and the model hyperparameters are dynamically adjusted through the validation set to optimize the model performance. S6. Model Validation and Evaluation: The trained model is evaluated using the test set, and the discrimination effect is verified by four indicators: accuracy, precision, recall and F1 score. S7. Fuzzy discrimination and confidence output: The real-time collected and preprocessed feature data is fed into the trained and optimized model to perform pitch angle fuzzy discrimination and output the confidence of the discrimination to support subsequent decision-making.

[0006] In step S1, the radar operates in FMCW mode, and the radar outputs in real time the target's horizontal angle α, elevation angle β, and the amplitude response values ​​[L1, L2, L3, L4, ... L...] of N channels. N Data was collected in 1° increments within the pitch angle range of -2φ° to +2φ°, and in 1° increments within the horizontal angle range of [-θ to θ]. During the data collection process, the horizontal angle α of the actual target was recorded. _true and pitch angle β _true This serves as the benchmark for subsequent labeling.

[0007] The feature data in step S2 is the amplitude response values ​​of the N receiving channels of the radar [L1, L2, L3, L4, ... L...]. N As channel response data, the target horizontal angle α and elevation angle β output by the corresponding radar are treated as two independent angle data. These three data constitute a set of feature data. Labeling rule: When the actual elevation angle β... _true When ∈ [-φ°, +φ°], it is labeled as "unambiguous state" (label 0); when β _true When ∉ [-φ°,+φ°], it is marked as "fuzzy state" (label 1).

[0008] In step S3, the channel response amplitude value L i Add random noise following a normal distribution N(0,σ²) (σ ranges from 0.01 to 0.05) to generate new amplitude response data. L i_noise = L i + N(0,σ²) Small angular offset perturbation: Add random offsets within the range of [-0.5°, +0.5°] to the horizontal angle α and the pitch angle β respectively to generate α. _disturb =α + Δα, β _disturb =β+Δβ (Δα and Δβ are uniformly distributed random numbers), simulating the angular deviation caused by the small displacement of the target in a real scene; In step S4, the original amplitude values ​​of the N channels and the amplitude values ​​of the enhanced data generated in S3 are normalized. The calculation expression for this normalization is as follows: L i '= (L i - L min ) / (L max -L min ) Where i is the channel index, L max The response values ​​for this group of channels are [L1, L2, L3, L4, ... L... N The maximum value in ], L min For the minimum value, L i L is the channel response value before normalization. i 'This is the value after normalization.

[0009] The raw horizontal and vertical angle data output by the radar, as well as the enhanced angle data generated in S3, are normalized. The calculation expression for this normalization is as follows: α'=(α+θ) / 2θ β'=(β+φ) / 2φ Where β is the horizontal angle value output by the radar, φ is the maximum value of the elevation angle detection range, and β' is the normalized elevation angle value.

[0010] In step S5, the preprocessed dataset is divided into a training set (for model parameter learning), a validation set (for hyperparameter tuning), and a test set (for final performance evaluation) in a ratio of 7:1:2. The division process uses stratified sampling to ensure that the ratio of fuzzy state data to unfuzzy state data is consistent in each set. Random forest model hyperparameter settings: The initial hyperparameters are 10 decision trees, the maximum depth of the decision trees is set to 8, the minimum number of sample splits is 5, and the minimum number of sample leaf nodes is 3; Hyperparameter optimization: Hyperparameters are adjusted using a grid search method based on the validation set. The search range includes the number of decision trees (5-20), maximum depth (5-15), and minimum number of sample splits (2-10). The F1 score of the validation set is used as the optimization objective to select the optimal combination of hyperparameters. Feature importance assessment: During training, the importance weight of each feature is recorded, and redundant features with a weight lower than 0.05 are removed (if they exist) to simplify the model structure and improve inference speed.

[0011] In step S6, the evaluation metrics are defined as follows: Accuracy = (True positives TP + True negatives TN) / (Total samples P + N); Precision = TP / (TP + False positives FP); Recall = TP / (TP + False negatives FN); F1 score = 2 × Precision × Recall / (Precision + Recall); Evaluation criteria: When the model's accuracy on the test set is ≥95% and its F1 score is ≥94%, the model is considered to have passed training; if the criteria are not met, return to step S3 to adjust the data augmentation strategy or step S5 to optimize the hyperparameters and retrain.

[0012] In step S7, the model adopts a random forest structure with optimal hyperparameter combination. The preprocessed feature data (Li', α', β') is input into the model, and the discrimination result is determined by the voting results of 10 decision trees (if the voting percentage is ≥50%, it is judged as the corresponding state). The confidence level is calculated for subsequent decision-making. Confidence level = (number of votes for the corresponding state) / total number of decision trees × 100%.

[0013] The beneficial effects of this invention are: in complex environments and multi-target scenarios, it differs from traditional methods and overcomes the shortcomings of existing technologies. By introducing radiation pattern features, it makes full use of the characteristics of radar hardware channels. By combining channel response radiation pattern features with machine learning models, it improves the performance of pitch angle ambiguity discrimination. At the same time, the computational load of feature extraction and model inference is small, which meets the requirements of radar real-time processing. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0016] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below.

[0017] like Figure 1 As shown, a method for ambiguity discrimination of elevation angle in millimeter-wave radar based on radiation pattern features and machine learning is described below. S1. Channel response pattern data acquisition: Acquire N channel pattern data corresponding to the angle reflector target within the entire angular range, the target's horizontal angle and elevation angle values ​​output by the radar, and the actual horizontal and elevation angles of the target. S2. Training data generation: Extract corresponding feature data from the collected channel radiation pattern data, and mark the data outside the radar detection elevation angle range as ambiguous and the data within the range as unambiguous. S3. Data Augmentation Processing: To address the issues of insufficient training data sample size or unbalanced distribution, the labeled feature data is augmented, including adding random noise and small angle shift perturbations. S4. Preprocessing of feature data: Normalize the original feature data and the enhanced data to eliminate the influence of dimensional differences; S5. Model Training and Optimization: Based on the preprocessed dataset, the training set, validation set, and test set are divided. The random forest model is used for training, and the model hyperparameters are dynamically adjusted through the validation set to optimize the model performance. S6. Model Validation and Evaluation: The trained model is evaluated using the test set, and the discrimination effect is verified by four indicators: accuracy, precision, recall and F1 score. S7. Fuzzy discrimination and confidence output: The real-time collected and preprocessed feature data is fed into the trained and optimized model to perform pitch angle fuzzy discrimination and output the confidence of the discrimination to support subsequent decision-making.

[0018] In complex environments and multi-target scenarios, this method differs from traditional methods and overcomes the shortcomings of existing technologies. By introducing radiation pattern features and making full use of the characteristics of radar hardware channels, it improves the performance of pitch angle ambiguity discrimination by combining channel response radiation pattern features with machine learning models. At the same time, the computational load of feature extraction and model inference is small, which meets the requirements of radar real-time processing.

[0019] In step S1, the radar operates in FMCW mode, and the radar outputs in real time the target's horizontal angle α, elevation angle β, and the amplitude response values ​​[L1, L2, L3, L4, ... L...] of N channels. N Data was collected in 1° increments within the pitch angle range of -2φ° to +2φ°, and in 1° increments within the horizontal angle range of [-θ to θ]. During the data collection process, the horizontal angle α of the actual target was recorded. _true and pitch angle β _true This serves as the benchmark for subsequent labeling.

[0020] The feature data in step S2 is the amplitude response values ​​of the N receiving channels of the radar [L1, L2, L3, L4, ... L...]. N As channel response data, the target horizontal angle α and elevation angle β output by the corresponding radar are treated as two independent angle data. These three data constitute a set of feature data. Labeling rule: When the actual elevation angle β... _trueWhen ∈ [-φ°, +φ°], it is labeled as "unambiguous state" (label 0); when β _true When ∉ [-φ°,+φ°], it is marked as "fuzzy state" (label 1).

[0021] In step S3, the channel response amplitude value L i Add random noise following a normal distribution N(0,σ²) (σ ranges from 0.01 to 0.05) to generate new amplitude response data. L i_noise = L i + N (0,σ²) Small angular offset perturbation: Add random offsets within the range of [-0.5°, +0.5°] to the horizontal angle α and the pitch angle β respectively to generate α. _disturb = α + Δα, β _disturb = β + Δβ (Δα and Δβ are uniformly distributed random numbers), simulating the angular deviation caused by the small displacement of the target in a real-world scenario; In step S4, the original amplitude values ​​of the N channels and the amplitude values ​​of the enhanced data generated in S3 are normalized. The calculation expression for this normalization is as follows: L i ' = (L i - L min ) / (L max - L min ) Where i is the channel index, L max The response values ​​for this group of channels are [L1, L2, L3, L4, ... L... N The maximum value in ], L min For the minimum value, L i L is the channel response value before normalization. i 'This is the value after normalization.

[0022] The raw horizontal and vertical angle data output by the radar, as well as the enhanced angle data generated in S3, are normalized. The calculation expression for this normalization is as follows: α'=(α+θ) / 2θ β'=(β+φ) / 2φ Where β is the horizontal angle value output by the radar, φ is the maximum value of the elevation angle detection range, and β' is the normalized elevation angle value.

[0023] In step S5, the preprocessed dataset is divided into a training set (for model parameter learning), a validation set (for hyperparameter tuning), and a test set (for final performance evaluation) in a ratio of 7:1:2. The division process uses stratified sampling to ensure that the ratio of fuzzy state data to unfuzzy state data is consistent in each set. Random forest model hyperparameter settings: The initial hyperparameters are 10 decision trees, the maximum depth of the decision trees is set to 8, the minimum number of sample splits is 5, and the minimum number of sample leaf nodes is 3; Hyperparameter optimization: Hyperparameters are adjusted using a grid search method based on the validation set. The search range includes the number of decision trees (5-20), maximum depth (5-15), and minimum number of sample splits (2-10). The F1 score of the validation set is used as the optimization objective to select the optimal combination of hyperparameters. Feature importance assessment: During training, the importance weight of each feature is recorded, and redundant features with a weight lower than 0.05 are removed (if they exist) to simplify the model structure and improve inference speed.

[0024] In step S6, the evaluation metrics are defined as follows: Accuracy = (True positives TP + True negatives TN) / (Total samples P + N); Precision = TP / (TP + False positives FP); Recall = TP / (TP + False negatives FN); F1 score = 2 × Precision × Recall / (Precision + Recall); Evaluation criteria: When the model's accuracy on the test set is ≥95% and its F1 score is ≥94%, the model is considered to have passed training; if it does not meet the criteria, return to step S3 to adjust the data augmentation strategy or step S5 to optimize the hyperparameters and retrain.

[0025] In step S7, the model adopts a random forest structure with optimal hyperparameter combination. The preprocessed feature data (Li', α', β') are input into the model, and the discrimination result is determined by the voting results of 10 decision trees (if the voting percentage is ≥50%, it is judged as the corresponding state). The confidence level is calculated for subsequent decision-making. Confidence level = (number of votes for the corresponding state) / total number of decision trees × 100%.

[0026] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely prisms of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for ambiguity discrimination of elevation angle in millimeter-wave radar based on radiation pattern features and machine learning, characterized in that: The specific steps are as follows: S1. Channel response pattern data acquisition: Acquire N channel pattern data corresponding to the angle reflector target within the entire angle range, the target horizontal angle and elevation angle values ​​output by the radar, and the actual horizontal and elevation angles of the target. S2. Training data generation: Extract corresponding feature data from the collected channel radiation pattern data, and mark the data outside the radar detection elevation angle range as ambiguous and the data within the range as unambiguous. S3. Data Augmentation Processing: To address the issues of insufficient training data sample size or unbalanced distribution, the labeled feature data is augmented. S4. Preprocessing of feature data: Normalize the original feature data and the enhanced data to eliminate the influence of dimensional differences; S5. Model Training and Optimization: Based on the preprocessed dataset, the training set, validation set, and test set are divided. The random forest model is used for training, and the model hyperparameters are dynamically adjusted through the validation set to optimize the model performance. S6. Model Validation and Evaluation: The trained model is evaluated using the test set, and the discrimination effect is verified by four indicators: accuracy, precision, recall and F1 score. S7. Fuzzy discrimination and confidence output: The real-time collected and preprocessed feature data is fed into the trained and optimized model to perform pitch angle fuzzy discrimination and output the confidence of the discrimination to support subsequent decision-making.

2. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 1, characterized in that: In step S1, the radar operates in FMCW mode, and the radar outputs in real time the target's horizontal angle α, elevation angle β, and the amplitude response values ​​[L1, L2, L3, L4, ... L...] of N channels. N Data is collected in steps of 1° within the pitch angle range of -2φ° to +2φ°, and in steps of 1° within the horizontal angle range of [-θ to θ].

3. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 1, characterized in that: The feature data in step S2 is the amplitude response values ​​[L1, L2, L3, L4, ... L] of the radar's N receiving channels. N As channel response data, the target horizontal angle α and elevation angle β output by the corresponding radar are taken as two independent angle data. The above three data constitute a set of characteristic data.

4. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 3, characterized in that: In step S3, the added processing includes adding random noise and small angular offset perturbations.

5. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 3, characterized in that: In step S3, the channel response amplitude value L i Add random noise following a normal distribution N(0,σ²) to generate new amplitude response data: L i_noise = L i + N (0,σ²) The value of σ ranges from 0.01 to 0.05; Small angular offset perturbation: Add random offsets within the range of [-0.5°, +0.5°] to the horizontal angle α and the pitch angle β respectively to generate α. _disturb =α + Δα, β _disturb =β+Δβ where Δα and Δβ are uniformly distributed random numbers, simulating the angular deviation caused by the small displacement of the target in the actual scene.

6. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 3, characterized in that: In step S4, the original amplitude values ​​of the N channels are normalized. The calculation expression for this normalization is as follows: L i '=(L i - L min ) / (L max - L min ) Where i is the channel index, L max The response values ​​for this group of channels are [L1, L2, L3, L4, ... L... N The maximum value in ], L min For the minimum value, L i L is the channel response value before normalization. i 'This is the value after normalization.

7. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 3, characterized in that: In step S4, the horizontal and vertical angle data output by the radar are normalized. The calculation expression for this normalization is as follows: α' = (α + θ) / 2θ β' = (β + φ) / 2φ Where β is the horizontal angle value output by the radar, φ is the maximum value of the elevation angle detection range, and β' is the normalized elevation angle value.

8. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 1, characterized in that: In step S5, the preprocessed dataset is divided into a training set for model parameter learning, a validation set for hyperparameter tuning, and a test set for final performance evaluation in a ratio of 7:1:

2. The division process uses stratified sampling.

9. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 1, characterized in that: In step S6, the evaluation index is defined as follows: Accuracy = (True positives TP + True negatives TN) / (Total samples P + N); Precision rate = TP / (TP + False positives FP); Recall rate = TP / (TP + False negatives FN); F1 score = 2 × precision × recall / (precision + recall); Evaluation criteria: The model is considered to have passed training when its accuracy on the test set is ≥95% and its F1 score is ≥94%. If the criteria are not met, return to step S3 to adjust the data augmentation strategy or step S5 to optimize the hyperparameters and retrain.

10. The millimeter-wave radar elevation angle ambiguity discrimination method based on radiation pattern features and machine learning according to claim 1, characterized in that: In step S7, the model adopts a random forest structure with optimal hyperparameter combination, inputs the real-time preprocessed feature data (Li', α', β') into the model, and determines the discrimination result through the voting results of 10 decision trees.

Images