A Method and Model for Ghost Suppression of Millimeter-Wave Radar Based on Machine Learning

Abstract

The present invention discloses a method and model for suppressing ghost images of millimeter-wave radar based on machine learning, belonging to the technical field of automatic target tracking. The present invention selects data feature parameters closely related to radar characteristics and subsequent target tracking algorithms. These features not only include the moving point cloud attributes directly affecting target generation, but also cover the environmental information around the target that does not directly form the target. The model of the present invention can identify and suppress ghost images more quickly and accurately. It can be applied to a specific scenario by tightening or adding some constraint conditions, and can also be applied to the target tracking task in a complex environment by relaxing the constraint conditions, improving the anti-interference ability of the system and improving the decision-making strategy for finally determining ghost images and real targets by the tracking algorithm in different scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic target tracking, and particularly to a method and model for suppressing millimeter-wave radar ghosts based on machine learning. Background Art

[0002] Millimeter-wave radar sensors play a crucial role in advanced driver assistance systems (ADAS) and autonomous vehicles (AV) due to their excellent penetration and robustness to adverse weather conditions. These systems rely on radar sensors to detect other vehicles, pedestrians, obstacles, etc. in the surrounding environment to achieve advanced functions such as collision warning, adaptive cruise control, and automatic parking. However, millimeter-wave radar systems face some challenges in target tracking and environmental perception. One of them is the so-called "ghost" phenomenon, which refers to false signals caused by factors such as multipath propagation, non-linear effects, or hardware limitations in radar signal processing. The presence of ghosts seriously interferes with the detection and tracking of real targets by the radar system, reducing the reliability and safety of the system.

[0003] Currently, the following problems exist in traditional ghost suppression methods:

[0004] Using radar point cloud fitting for drivable area boundary detection and road edge detection, although it has some effect to a certain extent, it lacks sufficient flexibility and adaptability to cope with the changing real-world traffic environment.

[0005] Suppressing through the physical characteristics of ghost targets or through statistical analysis of features has limitations in accurately identifying and suppressing ghosts due to the complexity and variability of radar signals.

[0006] Based on this, the present invention designs a method and model for suppressing millimeter-wave radar ghosts based on machine learning to solve the above problems. Summary of the Invention

[0007] In view of the above-mentioned drawbacks of the prior art, the present invention provides a method and model for suppressing millimeter-wave radar ghosts based on machine learning.

[0008] To achieve the above object, the present invention is realized through the following technical solutions:

[0009] A method for suppressing millimeter-wave radar ghosts based on machine learning, comprising the following steps:

[0010] Step 1: Collect millimeter-wave radar target tracking data under various environmental conditions and preprocess it. Retrieve the data sequence corresponding to each cycle target ID in the HDF5 sequence file. Perform feature processing on the data by retrieving the sequence name and target ID. Then extract the characteristic parameters that can effectively distinguish target signals from interference signals. Next, screen the characteristic parameters and select the characteristic parameters with higher scores as the main input basis for the dataset.

[0011] Step 2: Randomly divide the dataset into a training set and a test set.

[0012] Step 3: In the sub-module of the automatic parameter tuning module, define the search space for characteristic parameters and use grid search to find the optimal combination of characteristic parameters. Use a weighted combination of the Gini impurity criterion and the information entropy criterion to obtain a comprehensive evaluation index α. By adjusting α, a trade-off can be made between the Gini impurity criterion and the information entropy criterion. Then set the evaluation criteria for splitting subtrees, the range of the number of base estimators, the range of the minimum number of samples in leaf nodes, and the range of the minimum number of samples required for further splitting of internal nodes. By reasonably setting these parameters, the training and prediction processes of the model can be controlled.

[0013] Step 4: Build a classifier model in the automatic parameter tuning module. Use the random forest algorithm to build multiple decision trees, and then make predictions and train the results of the integrated decision tree by the majority voting principle.

[0014] Step 5: After the decision tree is trained, use the confusion matrix and the ROC curve to input the test set samples into the classifier model to calculate and compare the recall rate, calculate the AUC data to evaluate the performance of the classifier model. Then adopt the Monte Carlo method to build a framework for automatic parameter tuning. With the computing power of the computer, directly simulate Steps 3 and 4 above, continuously sample to find the characteristic parameters of the classifier that meet the expected recall rate and specificity, select the characteristic parameters with the most balanced recall rate and specificity, gradually approach the best result, and finally further screen out the best model characteristic parameters.

[0015] Step 6: Convert the best model characteristic parameters into a.h file, and the generated.h file is embodied in the data structure of a binary tree.

[0016] Step 7: In the radar-based object detection, tracking, and environmental perception module, integrate the ghost suppression model into the target tracking sub-module. The target tracking sub-module inputs the characteristic parameters into the target update module for calculation. After reading the calculated characteristic parameters in the target update module, call the.h file to classify the target as a ghost target or a valid target. By recursively traversing each binary tree, make a decision according to the characteristic parameters and thresholds of the current node, and post-process the classified results. After counting with a counter, obtain a stable output result.

[0017] Further, in step one, the process of data preprocessing: comparing the collected radar target tracking data with the visual ground truth, and making annotations to distinguish ghost targets and real targets;

[0018] If the visual ground truth matches the target output of the tracking model, set the is_valid label to 1;

[0019] If the visual ground truth does not match the target output of the tracking model, set the is_valid label to 0;

[0020]

[0021] Where, Z: the target to be annotated, X: the visual ground truth target, Y: the model output target;

[0022] None: there is no matching target; Otherwise: the complement of X i ∩Y j ≠None

[0023] Further, in step one, the feature parameters include the maximum radar cross-sectional area in the i point cloud clusters associated with the target in the current period, the average value of the maximum radar cross-sectional area in the point cloud clusters associated with the target from the creation period m to the current period n of maintenance, the maximum radar cross-sectional area in the point cloud clusters associated with the target from the creation period m to the current period n of maintenance, the maximum SNR in the point cloud clusters associated with the target, the ratio of the number and area of static targets between the sensor installation position and the tracked target, and the target movement trajectory feature.

[0024] Further, the calculation formula for the maximum radar cross-sectional area in the i point cloud clusters associated with the target in the current period is:

[0025] object_track.max_rcs = max(associated_det 0 .rcs, associated_det 1 .rcs

[0026] …associated_det i-1 .rcs, associated_det i .rcs)

[0027] associated_det i : the i-th point cloud associated with the target;

[0028] The calculation formula for the average value of the maximum radar cross-sectional area in the point cloud clusters associated with the target from the creation period m to the current period n of maintenance is:

[0029]

[0030] The calculation formula for the maximum radar cross - sectional area in the point cloud cluster associated with the target from the creation cycle m to the current maintenance cycle n is as follows:

[0031]

[0032] The calculation formula for the ratio of the number of static targets to the area between the sensor installation position and the tracked target is:

[0033]

[0034] area_x = object_track.pos_x - sensor.pos_x

[0035] area_y = object_track.pos_y - sensor.pos_y

[0036] Among them, object_track.pos_x: the x - coordinate of the target position in the coordinate system with the center of the vehicle's rear axle as the origin;

[0037] object_track.pos_y: the y - coordinate of the target position in the coordinate system with the center of the vehicle's rear axle as the origin;

[0038] sensor.pos_x: the x - coordinate of the radar installation position;

[0039] sensor.pos_y: the y - coordinate of the radar installation position;

[0040] area_x: the difference in the x - direction between the sensor installation position and the tracked target;

[0041] area_y: the difference in the y - direction between the sensor installation position and the tracked target;

[0042] The calculation of the target motion trajectory characteristics is obtained by calculating the first - order derivative and the second - order derivative of the target position.

[0043] Furthermore, in step three:

[0044] When α is 0, it tends to use the information entropy criterion as the node division criterion;

[0045] When α is 1, it tends to use the Gini impurity criterion as the node division criterion.

[0046] Furthermore, the calculation method of node division:

[0047] criterion = α * Gini+(1 - α) * Entropy

[0048] Gini = ∑p(x)(1 - p(x))

[0049] Entropy = -∑p(x)log 2 p(x)

[0050] Among them, criterion is the calculation criterion for node division, Gini is the Gini coefficient, Entropy is the information entropy coefficient, and p(x) is the probability that the class is x.

[0051] Furthermore, in step six, the data structure of the binary tree includes the TreesNode structure and the Trees structure.

[0052] Furthermore, in step six, the TreesNode structure is as follows:

[0053] typedef struct{

[0054] s8_t feature;

[0055] f32_t value;

[0056] s16_t left;

[0057] s16_t right;

[0058] }TreesNode;

[0059] The TreesNode structure is used to represent a node of a binary tree;

[0060] Feature: An 8-bit signed integer type (s8_t), representing the feature number checked by the current node;

[0061] Value: A 32-bit floating-point type (f32_t), representing the threshold of the current node;

[0062] Left: A 16-bit signed integer type (s16_t), representing the index of the left child node of the current node in the nodes array; if it is negative, it means that this node is a leaf node and has no further child nodes;

[0063] Right: A 16-bit signed integer type (s16_t), representing the index of the right child node of the current node in the nodes array; if it is negative, it means that this node is a leaf node and has no further child nodes.

[0064] Furthermore, the Trees structure is as follows:

[0065] typedef struct{

[0066] s32_t n_nodes;

[0067] TreesNode* nodes;

[0068] s32_t n_trees;

[0069] s32_t* tree_roots;

[0070] } Trees;

[0071] The Trees structure is used to represent a forest composed of multiple binary trees;

[0072] n_nodes: A 32-bit signed integer type (s32_t), representing the total number of nodes of all binary trees in the forest;

[0073] nodes: A pointer to an array of TreesNode structures, storing the node information of all binary trees;

[0074] n_trees: A 32-bit signed integer type (s32_t), representing the number of binary trees in the forest;

[0075] tree_roots: A pointer to an array of 32-bit signed integer type (s32_t), storing the indices of the root nodes of each binary tree in the nodes array.

[0076] To better achieve the object of the present invention, the present invention also provides a ghost suppression model constructed by using the above method.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0078] 1. The present invention selects data feature parameters closely related to radar characteristics and subsequent target tracking algorithms. These data feature parameters not only include the moving point cloud attributes that directly affect target generation, but also cover the environmental information that does not directly form the target but surrounds the target.

[0079] 2. The model generated by the present invention can be implemented by binary trees, has strong interpretability, is easy to be deployed in the target tracking sub-module, and occupies less memory.

[0080] 3. Due to the existence of Step 1 and Step 2, that is, the data feature parameters and limiting conditions are transparent to developers, the code has strong interpretability. When there is a large deviation between the output result and the expectation, developers can quickly locate the problem and modify the code, greatly reducing the number of times and time of data training in the process of model algorithm tuning.

[0081] 4. The model of the present invention can identify and suppress ghosts more quickly and accurately. It can be applied to a specific scenario by tightening or adding some constraint conditions, and can also be applied to the target tracking task in a complex environment by relaxing the constraint conditions, improving the anti-interference ability of the system and improving the decision-making strategy for finally determining ghosts and real targets in the tracking algorithm under different scenarios.

[0082] 5. The present invention uses an automated parameter tuning design, which can reduce blind parameter tuning and quickly and effectively output the best model. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0084] Figure 1 It is a flowchart of a method for suppressing ghosts of a millimeter-wave radar based on machine learning according to the present invention;

[0085] Figure 2 It is an evaluation diagram of the confusion matrix model of the present invention;

[0086] Figure 3 It is a simulation diagram of the effect of suppressing ghost targets by machine learning of the present invention in a tunnel scenario;

[0087] Figure 4 It is a simulation diagram of the effect of suppressing ghost targets by traditional suppression methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0088] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0089] Embodiment 1

[0090] In some embodiments, please refer to the accompanying specification Figure 1 , a method for suppressing ghosts of a millimeter-wave radar based on machine learning, includes the following steps:

[0091] Step 1: Collect radar target tracking data under various environmental conditions and preprocess it; various environmental conditions: the vehicle is stationary and moving in tunnels, viaducts, and beside fences, etc.

[0092] Process of data preprocessing: Compare the collected radar target tracking data with the visual ground truth, and perform annotation to distinguish ghost targets and real targets.

[0093] If the visual ground truth matches the target output of the tracking model, set the is_valid label to 1; otherwise, set it to 0.

[0094]

[0095] Z: Targets to be annotated; X: Visual ground truth targets; Y: Model output targets.

[0096] None: No matching targets; Otherwise: The complement of X i ∩Y j ≠None.

[0097] Retrieve the data sequence corresponding to each cycle target ID in the HDF5 sequence file, perform feature processing on the data by retrieving the sequence name and target ID, and then extract the feature parameters that can effectively distinguish target signals and interference signals.

[0098] For each feature parameter, set the interval range in segments and maintain a corresponding counter. Then use the counter corresponding to the feature parameter for feature scoring and screening to adaptively learn and update the recognition ability of the feature. The initial value of the counter is 0. When a certain feature parameter is within this interval, the corresponding counter is incremented by 1. Calculate the counter values of each feature parameter, calculate the ratio of the occurrence frequencies of each feature parameter in the valid target and ghost target sets as the score of this feature. Finally, select the feature parameter with a higher score as the main input basis for the dataset for target recognition.

[0099] Feature parameters include:

[0100] Calculate the maximum RCS (Radar Cross Section) in the point cloud cluster associated with the target in the current cycle, that is, the maximum radar cross section and the mean value in the point cloud cluster associated with the target from creation to maintenance in the current cycle; this kind of dynamic time series analysis can better capture the motion characteristics of the target and help distinguish persistent real targets from transient interference signals.

[0101] RCS is an index to measure the ability of a target to reflect radar signals in the radar receiving direction. The larger the RCS, the easier it is for the object to be detected.

[0102] The calculation formula for the maximum radar cross section in the i point cloud clusters associated with the target in the current cycle is:

[0103] object_track.max_rcs = max(associated_det 0 .rcs, associated_det 1 .rcs…associated_det i-1 .rcs, associated_det i .rcs)

[0104] associated_det i : the i-th point cloud associated with the label;

[0105] The average value of the maximum radar cross-sectional area in the point cloud cluster associated with the target from the creation cycle m to the current maintenance cycle n is:

[0106]

[0107] The calculation formula for the maximum radar cross-sectional area in the point cloud cluster associated with the target from the creation cycle m to the current maintenance cycle n is:

[0108]

[0109] The maximum SNR (SIGNAL-NOISE RATIO) in the point cloud cluster associated with the target. The signal-to-noise ratio represents the transmission quality of the signal under noise interference. Generally, the larger the signal-to-noise ratio, the better the signal quality;

[0110] Judge whether there is a reflecting mirror by calculating the ratio of the number and area of static targets between the sensor installation position and the tracked target;

[0111] The calculation formula for the ratio of the number and area of static targets between the sensor installation position and the tracked target is:

[0112]

[0113] area_x = object_track.pos_x - sensor.pos_x

[0114] area_y = object_track.pos_y - sensor.pos_y

[0115] Among them, object_track.pos_x: the x coordinate of the target position in the coordinate system with the center of the vehicle's rear axle as the origin;

[0116] object_track.pos_y: the y coordinate of the target position in the coordinate system with the center of the vehicle's rear axle as the origin;

[0117] sensor.pos_x: The x-coordinate of the radar installation position;

[0118] sensor.pos_y: The y-coordinate of the radar installation position;

[0119] area_x: The difference in the x-direction between the sensor installation position and the tracked target;

[0120] area_y: The difference in the y-direction between the sensor installation position and the tracked target.

[0121] Calculate the first-order derivatives (dx / dt, dy / dt) and second-order derivatives (d 2 x / dt 2 , d 2 y / dt 2 ) of the target position to obtain the target motion trajectory characteristics, which are used to describe the smoothness of the target motion trajectory. Real targets usually have relatively stable and continuous motion trajectories, while the trajectories of ghosts are relatively messy and discontinuous.

[0122] In addition to the above-mentioned characteristic parameters, the characteristic parameters also include: the target lateral velocity (velX_mps);

[0123] the target longitudinal velocity (velY_mps);

[0124] the first-order derivative of the target lateral velocity (first_order_derivative_velX_mps);

[0125] the second-order derivative of the target lateral velocity (second_order_derivative_velX_mps);

[0126] the first-order derivative of the target longitudinal velocity (first_order_derivative_velY_mps), etc.

[0127] Step 2: Randomly divide the dataset into a training set and a test set; the division method is as follows: Combine the id (identity document, identity identifier) and timestamp (timestamp signal) of the targets in the dataset as the unique identifier, then divide the targets according to a certain ratio (80% training set, 20% test set), and finally extract the remaining relevant information through this unique identifier.

[0128] Step 3: Before entering the automated parameter tuning module, fix the random_state (random state) parameter to ensure the reproducibility of the results. The random_state parameter is used to control the randomness in the random forest. By fixing the random_state parameter, the randomness of these random processes can be controlled, making the random forest model generated each time the same;

[0129] Determine the value of the class_weight (weight of sample classes) parameter according to the quantity ratio of valid targets to ghost targets in the training set samples. By setting the class_weight parameter, the weight of the minority class can be increased, making the model pay more attention to the minority class, thereby improving the performance of the overall model;

[0130] Since the number of samples in different classes in the dataset varies greatly, this may cause the model to be biased towards the majority class during training and ignore the minority class. The class_weight parameter can be used to effectively handle the class imbalance problem;

[0131] After that, in the sub-module of the automated parameter tuning module, define the search space for feature parameters and use grid search to find the optimal combination of feature parameters: Use a weighted combination of the Gini impurity criterion and the information entropy criterion to obtain a comprehensive evaluation index α. By adjusting α, a trade-off can be made between these two aspects;

[0132] The Gini criterion focuses on local purity, and the information entropy criterion focuses on global uncertainty. Both are indicators reflecting data purity, but with different focuses;

[0133] In the initial stage, the data has not been fully divided and is still relatively chaotic as a whole. At this time, α approaches 0, that is, it is more inclined to use the information entropy as the division criterion; as the decision tree grows continuously, the data has been well divided at each node. At this time, α approaches 1, that is, it is more appropriate to use the Gini criterion as the division criterion because the Gini criterion pays more attention to local purity and can better fine-tune and optimize the division of each node. The calculation formula for node division is:

[0134] criterion = α * Gini + (1 - α) * Entropy

[0135] Gini = ∑p(x)(1 - p(x))

[0136] Entropy = -∑p(x)log 2 p(x)

[0137] Among them, criterion is the calculation standard for node division, Gini is the Gini coefficient, Entropy is the information entropy coefficient, and p(x) is the probability of class x.

[0138] Set the evaluation criterion for dividing subtrees (criterion), the range of the number of base estimators (n_estimators), the range of the minimum number of samples in leaf nodes (min_samples_leaf), and the range of the minimum number of samples required for further partitioning of internal nodes (min_samples_split);

[0139] The criterion parameter is used to specify the splitting criterion of the decision tree;

[0140] The n_estimators parameter is used to specify the number of decision trees in the random forest;

[0141] The min_samples_split parameter is used to specify the minimum number of samples required for node splitting. If the number of samples in a node is less than this value, the node will not be split.

[0142] The min_samples_leaf parameter is used to specify the minimum number of samples required for leaf nodes. If the number of samples in a leaf node is less than this value, the leaf node will be merged.

[0143] By reasonably setting these parameters, the training and prediction processes of the model can be controlled, and the performance and generalization ability of the model can be improved.

[0144] Step 4: Build a classifier model in the automatic hyperparameter tuning module. Use the random forest algorithm to build multiple decision trees, and then train the prediction results of the integrated decision tree by the majority voting principle; each tree is built based on a random subset of the training set, and each tree grows independently during the training process until it cannot be further split or reaches a preset stopping condition;

[0145] At the same time, pruning is used to prevent the overfitting phenomenon that the decision tree classifies the training data accurately but has poor classification effect on unknown test data.

[0146] Step 5: After the decision tree is trained, use the confusion matrix and the ROC (Receiver Operating Characteristic Curve) curve to input the test set samples into the classifier model, calculate the AUC (Area Under Curve, the area enclosed by the ROC curve and the coordinate axes) to evaluate the performance of the classifier model. Then, adopt the Monte Carlo method to build a framework for automatic parameter tuning. With the computing power of the computer, directly simulate Steps 3 and 4 above, continuously sample to find the classifier parameters that meet the expected recall rate and specificity, select the parameters with the most balanced recall rate and specificity, gradually approach the optimal result, and finally further screen out the optimal model parameters.

[0147] For each combination of feature parameters, train the model using the training set and evaluate its performance on the test set. When evaluating the model, taking the recall rate as the first evaluation criterion is to minimize the situation of wrongly suppressing valid targets as ghosts. According to the model evaluation results and combined with the AUC value, the optimal combination of feature parameters can be selected to achieve the best balance between the recall rate and the specificity.

[0148]

[0149] TP (True Positive): True positive, that is, the number of samples correctly predicted as the positive class by the model;

[0150] FN (False Negative): False negative, that is, the number of samples wrongly predicted as the negative class by the model;

[0151]

[0152] TN (True Negative): True negative, that is, the number of samples correctly predicted as the negative class by the model;

[0153] FP (False Positive): False positive, that is, the number of samples wrongly predicted as the positive class by the model.

[0154] Step 6: Convert the optimal model feature parameters into a.h file for easy integration and deployment in an embedded system or other hardware platforms, and integrate them into the target tracking sub-module to perform classification judgment on each target in real time to suppress ghost interference and improve the accuracy and reliability of target tracking. Use the target dynamic information calculated by the tracking module to directly feedback to the classification model. At the same time, directly feedback the classification result to the tracking module and combine it with multiple calculated attributes to promote the optimization of the overall classification performance.

[0155] The generated.h file is represented in the data structure of a binary tree. The data structure of the binary tree includes the TreesNode structure and the Trees structure. Compared with using traditional if-else (if conditional statements) to judge the value of each node, it can save about 80% of the occupied memory space and also improve the calculation speed.

[0156] This binary tree data structure can efficiently represent and deploy machine learning models and is very suitable for application scenarios on embedded systems and other hardware platforms.

[0157] The data structure of the binary tree is as follows:

[0158] The TreesNode structure is as follows:

[0159] typedef struct{

[0160] s8_t feature;

[0161] f32_t value;

[0162] s16_t left;

[0163] s16_t right;

[0164] }TreesNode;

[0165] The TreesNode structure is used to represent a node of a binary tree;

[0166] Feature: An 8-bit signed integer type (s8_t), representing the feature number checked by the current node;

[0167] Value: A 32-bit floating-point type (f32_t), representing the threshold of the current node;

[0168] Left: A 16-bit signed integer type (s16_t), representing the index of the left child node of the current node in the nodes array; if it is negative, it means this node is a leaf node and has no further child nodes;

[0169] Right: A 16-bit signed integer type (s16_t), representing the index of the right child node of the current node in the nodes array; if it is negative, it means this node is a leaf node and has no further child nodes.

[0170] The Trees structure is as follows:

[0171] typedef struct{

[0172] s32_t n_nodes;

[0173] TreesNode* nodes;

[0174] s32_t n_trees;

[0175] s32_t* tree_roots;

[0176] } Trees;

[0177] The Trees structure is used to represent a forest composed of multiple binary trees;

[0178] n_nodes: A 32-bit signed integer type (s32_t), representing the total number of nodes of all binary trees in the forest;

[0179] nodes: A pointer to an array of TreesNode structures, storing the node information of all binary trees;

[0180] n_trees: A 32-bit signed integer type (s32_t), representing the number of binary trees in the forest;

[0181] tree_roots: A pointer to an array of 32-bit signed integer type (s32_t), storing the index of the root node of each binary tree in the nodes array.

[0182] Step 7: In the radar-based object detection and tracking and environmental perception module, integrate the generated ghost suppression model into the target update sub-module. The target tracking sub-module inputs the feature parameters into the target update module for calculation. After reading the calculated feature parameters in the target update module, call the.h file to classify the target as a ghost target or a valid target. By recursively traversing each binary tree, make a decision based on the features and thresholds of the current node, post-process the classified results, and obtain a stable output result through counting with a counter.

[0183] Save the output result in is_ghost_by_machine_learning (whether the target is determined to be a ghost by machine learning). Subsequently, is_ghost_by_machine_learning can be combined with other attributes of the target, such as is_crossing (whether the target is crossing), is_created_near_fov (whether the target is created near the FOV edge), etc., to assign is_ghost (whether the target is a ghost) for targets in different complex environments.

[0184] To better achieve the objectives of the present invention, the present invention also provides a ghost suppression model constructed by using the above method.

[0185] Embodiment 2

[0186] In some embodiments, as Figure 2 shown, a confusion matrix is used to evaluate the finally generated model to calculate the F1_Score (the harmonic mean of precision and recall, which can reflect the overall performance of the model in a balanced manner). The calculation method of the F1_Score is as follows:

[0187]

[0188] From Figure 2 the data in, the value of the F1_Score can be obtained as 0.984.

[0189] Embodiment 3

[0190] In some embodiments, as Figures 3 - 4 shown, Figure 3 is a simulation diagram of the effect of suppressing ghost targets by traditional suppression methods, Figure 4 and a simulation diagram of the effect of suppressing ghost targets by using machine learning in a tunnel scenario. It can be seen that the present invention solves the limitation that free space cannot suppress ghost targets, and successfully suppresses the false targets generated by moving clutter points caused by angle measurement errors.

[0191] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A millimeter wave radar ghost suppression method based on machine learning, characterized in that: The following steps are involved: Step 1: Collect and preprocess the data of millimeter-wave radar target tracking under various environmental conditions, retrieve the data sequence corresponding to the target ID of each cycle in the HDF5 sequence file, perform feature processing on the data by retrieving the sequence name and target ID, and then extract the feature parameters that effectively distinguish the target signal and the interference signal, and then screen the feature parameters, and select the feature parameters with higher scores as the main input basis of the data set; Step 2: Randomly divide the data set into training set and test set; Step 3: In the submodule of the automatic parameter adjustment module, define the feature parameter search space and use grid search to find the optimal feature parameter combination: use the weighted combination of the Gini impurity criterion and the information entropy criterion to obtain a comprehensive evaluation index α, and by adjusting α, make a trade-off between the Gini impurity criterion and the information entropy criterion, and then set the evaluation criteria for dividing the subtree, the number range of the base evaluator, the range of the minimum number of samples for the leaf node, and the range of the minimum number of samples required for the internal node to be re-divided; by setting these parameters reasonably, the training and prediction process of the model can be controlled; Step 4: Build a classifier model in the automatic parameter adjustment module, use the random forest algorithm to build multiple decision trees, and then use the majority voting principle to determine the results of the integrated decision tree for training; Step 5: After the decision tree is trained, the test set samples are input into the classifier model to calculate and compare the recall rate through the confusion matrix and ROC curve, and the AUC data is calculated to evaluate the performance of the classifier model. Then, the Monte Carlo method is used to build an automatic parameter adjustment framework. With the help of the computer's computing power, the above steps 3 and 4 are directly simulated. The feature parameters of the classifier are constructed by continuous sampling to find the expected recall rate and specificity, and the feature parameters with the most balanced recall rate and specificity are selected to gradually approach the best result. Finally, the best model feature parameters are further screened. Step 6: Convert the optimal model feature parameters into a .h file. The generated .h file is represented by a binary tree data structure. Step 7: In the radar-based object detection and tracking and environment perception modules, the ghost suppression model is integrated into the target tracking submodule. The target tracking submodule inputs the feature parameters into the target update module for calculation. After reading the feature parameters calculated in the target update module, the .h file is called to classify the target into a ghost target or a valid target. Each binary tree is recursively traversed, and a decision is made based on the feature parameters and threshold of the current node. The classified results are post-processed and a stable output result is obtained after counting with a counter.

2. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 1, characterized in that: In step 1, the data preprocessing process: the collected radar target tracking data is compared with the visual truth value, and annotated to distinguish ghosts from real targets; If the visual truth matches the target output of the tracking model, the is_valid label is set to 1; If the visual truth does not match the target output of the tracking model, the is_valid label is set to 0; Where Z: target to be labeled, X: visual truth target, Y: model output target; None: No matching target exists; Otherwise: X i ∩Y j ≠ the complement of None.

3. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 2, characterized in that: In step one, the characteristic parameters include the maximum radar cross-section area of ​​the i point cloud clusters associated with the target in the current cycle, the average of the maximum radar cross-section areas of the point cloud clusters associated with the target from the creation cycle m to the current maintenance cycle n, the maximum radar cross-section area of ​​the point cloud clusters associated with the target from the creation cycle m to the current maintenance cycle n, the maximum SNR of the point cloud cluster associated with the target, the ratio of the number and area of ​​static targets between the sensor installation position and the tracked target, and the target motion trajectory characteristics.

4. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 3 is characterized in that: The calculation formula for the largest radar cross-section among the i point cloud clusters associated with the current period target is: object_track.max_rcs =max(associated_det0.rcs,associated_det1.rcs …associated_it i-1 .rcs,associated_it i .rcs) associated_det i : The i-th point cloud associated with the target; The calculation formula for the mean value of the maximum radar cross-section in the point cloud group associated with the target from the creation cycle m to the current maintenance cycle n is: The formula for calculating the maximum radar cross-section of the point cloud group associated with the target from the creation cycle m to the current maintenance cycle n is: The calculation formula for the ratio of the number of static targets to the area between the sensor installation position and the tracking target is: area_x=object_track.pos_x-sensor.pos_x area_y=object_track.pos_y-sensor.pos_y Among them, object_track.pos_x: the x coordinate of the target position in the coordinate system with the center of the vehicle's rear axle as the origin; object_track.pos_y: the y coordinate of the target position in the coordinate system with the center of the vehicle's rear axle as the origin; sensor.pos_x: x coordinate of the radar installation position; sensor.pos_y: y coordinate of the radar installation position; area_x: the difference in the x direction between the sensor installation position and the tracking target; area_y: the difference in y direction between the sensor installation position and the tracking target; The target motion trajectory characteristics are calculated by calculating the first-order derivative and second-order derivative of the target position.

5. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 4, characterized in that: In step three: When α is 0, the information entropy criterion is preferred as the node division criterion; When α is 1, the Gini impurity criterion is used as the node partitioning criterion.

6. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 5, characterized in that: Calculation method for node division: criterion=α*Gini+(1-α)*Entropy Gini=∑p(x)(1-p(x)) Entropy=-∑p(x)log2p(x) Among them, criterion is the calculation standard for node division, Gini is the Gini coefficient, Entropy is the information entropy coefficient, and p(x) is the probability of category x.

7. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 6, characterized in that: In step six, the data structure of the binary tree includes the TreesNode structure and the Trees structure.

8. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 7, characterized in that: The TreesNode structure is as follows: typedef struct{ s8_t feature; f32_t value; s16_t left; s16_t right; }TreesNode.

9. The method for suppressing ghost images of millimeter wave radar based on machine learning according to claim 8, characterized in that: The Trees structure is as follows: typedef struct{ s32_t n_nodes; TreesNode* nodes; s32_t n_trees; s32_t*tree_roots; }Trees.

10. A ghost suppression model constructed using the machine learning-based millimeter wave radar ghost suppression method according to claim 9.

Images