Vehicle-mounted millimeter wave radar target classification method and device and non-transitory computer readable storage medium

Abstract

Embodiments of the present application disclose a kind of vehicle-mounted millimeter wave radar target classification method, device and non-transitory computer readable storage medium, wherein, method includes: step S1, obtains the real-time point track and track information of vehicle-mounted millimeter wave radar;Step S2, calculate the preset multi-dimensional classification feature based on the point track and track information;Step S3, obtain the decision tree model pre-trained and pruned, the parameter and structure of the decision tree model are stored in the form of non-leaf node and leaf node and are deployed in vehicle-mounted embedded unit;Step S4, the multi-dimensional classification feature is input into the decision tree model, and the classification result of moving target is obtained by decision traversal.This application solves the problem that robustness, interpretability and real-time performance are difficult to be considered in prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle technology, specifically to a vehicle-mounted millimeter-wave radar target classification method, device, and non-transitory computer-readable storage medium. Background Technology

[0002] As Advanced Driver Assistance Systems (ADAS) and intelligent driving technologies evolve to higher levels, automotive millimeter-wave radar, with its all-weather, all-time operation capabilities and high-precision measurement of target distance and speed, has become an irreplaceable core sensor in environmental perception systems. In complex traffic scenarios, radar systems not only need to reliably detect targets but also accurately identify and classify target types (such as vehicles, pedestrians, and bicycles). This classification capability directly constitutes a key prerequisite for the correct decision-making and control of functions such as Automatic Emergency Braking (AEB), Adaptive Cruise Control (ACC), and Forward Collision Warning (FCW). Therefore, its accuracy and reliability are crucial to driving safety.

[0003] Currently, target classification methods based on millimeter-wave radar mainly rely on a series of features extracted from radar echoes. Common features include the target's radar cross-section (RCS), micro-Doppler spectral characteristics, point cloud distribution, and motion trajectory. A traditional approach is to use a rule-based classification method based on fixed thresholds, where target classification is achieved by manually setting discrimination intervals for specific features. For example, determining whether a target is a vehicle or a pedestrian is based solely on the numerical range of the RCS. However, in real-world scenarios, the target's RCS and other features change drastically with factors such as target attitude, relative distance, environmental clutter, and obstruction. A typical dangerous situation is that the RCS of a pedestrian at close range may be on the same order of magnitude as that of a small vehicle at a distance. In this case, a single threshold rule is highly susceptible to confusion, leading to classification errors and serious safety risks. This method, due to its over-reliance on the accuracy of a few features, suffers from poor robustness and a high misclassification rate in complex multi-target scenarios such as dense targets and obstructions.

[0004] To overcome the limitations of fixed threshold methods, another approach has been introduced using machine learning models, such as Support Vector Machines (SVMs) or shallow neural networks. These methods construct classifiers by fusing features from multiple dimensions, including velocity, acceleration, target size estimation, and reflection intensity distribution, resulting in significantly improved classification accuracy compared to fixed threshold methods. However, these models typically behave as "black box" systems, with internal classification boundaries often being complex nonlinear mappings that are difficult to understand intuitively, and the decision-making process lacks transparency. In the context of automotive applications emphasizing functional safety and Safety for Expected Functions (SOTIF), this lack of interpretability makes it difficult for engineers to trace and verify the causes of each classification result. When misclassification occurs, debugging and model optimization become extremely difficult, hindering the rapid iteration and certification of safety-critical systems. Furthermore, the increased complexity of the model structure often leads to higher computational resource overhead, adding challenges to the real-time performance of automotive embedded processing units and software deployment.

[0005] In summary, existing target classification technologies based on millimeter-wave radar have the following main shortcomings: First, rule-based classification methods based on fixed thresholds have poor robustness and are difficult to cope with non-stationary changes in target features in real-world dynamic scenarios. They also suffer from high classification error rates in complex multi-target situations, directly impacting functional safety. Second, while methods based on complex machine learning models improve classification performance, they suffer from poor model interpretability, difficulty in debugging and verification, and a significant contradiction exists between the real-time requirements and resource overhead of vehicle deployment, making it difficult to meet the demands for rapid deployment and continuous iteration in safe and reliable applications. Third, the efficiency of comprehensively utilizing multi-dimensional radar features is low. Existing methods either rely on fragile single-dimensional or low-dimensional linear judgments or uninterpretable high-dimensional nonlinear mappings, lacking a classification mechanism that can effectively integrate multi-dimensional radar features while possessing a clear logical hierarchy and good interpretability. Summary of the Invention

[0006] The technical problem to be solved by the embodiments of the present invention is to provide a vehicle-mounted millimeter-wave radar target classification method, device and non-transitory computer-readable storage medium, so as to ensure high classification accuracy, good interpretability, easy debugging and verification, and meet the real-time requirements of vehicle-mounted platforms.

[0007] To address the aforementioned technical problems, as one aspect of the present invention, a target classification method for vehicle-mounted millimeter-wave radar is provided, comprising: Step S1: Obtain real-time point and track information from the vehicle-mounted millimeter-wave radar; Step S2: Calculate the preset multidimensional classification features based on the point and track information; Step S3: Obtain the pre-trained and pruned decision tree model. The parameters and structure of the decision tree model are deployed in the vehicle embedded unit in the form of separate storage of non-leaf nodes and leaf nodes. Step S4: Input the multidimensional classification features into the decision tree model, and obtain the classification result of the moving target through decision traversal.

[0008] Preferably, in step S2, the preset multidimensional classification features include at least one of the following dimensions or any combination thereof: the radar cross-section of the target and its statistics, the estimated size of the target, the target's speed, acceleration, micro-Doppler spectrum distribution features, point cloud distribution morphology features, the number of frames the target exists in and the trajectory length, as well as the mean and variance of the above features within a preset time window.

[0009] Preferably, in step S4, the classification result of the moving target specifically includes four categories: large vehicles, small vehicles, two-wheeled vehicles, and pedestrians.

[0010] Preferably, in step S3, the decision tree model is obtained through the following training process: We collect vehicle millimeter-wave radar spot and track data and synchronous video data under various driving scenarios. We label the track with target type through video data and extract multi-dimensional classification features of each track in each frame to form a sample dataset containing features and corresponding labels. The label categories include large vehicles, small vehicles, two-wheeled vehicles and pedestrians. After arranging the sample dataset in chronological order, a predetermined proportion of the data at the beginning of each category is taken as the training and validation set, and the remaining data at the end is taken as the test set, so that the test set is independent of the training and validation set in time. The decision tree model is trained based on the training and validation set. A class weighting mechanism is used to handle data imbalance, and feature selection and hyperparameter configuration are performed. The fully grown decision tree is then pruned using cross-validation. The pruned model is evaluated using the test set until the classification accuracy meets the preset index.

[0011] Preferably, the category weighting mechanism is as follows: count the number of samples in each category in the training set, set a weight for each category, and the weight is the reciprocal of the proportion of the number of samples in that category to the total number of samples.

[0012] Preferably, the feature selection adopts an iterative selection method based on feature importance: in the first round, all features are used for training to obtain the importance weight of each feature; in subsequent rounds, the top N features with higher importance weights are selected from the results of the previous round for training, where N is a preset positive integer and N is not greater than the total number of features M.

[0013] Preferably, in step S3, the method of storing non-leaf nodes and leaf nodes separately specifically means: Store the feature indices and decision thresholds of all non-leaf nodes in the first array, and store the target category labels corresponding to all leaf nodes in the second array; During the decision traversal, at each non-leaf node, the input feature value is obtained according to the corresponding feature index in the first array, and the branch jump is determined after comparing it with the corresponding decision threshold until the leaf node is reached. Then, the corresponding category label is read from the second array as the classification result output.

[0014] Preferably, the first array and the second array are written into the header file as static array constants, serving as the parameters and structure of the decision tree model, and are integrated into the application code of the vehicle-mounted embedded unit.

[0015] The present invention also provides a vehicle-mounted millimeter-wave radar target classification device, comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the vehicle-mounted millimeter-wave radar target classification method.

[0016] The present invention also provides a non-transitory computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the vehicle-mounted millimeter-wave radar target classification method.

[0017] Implementing the embodiments of this invention has the following beneficial effects: This invention significantly improves the robustness and scene adaptability of target classification: This invention abandons the judgment method that relies on a single or a few fixed thresholds, and instead adopts a multi-dimensional feature fusion classification mechanism based on decision trees. By incorporating multi-dimensional features such as radar cross-section, velocity, size estimation, and micro-Doppler into a hierarchical conditional judgment framework, the classification decision can comprehensively consider the various physical and motion characteristics of the target, effectively avoiding misjudgments caused by fluctuations in a single feature value due to factors such as distance and attitude. In complex and ever-changing real-world traffic scenarios, it can maintain stable and reliable classification performance for various targets such as large vehicles, small vehicles, two-wheeled vehicles, and pedestrians. This invention also achieves high interpretability and easy debugging of the classification decision process: This invention uses a decision tree as the classification model, and its tree-like judgment structure naturally possesses "white-box" characteristics: each classification inference from the root node to the leaf node forms a clear feature judgment path, and each judgment node corresponds to an intuitive feature threshold comparison rule. When misclassification occurs, engineers can trace the classification criteria node by node along the path to quickly locate the root cause of the problem and perform targeted optimizations. This greatly reduces the complexity of algorithm debugging and iteration, fully meeting the stringent requirements of automotive functional safety for interpretability and verifiability. This invention achieves an effective balance between high classification accuracy and the real-time performance and low resource consumption of the in-vehicle platform. The decision tree model has a simplified structure, and online inference involves only a limited number of feature value comparison operations. Its computational complexity is far lower than that of neural network models, and it can achieve low-latency real-time inference on in-vehicle embedded processors without dedicated acceleration hardware. At the same time, by storing non-leaf nodes and leaf nodes separately, the model compression deployment method further saves about 40% of storage space. This allows the method to maintain high classification accuracy while also possessing the advantages of high efficiency and low power consumption, making it easy to deploy on a large scale on resource-constrained in-vehicle platforms. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a vehicle-mounted millimeter-wave radar target classification method according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the training, pruning optimization, and deployment process of the decision tree model used in Embodiment 1 of the present invention.

[0021] Figure 3 This is a schematic diagram of the confusion matrix of the validation set model in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram illustrating the importance ranking of features selected through training in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the pruning curve for pruning optimization of the model in an embodiment of the present invention.

[0024] Figure 6 This is a schematic diagram of the confusion matrix of the test set model in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following illustrative embodiments and descriptions are only used to explain the technical principles and feasible solutions of the present invention, and are not intended to limit the present invention. Moreover, in the absence of conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Please refer to Figure 1 As shown, Embodiment 1 of the present invention provides a target classification method for vehicle-mounted millimeter-wave radar, including: Step S1: Obtain real-time point and track information from the vehicle-mounted millimeter-wave radar; Step S2: Calculate the preset multidimensional classification features based on the point and track information; Step S3: Obtain the pre-trained and pruned decision tree model. The parameters and structure of the decision tree model are deployed in the vehicle embedded unit in the form of separate storage of non-leaf nodes and leaf nodes. Step S4: Input the multidimensional classification features into the decision tree model, and obtain the classification result of the moving target through decision traversal.

[0027] As described above, this invention addresses the problems of poor robustness of fixed threshold rules, insufficient interpretability of complex machine learning models, and deployment difficulties in existing target classification schemes. It proposes a target classification method centered on a decision tree, achieving both high classification accuracy and good interpretability. When this method performs real-time inference on a real-world in-vehicle embedded platform, it includes the following main steps: Step S1: Obtain real-time point and track information from the vehicle-mounted millimeter-wave radar.

[0028] In this step, the perception data output by the vehicle-mounted millimeter-wave radar is acquired in real time. Specifically, the millimeter-wave radar continuously tracks moving targets within the detection area, outputting point trace information and trajectory information for each frame. Point trace information includes basic measurement values ​​such as the target reflection point position, velocity, and reflection intensity obtained from a single radar measurement; trajectory information is the tracking trajectory data formed by associating multiple consecutive frames of point traces of the same target, including track identifiers (Track ID), target position, velocity, acceleration, heading angle, and other motion state estimates, as well as trajectory correlations. This information forms the data foundation for subsequent feature calculations.

[0029] Step S2: Calculate the preset multidimensional classification features based on the point and track information.

[0030] After obtaining real-time point and track information, a set of multi-dimensional features for classification is calculated in real time for each moving target track in each frame. These features are physical and statistical features extracted from the raw radar measurement data that have the ability to distinguish different target types.

[0031] As an example, multidimensional classification features include, but are not limited to, the following dimensions: the target's radar cross-section (RCS) and its statistical value, target size estimation parameters, target velocity, acceleration, micro-Doppler spectrum distribution characteristics, point cloud distribution morphology characteristics, the number of frames in which the target exists and the trajectory length, as well as derived features such as the mean and variance of the above features within a time window. These features characterize the differences in physical and motion characteristics exhibited by different target types, such as large vehicles, small vehicles, two-wheeled vehicles, and pedestrians, in radar echoes from different perspectives. For example, large vehicles typically have larger RCS and size estimates, while pedestrians exhibit smaller RCS and unique micro-Doppler characteristics, and the lateral maneuverability and acceleration variation patterns of two-wheeled vehicles differ from other categories. By calculating such a set of pre-selected multidimensional features, discriminative input can be provided to the decision tree model.

[0032] Step S3: Obtain the pre-trained and pruned decision tree model. The parameters and structure of the decision tree model are deployed in the vehicle embedded unit in the form of separate storage of non-leaf nodes and leaf nodes.

[0033] In actual operation, the classification method uses a decision tree model that has been trained offline and deployed to the in-vehicle processing unit. The training and optimization process of this model will be described in detail later. It is worth noting that, to meet the stringent requirements of the in-vehicle embedded platform for storage space and computational efficiency, the decision tree model employs an optimized storage structure during deployment. Specifically, the decision tree consists of two types of nodes: non-leaf nodes and leaf nodes. Non-leaf nodes store the feature index and decision threshold of the current node, used to determine the branch direction during inference based on the comparison between feature values ​​and thresholds. Leaf nodes store the target category label corresponding to that leaf. During model deployment, the parameters of the non-leaf nodes and leaf nodes of the tree structure are stored separately as independent arrays or tables, rather than directly storing the complete tree structure. Compared to the overhead of storing the child node pointers and other tree structure elements of each node, this separate storage method can significantly save storage space, with actual measurements showing a saving of approximately 40% in model storage overhead. After this compression, the decision tree model parameters and structural information are exported as a .h header file and embedded into the application code of the in-vehicle embedded unit as static array constants.

[0034] Step S4: Input the multidimensional classification features into the decision tree model, and obtain the classification result of the moving target through decision traversal.

[0035] During real-time inference, the multidimensional classification feature vector of a trajectory calculated in step S2 is input into the deployed decision tree model, and decision traversal is performed layer by layer along the tree structure starting from the root node. Specifically, at each non-leaf node, the feature index corresponding to that node is read, the feature value of the corresponding dimension in the input feature vector is obtained, and compared with the decision threshold stored in the node; based on the comparison result, the model jumps to the next level node along the left or right subtree branch. This process is repeated until a leaf node is reached. The target category label stored in this leaf node is the target type of the current trajectory. The final output classification result includes at least one of the following categories: large vehicles, small vehicles, two-wheeled vehicles, and pedestrians.

[0036] Through the above steps, this invention achieves real-time classification of moving targets using millimeter-wave radar on an in-vehicle embedded platform. Compared to traditional fixed threshold rules, decision trees can perform hierarchical, conditional classification based on multi-dimensional features, avoiding misjudgments caused by fluctuations in a single feature and exhibiting stronger robustness. Furthermore, compared to black-box models such as neural networks, the decision-making process of decision trees is inherently interpretable: each classification decision can be clearly traced back to the feature determination path from the root node to the leaf node, making it easy for engineers to understand, verify, and debug. In addition, the decision tree model itself has low computational cost and fast traversal speed, making it naturally suitable for real-time in-vehicle applications.

[0037] Below, in conjunction with Figure 2The training, pruning, optimization, and deployment process of the decision tree model used in step S3 above is described in detail. It should be understood that this training process itself can be completed on a server or offline computing platform, without consuming real-time operating resources of the in-vehicle system.

[0038] (I) Dataset Construction First, a feature dataset needs to be constructed for model training and evaluation. The specific process is as follows: We collect vehicle-mounted millimeter-wave radar data in various real-world driving scenarios, including but not limited to urban road scenarios, highway scenarios, and collaborative data collection scenarios. The collected data includes: lists of points and tracks output by the millimeter-wave radar, track correlation information, and synchronously acquired forward real-time video data. The video data serves as a basis for providing ground truth labels for target types in subsequent steps.

[0039] By simultaneously replaying radar and video data, the frame number range and corresponding track ID of each moving target track appearing in the scene are recorded. The actual target type of each track is then determined manually based on the video images. Target type tags include at least four categories: large vehicles, small vehicles, two-wheeled vehicles, and pedestrians. The recorded tag information is stored in a tag CSV file.

[0040] Using MATLAB or other data processing scripts, the radar data is parsed packet by packet, including point traces, tracks, and track association information, and the corresponding label CSV files are loaded. For each track instance present in each frame, a set of features that may be effective for classification is calculated, forming a sample. All samples constitute a feature dataset, where each sample contains feature values ​​across multiple dimensions, and the last column is the corresponding target type label.

[0041] (II) Dataset Partitioning To avoid bias in model evaluation caused by temporal continuity, the dataset is divided chronologically. After arranging the sample dataset in chronological order, a predetermined proportion of the data at the beginning of each category is used as the training and validation set, and the remaining data at the end is used as the test set, making the test set temporally independent of the training and validation sets.

[0042] Specifically, all data packets are arranged chronologically. For each category, the first 70% of the data samples are assigned to the training and validation set, and the last 30% are assigned to the test set. This division ensures that the test set is completely independent of the training and validation sets in time, and can more realistically reflect the model's ability to generalize to data from unknown future scenarios. It also avoids the problem of overly optimistic evaluations caused by random sampling, where samples from similar time periods appear in both the training and test sets.

[0043] (III) Model Training and Evaluation (1) Handling data imbalance.

[0044] In real-world road scenarios, the number of samples from small vehicles is typically far greater than that from large vehicles, two-wheeled vehicles, and pedestrians. To mitigate the negative impact of data imbalance on model training, a class weight compensation mechanism can be employed. Specifically, the number of samples in each category in the training set is statistically analyzed, and a training weight is assigned to each category. This weight is the reciprocal of the proportion of samples from that category to the total number of samples. This setting amplifies the contribution of the category with the smaller sample size to the training loss function, thus prompting the model to focus more on the correct classification of the minority category.

[0045] For categories with severely insufficient sample sizes, such as pedestrians and two-wheeled vehicles, additional data collected specifically in collaborative scenarios can be added to further enrich the sample diversity of these few categories.

[0046] (2) Division of training and validation data.

[0047] Within the training and validation sets, 70% of the samples are randomly sampled as the actual training set, and the remaining 30% are used as the model validation set. The validation set is used to monitor model performance during training and to assist in hyperparameter tuning and model selection.

[0048] (3) Feature selection and parameter configuration.

[0049] Each sample in the dataset contains M features that may be effective for classification, but not all features have equal discriminative power. To build a concise and efficient decision tree model, N features (N≤M) with higher importance weights can be selected. Feature selection can be performed in two ways: one is to use a recursive feature elimination algorithm to automatically select a specified number of superior feature subsets; the other is to use a manual iterative selection method based on feature importance, that is, all M features are used in the first round of training, and after training, the importance weight histogram of each feature is obtained, based on which the top N features with higher importance are selected for subsequent rounds of training. In addition, during the training process, hyperparameters such as the maximum tree depth and the minimum number of leaf node samples need to be set to control the tree size and the degree of overfitting.

[0050] (4) Model training and pruning.

[0051] During the training phase, a relatively large tree depth can be initially set to allow the decision tree to grow fully and obtain a relatively complete classification tree. Subsequently, cross-validation is used to prune the grown decision tree. Pruning involves removing subtree branches that do not contribute significantly to improving validation accuracy from the bottom up and replacing them with leaf nodes. This reduces the model size while maintaining classification accuracy, lowers the risk of overfitting, and improves the model's generalization ability on unknown data.

[0052] (5) Model evaluation.

[0053] The performance of the pruned decision tree model is evaluated using a test set that is independent of the training process. The model's classification accuracy is calculated, and a confusion matrix is ​​used to visualize the model's classification accuracy for each category and the confusion between categories. Simultaneously, histograms of feature importance are output to perform feature importance analysis and verify the effectiveness and rationality of the selected features.

[0054] (6) Iterative optimization.

[0055] The model evaluation results determine whether the preset accuracy target has been met. For example, if the average classification accuracy of the pruned model on the test set reaches over 98%, training ends and the model deployment phase begins. If the target is not met, the process returns to the feature selection and parameter configuration steps, adjusts the selected feature set or hyperparameter settings, and retrains and evaluates until the requirements are met. This judgment process can be manually decided by R&D personnel based on the evaluation results.

[0056] (iv) Model storage and deployment After training, the finalized decision tree model parameters and structure are exported for deployment on the vehicle platform. To conserve valuable storage resources in the embedded unit, the aforementioned strategy of storing non-leaf nodes and leaf nodes separately is adopted: feature indices and decision thresholds for all non-leaf nodes, and category labels for all leaf nodes are extracted and stored in two independent arrays, and the traversal logic of the decision tree is reconstructed from the data. These arrays are written into a .h header file and defined as static array constants in C language. Integrating this header file into the vehicle embedded software project, along with the corresponding feature calculation and decision traversal functions, enables real-time millimeter-wave radar moving target classification on the vehicle radar signal processing or perception fusion unit.

[0057] To verify the effectiveness of the method proposed in this patent, road test data from various scenarios were collected for verification.

[0058] A total of 670,068 samples were collected, with 70% (469,054 samples) assigned to the training and validation sets, and 30% (201,014 samples) assigned to the test set. The distribution of each category in the training and validation sets is shown below:

[0059] Visualization of the confusion matrix of the validation set model, as shown below Figure 3 As shown, in the confusion matrix, rows represent the true class, columns represent the predicted class, the percentage in the first column on the right is the target classification success accuracy, the second column is the classification failure accuracy, the diagonal of the matrix is ​​the number of correctly classified samples, and the remaining positions are the number of misclassified samples.

[0060] The training process ultimately selected 15 high-scoring features, and these features were ranked by importance. The results are as follows: Figure 4 As shown.

[0061] The model is then pruned and optimized, and the pruning curve is as follows: Figure 5 As shown, the optimal pruning level is 120, and the total number of nodes after pruning is 2061.

[0062] The accuracy of the classification model is tested using a test set. The confusion matrix of the test set is as follows: Figure 5 As shown, the overall classification accuracy rate is 88.5%, the accuracy rate for two-wheeled vehicles is 82.4%, the accuracy rate for small vehicles reaches 89%, the accuracy rate for large vehicles is 86.8%, and the accuracy rate for pedestrians reaches 92.7%.

[0063] The following is the distribution of each category in the test set:

[0064] Visualization of the confusion matrix of the test set model, as shown below Figure 6 As shown.

[0065] Corresponding to the vehicle-mounted millimeter-wave radar target classification method described in Embodiment 1 of the present invention, Embodiment 2 of the present invention also provides a vehicle-mounted millimeter-wave radar target classification device, comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the vehicle-mounted millimeter-wave radar target classification method as described in Embodiment 1.

[0066] Corresponding to the vehicle-mounted millimeter-wave radar target classification method described in Embodiment 1 of the present invention, Embodiment 3 of the present invention also provides a non-transitory computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the vehicle-mounted millimeter-wave radar target classification method as described in Embodiment 1.

[0067] Preferably, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the device, connecting various parts of the device through various interfaces and lines.

[0068] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0069] It should be noted that the above-mentioned device may include, but is not limited to, a processor and a memory. Those skilled in the art will understand that the above description does not constitute the only limitation on the device structure.

[0070] For the working principle and process of the above embodiments, please refer to the description of Embodiment 1 of the present invention, which will not be repeated here.

[0071] As can be seen from the above description, compared with the prior art, the beneficial effects of the present invention are as follows: The present invention significantly improves the robustness and scene adaptability of target classification: The present invention abandons the judgment method that relies on a single or a few fixed thresholds, and instead adopts a multi-dimensional feature fusion classification mechanism based on decision trees. By incorporating multi-dimensional features such as radar cross-section, velocity, size estimation, and micro-Doppler into a hierarchical conditional judgment framework, the classification decision can comprehensively consider the multiple physical and motion characteristics of the target, effectively avoiding misjudgments caused by fluctuations in a single feature value with factors such as distance and attitude. In complex and ever-changing real traffic scenarios, it can maintain stable and reliable classification performance for various targets such as large vehicles, small vehicles, two-wheeled vehicles, and pedestrians. The present invention also achieves a high degree of interpretability and easy debugging of the classification decision process: The present invention uses a decision tree as the classification model, and its tree-like judgment structure naturally has the characteristics of a "white box": each classification reasoning from the root node to the leaf node forms a clear feature judgment path, and each judgment node corresponds to an intuitive feature threshold comparison rule. When misclassification occurs, engineers can trace the classification criteria node by node along the path to quickly locate the root cause of the problem and perform targeted optimizations. This greatly reduces the complexity of algorithm debugging and iteration, fully meeting the stringent requirements of automotive functional safety for interpretability and verifiability. This invention achieves an effective balance between high classification accuracy and the real-time performance and low resource consumption of the in-vehicle platform. The decision tree model has a simplified structure, and online inference involves only a limited number of feature value comparison operations. Its computational complexity is far lower than that of neural network models, and it can achieve low-latency real-time inference on in-vehicle embedded processors without dedicated acceleration hardware. At the same time, by storing non-leaf nodes and leaf nodes separately, the model compression deployment method further saves about 40% of storage space. This allows the method to maintain high classification accuracy while also possessing the advantages of high efficiency and low power consumption, making it easy to deploy on a large scale on resource-constrained in-vehicle platforms.

[0072] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many variations under the guidance of the present invention without departing from the inventive spirit and scope of protection of the claims. These variations are all within the scope of protection of the present invention.

Claims

1. A target classification method for vehicle-mounted millimeter-wave radar, characterized in that, include: Step S1: Obtain real-time point and track information from the vehicle-mounted millimeter-wave radar; Step S2: Calculate the preset multidimensional classification features based on the point and track information; Step S3: Obtain the pre-trained and pruned decision tree model. The parameters and structure of the decision tree model are deployed in the vehicle embedded unit in the form of separate storage of non-leaf nodes and leaf nodes. Step S4: Input the multidimensional classification features into the decision tree model, and obtain the classification result of the moving target through decision traversal.

2. The method according to claim 1, characterized in that, In step S2, the preset multidimensional classification features include at least one of the following dimensions or any combination thereof: the radar cross-section of the target and its statistics, the estimated size of the target, the target's speed, acceleration, micro-Doppler spectrum distribution features, point cloud distribution morphology features, the number of frames the target exists in and the trajectory length, as well as the mean and variance of the above features within a preset time window.

3. The method according to claim 1, characterized in that, In step S4, the classification results of the moving target specifically include four categories: large vehicles, small vehicles, two-wheeled vehicles, and pedestrians.

4. The method according to claim 1, characterized in that, In step S3, the decision tree model is obtained through the following training process: We collect vehicle millimeter-wave radar spot and track data and synchronous video data under various driving scenarios. We label the track with target type through video data and extract multi-dimensional classification features of each track in each frame to form a sample dataset containing features and corresponding labels. The label categories include large vehicles, small vehicles, two-wheeled vehicles and pedestrians. After arranging the sample dataset in chronological order, a predetermined proportion of the data at the beginning of each category is taken as the training and validation set, and the remaining data at the end is taken as the test set, so that the test set is independent of the training and validation set in time. The decision tree model is trained based on the training and validation set. A class weighting mechanism is used to handle data imbalance, and feature selection and hyperparameter configuration are performed. The fully grown decision tree is then pruned using cross-validation. The pruned model is evaluated using the test set until the classification accuracy meets the preset index.

5. The method according to claim 4, characterized in that, The category weighting mechanism is as follows: count the number of samples in each category in the training set, and set a weight for each category. The weight is the reciprocal of the proportion of the number of samples in that category to the total number of samples.

6. The method according to claim 4, characterized in that, The feature selection adopts an iterative selection method based on feature importance: in the first round, all features are used for training to obtain the importance weight of each feature; in subsequent rounds, the top N features with higher importance weights are selected from the results of the previous round for training, where N is a preset positive integer and N is not greater than the total number of features M.

7. The method according to claim 1, characterized in that, In step S3, the method of storing non-leaf nodes and leaf nodes separately specifically means: Store the feature indices and decision thresholds of all non-leaf nodes in the first array, and store the target category labels corresponding to all leaf nodes in the second array; During the decision traversal, at each non-leaf node, the input feature value is obtained according to the corresponding feature index in the first array, and the branch jump is determined after comparing it with the corresponding decision threshold until the leaf node is reached. Then, the corresponding category label is read from the second array as the classification result output.

8. The method according to claim 7, characterized in that, The first array and the second array are written into the header file as static array constants, serving as the parameters and structure of the decision tree model, and are integrated into the application code of the vehicle embedded unit.

9. A vehicle-mounted millimeter-wave radar target classification device, characterized in that, include: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the vehicle-mounted millimeter-wave radar target classification method as described in any one of claims 1-8.

10. A non-transitory computer-readable storage medium, characterized in that, It stores computer instructions, which, when executed by a processor, implement the vehicle-mounted millimeter-wave radar target classification method as described in any one of claims 1-8.

Images