A millimeter wave radar target classification method

Abstract

This invention relates to the field of radar signal processing and recognition technology, specifically a millimeter-wave radar target classification method, comprising: receiving raw point cloud data output by a millimeter-wave radar and filtering outomas; extracting nine-dimensional features from the processed targets to form a feature vector X=[x0, x1, x2…x8]; pre-screening targets using manual experience rules; classifying targets using a two-level decision tree architecture; performing geometric feature-assisted verification for near-range targets; establishing a confidence sliding window mechanism to directly determine the classification result as a large vehicle target based on the target size; performing category correction and post-processing optimization on the classification results; preparing training data and configuring manual weights and the importance of fused features; calculating sample weights and optimizing leaf node sizes; training and pruning the decision tree, and generating C code. This invention overcomes the problems of disconnect between the decision tree and manual intervention, and the mismatch between the model and the actual algorithm in current millimeter-wave radar target classification methods.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and recognition technology, specifically a millimeter-wave radar target classification method. Background Technology

[0002] Millimeter-wave radar, due to its strong anti-interference capability and fast response speed, is an important sensor in vehicle intelligent driving assistance systems. Target classification using millimeter-wave radar has always been a challenge, as its accuracy directly affects the reliability and safety of driving decisions. Current technologies for millimeter-wave radar target classification mainly suffer from the following problems: disconnect between decision trees and human intervention; and mismatch between the model and the actual algorithm. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a millimeter-wave radar target classification method.

[0004] The technical problem to be solved by this invention is achieved by the following technical solution: A millimeter-wave radar target classification method includes the following steps: Step 1: Receive the raw point cloud data output by the millimeter-wave radar and perform outlier filtering; Step 2: Perform nine-dimensional feature extraction on the processed target to form a feature vector X=[x0, x1, x2…x8], where x0 is the velocity standard deviation, x1 is the unit echo intensity, x2 is the point velocity standard deviation, x3 is the acceleration, x4 is the target distance, x5 is the number of associated points, x6 is the matching point density per unit area, x7 is the normalized echo intensity, and x8 is the target area. Step 3: Apply manual experience rules to pre-screen the targets, and use the not_class array to record the targets that are not of one or more types; Step 4: Use a two-level decision tree architecture to perform classification and output the classification labels and confidence. Step 5: Perform geometric feature verification on close-range targets; Step 6: Establish a trust-based sliding window mechanism to directly determine the classification result as a large vehicle based on the target size; Step 7: Perform category correction and post-processing optimization on the classification results; Step 8: Prepare training data and configure artificial weights and the importance of fused features; Step 9: Calculate sample weights and optimize leaf node size; Step 10: Train the decision tree and prune it to generate C code.

[0005] As a further improvement of the present invention, the outlier filtering method in step 1 is as follows: Calculate the radial velocity difference between each associated point and all other associated points. If the velocity difference exceeds 5 km / h, it is identified as an anomaly and filtered out.

[0006] As a further improvement to the present invention, the manual experience rule in step 3 is specifically as follows: Set the target category as 0=PEOPLE, 1=TWO_THREE_CAR, 2=FOUR_CAR, 3=BIG_CAR; Non-pedestrian detection: When the target area is >22m 2 Or the target distance is greater than 100m, or the standard deviation of the velocity is greater than 10, or the ground velocity is greater than or equal to 33km / h, or the acceleration is greater than 1.8m / s². 2 Or when the unit echo intensity is greater than -10dB, set not_class[0]=CLU_NOT_PEOPLE to clear the pedestrian trust window; Non-two-wheeled vehicle determination: When the speed standard deviation is >15 or the target area is >40m² 2 When the density of matching points per unit area is greater than 27, the target distance is greater than 115m, the unit echo intensity is greater than -10dB, or the maximum length is greater than 5.5m, set not_class[1]=CLU_NOT_TWO_THERR_CAR and clear the trust window of the two-wheeled vehicle; Non-car identification: When the target area is >60m² 2 When the density of matching points per unit area is greater than or the intensity of echo per unit area is greater than -10dB, set not_class[2]=CLU_NOT_CAR and clear the car trust window; Large vehicles are directly identified when the cluster category is CLU_BIG_CAR or the target area is >60m². 2 If the maximum length is greater than 7m, it is directly identified as a large vehicle and will not participate in subsequent decision tree classification.

[0007] As a further improvement to the present invention, the classification process of the two-level decision tree architecture in step 4 is as follows: First, a coarse classification is performed using the first level. Based on the output, the second level is selected and called. Large vehicles are directly determined by manual rules and do not participate in the decision tree classification. The first level distinguishes between pedestrians and cars. The input feature vector X is output with label=0 indicating non-cars and label=1 indicating non-pedestrians. The second level includes two types. The first type is called when the first level outputs label=0. The feature vector X is input again, and the output label=0 indicates a pedestrian, while the output label=1 indicates a two-wheeled vehicle. The second method: When the first-level output label=1 is called, the feature vector X is input again, and the output label=0 indicates a two-wheeled vehicle, and label=1 indicates a car.

[0008] As a further improvement to the present invention, the geometric feature-assisted verification in step 5 specifically includes: For close-range targets, extract the target center position and heading angle, rotate all associated points to the target's local coordinate system, and calculate the rotated coordinates as follows: x_raw=(x-x_ref)*cosθ-(y-y_ref)*sinθ, y_raw=(x-x_ref)*sinθ+(y-y_ref)*cosθ, Where x_raw is the X coordinate value after rotation, x is the X coordinate value of the nearby target, x_ref is the X coordinate value of the target center position, θ is the heading angle, y is the y coordinate value of the nearby target, y_ref is the y coordinate value of the target center position, and y_raw is the y coordinate value after rotation. Obtain the rotated x_min, x_max, y_min, y_max boundaries, and calculate the new length and width dimensions. new Hekuan new When the length new <1m and width new If the distance is less than 1 meter, it is considered a pedestrian; otherwise, it is considered a two-wheeled vehicle. The historical classification results for pedestrians and two-wheeled vehicles are statistically analyzed. When the number of pedestrians is greater than 50 or the number of two-wheeled vehicles is greater than 25, the classification results are fixed.

[0009] As a further improvement of the present invention, the process of establishing the trust sliding window decision mechanism in step 6 is as follows: Assign four trust counters to each target; Trust accumulation: For each target classification result in each frame, its confidence score is divided by 10 and then accumulated into the corresponding trust counter; Manual rule triggering: The trust_artificial_flag array is used to record the count of manual rule triggering. When the count reaches the threshold, i.e. _0_threshold=3, _1_threshold=5, _2_threshold=5, _3_threshold=5, the classification result is fixed. Trajectory lifespan determination: When trajectory lifespan life_cnt > 60, the category corresponding to the maximum value in the trust counter is selected as the final result and locked; Low-speed protection: When the ground speed temp_v≥10.8km / h, set not_class[0]=CLU_NOT_PEOPLE to exclude pedestrians.

[0010] As a further improvement to the present invention, the post-processing optimization in step 7, which involves category correction, specifically includes: Pedestrian to Two-Wheeled Vehicle Correction: When the width of the pedestrian target is >1.9m, or the length is >2.5m, or the speed is >9km / h, the pedestrian is corrected to be classified as a two-wheeled vehicle. Correction from two-wheeled vehicle to car: When the width of the two-wheeled vehicle target is greater than 2.5m and the length is greater than 3.5m, or the speed is greater than 60km / h, it is corrected to be classified as a car. Correction from passenger car to large vehicle: When the target length of a passenger car is greater than 7m, it is corrected to be classified as a large vehicle.

[0011] As a further improvement of the present invention, the specific steps in step 8 for configuring the importance of artificial weights and fused features are as follows: The manual weights for each feature are configured using the MANUAL_FEATURE_WEIGHTS dictionary, with values ​​ranging from [0, 1], where None indicates no intervention; the intervention intensity is controlled by the MANUAL_WEIGHT_ALPHA parameter, with alpha values ​​ranging from [0, 1], where alpha=0 indicates complete trust in the model output, and alpha=1 indicates complete trust in the manual configuration. The formula for fusing feature importance is: final_weights[i] = alpha * manual_weight[i] + (1 - alpha) * model_weights[i]; Use the fusion value when manual_weights[i] is not None; otherwise, use the model weights directly.

[0012] As a further improvement of the present invention, the sample weight calculation process in step 9 is as follows: calculate the base sample weights base_weights using class_weight='balanced'; calculate the feature adjustment coefficient: feature_adjustment=1.0+alpha*sum(xi*wi) / sum(wi), where xi is the sample feature value and wi is the fused feature weight; the final sample weights sample_weights=base_weights*feature_adjustment; The leaf node size optimization process is as follows: A two-stage search strategy is adopted. In the first stage, several candidate values ​​are generated on a logarithmic scale, and each candidate value is evaluated using 5-fold hierarchical cross-validation. The second stage involves a coarse search of several candidate values ​​generated near the best quality value, which are then evaluated using 5-fold cross-validation. The 1-SD rule is applied for selection: the highest accuracy is calculated and its standard deviation is subtracted to obtain a threshold. Among all candidate values ​​whose accuracy exceeds the threshold, the largest leaf node size is selected as the optimal value.

[0013] As a further improvement of the present invention, the process of training the decision tree and pruning it in step 10 is as follows: Train the decision tree using the optimal leaf node size, with the following parameters: max_depth=8, min_samples_leaf=optimal leaf node size, min_samples_split=2*optimal leaf node size; After training, cost complexity pruning is performed: the cost complexity pruning path of the decision tree is calculated, generating a series of pruned trees with different ccp_alpha values. The accuracy is evaluated on the test set, and the pruned tree with the highest accuracy on the test set is selected as the final model.

[0014] The beneficial effects of this invention are: This invention provides a millimeter-wave radar target classification method. Compared with existing technologies, this invention significantly improves classification accuracy; fully automates model optimization; integrates data and deployment; performs secondary verification using geometric features; and exhibits strong dynamic adaptability. It overcomes the problems of decision tree separation and manual intervention, as well as the mismatch between the model and the actual algorithm in current millimeter-wave radar target classification methods. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of nine-dimensional feature extraction; Figure 2 This is a schematic diagram of a two-level decision tree architecture; Figure 3 This is a schematic diagram of the trust level sliding window mechanism; Figure 4 Automatic generation and optimization of flowcharts for decision trees; Figure 5 This is a schematic diagram illustrating the fusion of manual weights and model weights. 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 in conjunction with the accompanying drawings and embodiments.

[0017] A millimeter-wave radar target classification method includes the following steps: Step 1: Receive the raw point cloud data output by the millimeter-wave radar and perform outlier filtering.

[0018] Specifically: Calculate the radial velocity difference between each associated point and all other associated points. If the velocity difference exceeds 5 km / h, it is identified as an anomaly and filtered out.

[0019] Step 2: Extract nine-dimensional features from the processed target, such as... Figure 1 As shown, the feature vector X = [x0, x1, x2…x8] is formed; where, x0 is the speed standard deviation: Based on a sliding window rr_window that stores 10 frames of content, the mean and variance of the speed values ​​of multiple frames within the sliding window are calculated to obtain the standard speed rr_std; x1 represents the unit echo intensity: target reflected signal intensity level; x2 is the standard deviation of point velocity: the standard deviation of velocity of the associated point cloud in a single frame, reflecting the consistency of motion within the target; x3 represents acceleration: calculated from the extended Kalman filter state vector, using the formula acceleration = sqrt(ekf_current_vector[S_iXa)). 2 +ekf_current_vector[S_iYa] 2 ); x4 represents the target distance: the radial distance between the target and the radar, match_value[S_RAW_R]. x5 represents the number of associated points: the number of original point clouds associated with the target, match_num; x6 represents the density of matching points per unit area: the ratio of the number of target associated points to the target area, match_num / (width*length); x7 represents the normalized echo intensity: the calculation formula is unit_echo_intensity / (R 4 )*100000, to eliminate the effect of distance attenuation; x8 represents the target area.

[0020] Step 3: Apply manual experience rules to pre-screen the targets, and use the not_class array to record the targets that are not of one or more types.

[0021] Specifically, the target categories are set as 0=PEOPLE, 1=TWO_THREE_CAR, 2=FOUR_CAR, and 3=BIG_CAR; Non-pedestrian detection: When the target area is >22m 2 Or the target distance is greater than 100m, or the standard deviation of the velocity is greater than 10, or the ground velocity is greater than or equal to 33km / h, or the acceleration is greater than 1.8m / s². 2Or when the unit echo intensity is greater than -10dB, set not_class[0]=CLU_NOT_PEOPLE to clear the pedestrian trust window; Non-two-wheeled vehicle determination: When the speed standard deviation is >15 or the target area is >40m² 2 When the density of matching points per unit area is greater than 27, the target distance is greater than 115m, the unit echo intensity is greater than -10dB, or the maximum length is greater than 5.5m, set not_class[1]=CLU_NOT_TWO_THERR_CAR and clear the trust window of the two-wheeled vehicle; Non-car identification: When the target area is >60m² 2 When the density of matching points per unit area is greater than or the intensity of echo per unit area is greater than -10dB, set not_class[2]=CLU_NOT_CAR and clear the car trust window; Large vehicles are directly identified when the cluster category is CLU_BIG_CAR or the target area is >60m². 2 If the maximum length is greater than 7m, it is directly identified as a large vehicle and will not participate in subsequent decision tree classification.

[0022] Step 4: Use a two-level decision tree architecture for classification, such as... Figure 2 As shown, the output is the category label and confidence.

[0023] Specifically, a coarse classification is first performed using the first level, and the second level is selected based on the output results. Large vehicles are directly determined by manual rules and do not participate in the decision tree classification.

[0024] The first level distinguishes between pedestrians and cars. The input feature vector X is output with label=0 indicating non-cars and label=1 indicating non-pedestrians. The second level includes two types. The first type is called when the first level outputs label=0. The feature vector X is input again, and the output label=0 indicates a pedestrian, while the output label=1 indicates a two-wheeled vehicle. The second method: When the first-level output label=1 is called, the feature vector X is input again, and the output label=0 indicates a two-wheeled vehicle, and label=1 indicates a car.

[0025] Each level of the decision tree outputs a quantified confidence score, with a value range of [0, 100].

[0026] Step 5: Perform geometric feature verification on nearby targets.

[0027] Specifically, for close-range targets (x≤30m and y≤30m), the target center position and heading angle are extracted, all associated points are rotated to the target's local coordinate system, and the rotated coordinates are calculated as follows: x_raw=(x-x_ref)*cosθ-(y-y_ref)*sinθ, y_raw=(x-x_ref)*sinθ+(y-y_ref)*cosθ, Where x_raw is the X coordinate value after rotation, x is the X coordinate value of the nearby target, x_ref is the X coordinate value of the target center position, θ is the heading angle, y is the y coordinate value of the nearby target, y_ref is the y coordinate value of the target center position, and y_raw is the y coordinate value after rotation.

[0028] Obtain the rotated x_min, x_max, y_min, y_max boundaries, and calculate the new length and width dimensions. new Hekuan new When the length new <1m and width new If the distance is less than 1 meter, it is considered a pedestrian; otherwise, it is considered a two-wheeled vehicle. The historical classification results for pedestrians and two-wheeled vehicles are statistically analyzed. When the number of pedestrians is greater than 50 or the number of two-wheeled vehicles is greater than 25, the classification results are fixed.

[0029] Step 6: Establish a trust level sliding window mechanism, such as... Figure 3 As shown, the classification result is directly determined to be a large vehicle based on the target size.

[0030] Specifically, four trust counters are assigned to each target; Trust accumulation: For each target classification result in each frame, its confidence score is divided by 10 and then accumulated into the corresponding trust counter; Manual rule triggering: The trust_artificial_flag array is used to record the count of manual rule triggering. When the count reaches the threshold, i.e. _0_threshold=3, _1_threshold=5, _2_threshold=5, _3_threshold=5, the classification result is fixed. Trajectory lifespan determination: When trajectory lifespan life_cnt > 60, the category corresponding to the maximum value in the trust counter is selected as the final result and locked; Low-speed protection: When the ground speed temp_v≥10.8km / h, set not_class[0]=CLU_NOT_PEOPLE to exclude pedestrians.

[0031] Step 7: Perform category correction and post-processing optimization on the classification results.

[0032] Specifically, the correction for pedestrians to two-wheeled vehicles: when the width of the pedestrian target is >1.9m, or the length is >2.5m, or the speed is >9km / h, the pedestrian is corrected to be classified as a two-wheeled vehicle; Correction from two-wheeled vehicle to car: When the width of the two-wheeled vehicle target is greater than 2.5m and the length is greater than 3.5m, or the speed is greater than 60km / h, it is corrected to be classified as a car. Correction from passenger car to large vehicle: When the target length of a passenger car is greater than 7m, it is corrected to be classified as a large vehicle.

[0033] Step 8: Prepare training data and configure artificial weights and the importance of fused features, such as... Figure 5 As shown.

[0034] Specifically, the configuration of artificial weights and the importance of fused features are as follows: The manual weights for each feature are configured using the MANUAL_FEATURE_WEIGHTS dictionary, with values ​​ranging from [0, 1], where None indicates no intervention; the intervention intensity is controlled by the MANUAL_WEIGHT_ALPHA parameter, with alpha values ​​ranging from [0, 1], where alpha=0 indicates complete trust in the model output, and alpha=1 indicates complete trust in the manual configuration. The formula for fusing feature importance is: final_weights[i] = alpha * manual_weight[i] + (1 - alpha) * model_weights[i]; Use the fusion value when manual_weights[i] is not None; otherwise, use the model weights directly.

[0035] Step 9: Calculate sample weights and optimize leaf node size, such as... Figure 4 As shown.

[0036] Specifically, the base sample weights are calculated using class_weight='balanced'; the feature adjustment coefficient is calculated as: feature_adjustment=1.0+alpha*sum(xi*wi) / sum(wi), where xi is the sample feature value and wi is the fused feature weight; the final sample weights are sample_weights=base_weights*feature_adjustment.

[0037] The leaf node size optimization process is as follows: A two-stage search strategy is adopted. In the first stage, several candidate values ​​are generated on a logarithmic scale. In this invention, 15 candidate values ​​are generated, ranging from 20 to 10,000. Each candidate value is evaluated using 5-fold hierarchical cross-validation. In the second stage, a coarse search is performed to generate several candidate values ​​near the best value. In this invention, 20 candidate values ​​are generated, with a search range of [0.3*optimal value, 3*optimal value]. Five-fold cross-validation is also used for evaluation. The 1-SD rule is applied for selection: the highest accuracy is calculated and its standard deviation is subtracted to obtain a threshold. Among all candidate values ​​whose accuracy exceeds the threshold, the largest leaf node size is selected as the optimal value.

[0038] Step 10: Train the decision tree and prune it to generate C code.

[0039] Specifically, the process of training and pruning the decision tree is as follows: train the decision tree using the optimal leaf node size, and configure the parameters max_depth=8, min_samples_leaf=optimal leaf node size, and min_samples_split=2*optimal leaf node size; After training, cost complexity pruning is performed: the cost complexity pruning path of the decision tree is calculated, generating a series of pruned trees with different ccp_alpha values. The accuracy is evaluated on the test set, and the pruned tree with the highest accuracy on the test set is selected as the final model.

[0040] The process of generating C code is as follows: the decision tree nodes are traversed recursively using the tree_to_c function, and the left child node, right child node index, feature index, and split threshold are recorded. Leaf node processing: Calculate the number of samples for each category of the leaf node to obtain the category probability distribution, select the category with the highest probability as the output label, and calculate the confidence level by multiplying the probability by 100. Internal node processing: Generates if-else conditional branch statements and outputs a C code file with the prototype void tree_model(constfloat*features, int8_t*label, uint8_t*confidence), which can be directly compiled and deployed to an automotive embedded platform.

[0041] 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 principles 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 the present invention is defined by the appended claims and their equivalents.

Claims

1. A millimeter-wave radar target classification method, characterized in that: Includes the following steps: Step 1: Receive the raw point cloud data output by the millimeter-wave radar and perform outlier filtering; Step 2: Perform nine-dimensional feature extraction on the processed target to form a feature vector X=[x0, x1, x2…x8], where x0 is the velocity standard deviation, x1 is the unit echo intensity, x2 is the point velocity standard deviation, x3 is the acceleration, x4 is the target distance, x5 is the number of associated points, x6 is the matching point density per unit area, x7 is the normalized echo intensity, and x8 is the target area. Step 3: Apply manual experience rules to pre-screen the targets, and use the not_class array to record the targets that are not of one or more types; Step 4: Use a two-level decision tree architecture to perform classification and output the classification labels and confidence. Step 5: Perform geometric feature verification on close-range targets; Step 6: Establish a trust-based sliding window mechanism to directly determine the classification result as a large vehicle based on the target size; Step 7: Perform category correction and post-processing optimization on the classification results; Step 8: Prepare training data and configure artificial weights and the importance of fused features; Step 9: Calculate sample weights and optimize leaf node size; Step 10: Train the decision tree and prune it to generate C code.

2. The millimeter-wave radar target classification method according to claim 1, characterized in that: The outlier filtering method in step 1 is as follows: Calculate the radial velocity difference between each associated point and all other associated points. If the velocity difference exceeds 5 km / h, it is identified as an anomaly and filtered out.

3. The millimeter-wave radar target classification method according to claim 1, characterized in that: The specific rules for manual experience in step 3 are as follows: Set the target category as 0=PEOPLE, 1=TWO_THREE_CAR, 2=FOUR_CAR, 3=BIG_CAR; Non-pedestrian detection: When the target area is >22m 2 Or the target distance is greater than 100m, or the standard deviation of the velocity is greater than 10, or the ground velocity is greater than or equal to 33km / h, or the acceleration is greater than 1.8m / s². 2 Or when the unit echo intensity is greater than -10dB, set not_class[0]=CLU_NOT_PEOPLE to clear the pedestrian trust window; Non-two-wheeled vehicle determination: when the speed standard deviation is >15 or the target area is >40m² 2 When the density of matching points per unit area is greater than 27, the target distance is greater than 115m, the unit echo intensity is greater than -10dB, or the maximum length is greater than 5.5m, set not_class[1]=CLU_NOT_TWO_THERR_CAR and clear the trust window of the two-wheeled vehicle; Non-car identification: When the target area is >60m² 2 When the density of matching points per unit area is greater than or the intensity of echo per unit area is greater than -10dB, set not_class[2]=CLU_NOT_CAR and clear the car trust window; Large vehicles are directly identified when the cluster category is CLU_BIG_CAR or the target area is >60m². 2 If the maximum length is greater than 7m, it is directly identified as a large vehicle and will not participate in subsequent decision tree classification.

4. The millimeter-wave radar target classification method according to claim 1, characterized in that: The classification process of the second-level decision tree architecture in step 4 is as follows: First, a coarse classification is performed using the first level. Based on the output, the second level is selected and called. Large vehicles are directly determined by manual rules and do not participate in the decision tree classification. The first level distinguishes between pedestrians and cars. The input feature vector X is output with label=0 indicating non-cars and label=1 indicating non-pedestrians. The second level includes two types. The first type is called when the first level outputs label=0. The feature vector X is input again, and the output label=0 indicates a pedestrian, while the output label=1 indicates a two-wheeled vehicle. The second method: When the first-level output label=1 is called, the feature vector X is input again, and the output label=0 indicates a two-wheeled vehicle, and label=1 indicates a car.

5. The millimeter-wave radar target classification method according to claim 1, characterized in that: The geometric feature-assisted verification in step 5 specifically includes: For close-range targets, extract the target center position and heading angle, rotate all associated points to the target's local coordinate system, and calculate the rotated coordinates as follows: x_raw=(x-x_ref)*cosθ-(y-y_ref)*sinθ, y_raw=(x-x_ref)*sinθ+(y-y_ref)*cosθ, Where x_raw is the X coordinate value after rotation, x is the X coordinate value of the nearby target, x_ref is the X coordinate value of the target center position, θ is the heading angle, y is the y coordinate value of the nearby target, y_ref is the y coordinate value of the target center position, and y_raw is the y coordinate value after rotation. Obtain the rotated x_min, x_max, y_min, y_max boundaries, and calculate the new length and width dimensions. new Hekuan new When the length new <1m and width new If the distance is less than 1 meter, it is considered a pedestrian; otherwise, it is considered a two-wheeled vehicle. The historical classification results for pedestrians and two-wheeled vehicles are statistically analyzed. When the number of pedestrians is greater than 50 or the number of two-wheeled vehicles is greater than 25, the classification results are fixed.

6. The millimeter-wave radar target classification method according to claim 1, characterized in that: The process of establishing the trust sliding window decision-making mechanism in step 6 is as follows: Assign four trust counters to each target; Trust accumulation: For each target classification result in each frame, its confidence score is divided by 10 and then accumulated into the corresponding trust counter; Manual rule triggering: The trust_artificial_flag array is used to record the count of manual rule triggering. When the count reaches the threshold, i.e. _0_threshold=3, _1_threshold=5, _2_threshold=5, _3_threshold=5, the classification result is fixed. Trajectory lifespan determination: When trajectory lifespan life_cnt > 60, the category corresponding to the maximum value in the trust counter is selected as the final result and locked; Low-speed protection: When the ground speed temp_v≥10.8km / h, set not_class[0]=CLU_NOT_PEOPLE to exclude pedestrians.

7. The millimeter-wave radar target classification method according to claim 1, characterized in that: The specific post-processing optimization in step 7, which involves category correction, is as follows: Pedestrian to Two-Wheeled Vehicle Correction: When the width of the pedestrian target is >1.9m, or the length is >2.5m, or the speed is >9km / h, the pedestrian is corrected to be classified as a two-wheeled vehicle. Correction from two-wheeled vehicle to car: When the width of the two-wheeled vehicle target is greater than 2.5m and the length is greater than 3.5m, or the speed is greater than 60km / h, it is corrected to be classified as a car. Correction from passenger car to large vehicle: When the target length of a passenger car is greater than 7m, it is corrected to be classified as a large vehicle.

8. The millimeter-wave radar target classification method according to claim 1, characterized in that: Step 8 involves configuring the importance of artificial weights and fused features as follows: The manual weights for each feature are configured using the MANUAL_FEATURE_WEIGHTS dictionary, with values ​​ranging from [0, 1], where None indicates no intervention; the intervention intensity is controlled by the MANUAL_WEIGHT_ALPHA parameter, with alpha values ​​ranging from [0, 1], where alpha=0 indicates complete trust in the model output, and alpha=1 indicates complete trust in the manual configuration. The formula for fusing feature importance is: final_weights[i] = alpha * manual_weight[i] + (1 - alpha) * model_weights[i]; Use the fusion value when manual_weights[i] is not None; otherwise, use the model weights directly.

9. The millimeter-wave radar target classification method according to claim 1, characterized in that: The sample weight calculation process in step 9 is as follows: use class_weight='balanced' to calculate the base sample weights base_weights; Calculate the feature adjustment coefficient: feature_adjustment=1.0+alpha*sum(xi*wi) / sum(wi), where xi is the sample feature value and wi is the fused feature weight; the final sample weights sample_weights=base_weights*feature_adjustment; The leaf node size optimization process is as follows: A two-stage search strategy is adopted. In the first stage, several candidate values ​​are generated on a logarithmic scale, and each candidate value is evaluated using 5-fold hierarchical cross-validation. The second stage involves a coarse search of several candidate values ​​generated near the best quality value, which are then evaluated using 5-fold cross-validation. The 1-SD rule is applied for selection: the highest accuracy is calculated and its standard deviation is subtracted to obtain a threshold. Among all candidate values ​​whose accuracy exceeds the threshold, the largest leaf node size is selected as the optimal value.

10. The millimeter-wave radar target classification method according to claim 1, characterized in that: The process of training and pruning the decision tree in step 10 is as follows: Train the decision tree using the optimal leaf node size, with the following parameters: max_depth=8, min_samples_leaf=optimal leaf node size, min_samples_split=2*optimal leaf node size; After training, cost complexity pruning is performed: the cost complexity pruning path of the decision tree is calculated, generating a series of pruned trees with different ccp_alpha values. The accuracy is evaluated on the test set, and the pruned tree with the highest accuracy on the test set is selected as the final model.

Images