Vehicle point cloud wind resistance coefficient integrated prediction method and system based on feature recognition

Abstract

The invention provides a vehicle point cloud wind resistance coefficient integrated prediction method and system based on feature recognition, and relates to the technical field of wind resistance coefficient prediction, and the method specifically comprises the steps: obtaining the point cloud data of a vehicle model, and carrying out the preprocessing of the point cloud data; constructing a classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting vehicle type categories and probability values corresponding to the vehicle type categories; constructing a plurality of specific models in one-to-one correspondence with the vehicle type categories and a generalization model; whether the maximum probability value in the vehicle type classification labels is higher than a preset threshold value or not is judged, if yes, a specific model corresponding to the vehicle type category is selected to conduct wind resistance coefficient prediction, and if not, a generalization model is selected to conduct wind resistance coefficient prediction; according to the method, the vehicle types are classified and matched with the proper wind resistance coefficient prediction model, so that the prediction speed and the accuracy of a prediction result are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of aerodynamic drag coefficient prediction, and in particular to an integrated prediction method and system for the aerodynamic drag coefficient of a whole vehicle based on feature recognition. Background Art

[0002] With the development of the automotive industry, aerodynamic performance, especially the aerodynamic drag coefficient of the whole vehicle, has become an important indicator for measuring vehicle performance, improving fuel economy and reducing emissions.

[0003] However, in the prior art, the prediction methods of the aerodynamic drag coefficient are mostly based on traditional computational fluid dynamics (CFD) simulations, wind tunnel experiments or mathematical modeling. However, these methods generally have problems such as complex calculations, long time consumption or reliance on a large amount of experimental data. Especially in the process of whole vehicle design, a large amount of wind tunnel experimental data or high-precision CFD calculations are required, making it difficult to quickly and real-time optimize the vehicle design scheme. Summary of the Invention

[0004] The purpose of the present invention is to provide an integrated prediction method and system for the aerodynamic drag coefficient of a whole vehicle based on feature recognition, so as to solve the problem in the above background art that the prior art requires a large amount of wind tunnel experimental data or high-precision CFD calculations and is difficult to quickly and real-time optimize the vehicle design scheme.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An integrated prediction method for the aerodynamic drag coefficient of a whole vehicle based on feature recognition, comprising: acquiring point cloud data of a whole vehicle model and preprocessing the point cloud data; constructing a classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting a vehicle type classification label, wherein the vehicle type classification label includes a vehicle type category and a probability value corresponding to the vehicle type category; constructing an aerodynamic drag coefficient prediction model library, wherein the aerodynamic drag coefficient prediction model library includes a number of dedicated models corresponding one-to-one to the vehicle type categories and a generalization model; judging whether the maximum probability value in the vehicle type classification label is higher than a preset threshold, if so, selecting the dedicated model corresponding to the vehicle type category for aerodynamic drag coefficient prediction, if not, selecting the generalization model for aerodynamic drag coefficient prediction.

[0006] Optionally, the steps of the preprocessing specifically include: performing regularization processing on the point cloud data; performing augmentation processing on the regularized point cloud data; performing maximum-minimum normalization processing on the augmented point cloud data.

[0007] Optionally, the step of constructing the classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting the vehicle type classification label specifically includes: constructing a PointNet classification network model, processing the point cloud data through a number of MLP layers with the same structure, generating a global feature vector through a max pooling layer, then extracting features and classifying through a fully connected layer, and finally outputting the probability value corresponding to each vehicle type category through a softmax layer.

[0008] Optionally, before the step of constructing the PointNet classification network model, it further includes: constructing a spatial transformation network model to perform a global spatial transformation on the point cloud data.

[0009] Optionally, the vehicle type classification label includes: a fastback vehicle type and its corresponding probability value ; a station wagon vehicle type and its corresponding probability value ; a notchback vehicle type and its corresponding probability value ; where , , , and .

[0010] Optionally, the prediction threshold is , and .

[0011] Optionally, the dedicated model is trained with the point cloud data of the corresponding vehicle type category, and the generalization model is trained with the point cloud data of all vehicle type categories.

[0012] Optionally, each model in the wind resistance coefficient prediction model library is based on a dynamic graph edge convolution network. Using the KNN algorithm to calculate the distance between the input point cloud data and its nearest neighbor points and construct an adjacency graph to capture the local features of the point cloud data; calculating the relative position vectors of each point through an edge convolution EdgeConv operation, and combining a multi-layer perceptron MLP to transform and update the local features to extract the global features of the vehicle point cloud; aggregating the local features and the global features, and outputting the wind resistance coefficient prediction value through a fully connected layer.

[0013] Optionally, the step of aggregating the local features and the global features includes: aggregating the local features and the global features through max pooling and / or average pooling.

[0014] On the other hand, the present invention also provides an integrated prediction system for the vehicle point cloud drag coefficient based on feature recognition, including: a preprocessing module for obtaining the point cloud data of the vehicle model and preprocessing the point cloud data; a classification module for constructing a classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting a vehicle type classification label, wherein the vehicle type classification label includes the vehicle type category and the probability value corresponding to the vehicle type category; a prediction module for constructing a drag coefficient prediction model library, wherein the drag coefficient prediction model library includes a number of dedicated models corresponding one-to-one to the vehicle type categories and a generalization model; a matching module for determining whether the maximum probability value in the vehicle type classification label is higher than a preset threshold. If so, select the dedicated model corresponding to the vehicle type category for drag coefficient prediction. If not, select the generalization model for drag coefficient prediction.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The present application first constructs a classification network model, inputs the preprocessed point cloud data into the classification network model for vehicle type classification, and outputs a vehicle type classification label; then, constructs a drag coefficient prediction model library, which includes a number of dedicated models corresponding one-to-one to the vehicle type categories and a generalization model; finally, selects and matches with the dedicated model or the generalization model through the probability value in the classification label, and then predicts the drag coefficient. By classifying the vehicle type and matching a suitable drag coefficient prediction model, compared with the traditional method of predicting through a unified model, which requires a large amount of calculation and has a large calculation error, the prediction speed and the accuracy of the prediction result are greatly improved. Description of the Drawings

[0016] Figure 1 It is a schematic flowchart of the method steps of the present invention.

[0017] Figure 2 It is a schematic diagram of the overall framework of the system of the present invention.

[0018] Figure 3 It is a schematic diagram of the construction of the dedicated model database of the present invention.

[0019] Figure 4 It is a schematic flowchart of the feature recognition workflow based on the PointNet network of the present invention.

[0020] Figure 5 It is a schematic diagram of the vehicle point cloud drag coefficient prediction model of the present invention.

[0021] Figure 6 It is a schematic diagram of the system structure of the present invention.

[0022] In the figure: 10 - preprocessing module, 20 - classification module, 30 - prediction module, 40 - matching module. Detailed implementation mode

[0023] Next, the solutions of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned accompanying drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances, so as to implement the embodiments of the present application described here. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily need to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0025] Those skilled in the art of this technology can understand that unless specifically stated, the singular forms "a", "an", "the" and "said" used here can also include the plural form. It should be further understood that the term "include" used in the specification of this application means the presence of features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say that an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used here can include wireless connection or wireless coupling. The phrase "and / or" used here includes all or any unit and all combinations of one or more related listed items.

[0026] Those skilled in the art of this technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used here have the same meaning as the general understanding of those of ordinary skill in the field to which this application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless specifically defined as here.

[0027] It should be understood that the sequence numbers and magnitudes of the steps in this embodiment do not mean the order of execution. The order of execution of each process is determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of this application.

[0028] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The following will detail the present application with reference to the drawings and in combination with the embodiments.

[0029] Please refer to Figure 1 - Figure 2 , a method for integrated prediction of the vehicle point cloud drag coefficient based on feature recognition according to the present invention, comprising: S100. Obtain the point cloud data of the vehicle model and preprocess the point cloud data.

[0030] Specifically, after starting the program, input the vehicle model that has undergone format conversion, and generate point cloud data by converting the vehicle model through a file. Each point cloud consists of multiple points, and each point contains spatial coordinates (x, y, z). These points present the three-dimensional geometric features of the vehicle in space; the preprocessing includes at least one of regularization, enhancement, maximum-minimum normalization, and denoising, which can improve the data quality and facilitate the subsequent network model to process the point cloud data.

[0031] S200. Construct a classification network model, input the preprocessed point cloud data into the classification network model for vehicle model classification, and output a vehicle model classification label, where the vehicle model classification label includes the vehicle model category and the probability value corresponding to the vehicle model category.

[0032] Specifically, construct a classification network model, use the point cloud samples that have undergone data processing for the training of the classification network model. After training, this model can be used to accurately identify the vehicle model category of the input point cloud data. The classification network model calculates the probability value corresponding to the vehicle class of the point cloud data through feature extraction and classification algorithms, and then matches the subsequent drag coefficient model through the probability value. Point cloud data usually appears as an unordered and sparse three-dimensional point set, and the geometric features between different vehicle models vary greatly, and these differences are crucial for the prediction of the drag coefficient. By classifying the input point cloud data and classifying it into the most suitable vehicle model category, a corresponding dedicated prediction model can be selected for subsequent drag coefficient calculation, thereby improving the prediction accuracy.

[0033] S300. Construct a drag coefficient prediction model library, where the drag coefficient prediction model library includes several dedicated models corresponding one-to-one to the vehicle model categories and a generalization model.

[0034] Specifically, to ensure efficient prediction of the drag coefficient on different types of point cloud data, the present invention constructs a drag coefficient prediction model library, which has multiple drag coefficient models. In this application, the drag coefficient prediction model library includes several specific models and a generalization model. The specific model is a drag coefficient prediction model obtained by training and detecting for a specific type of vehicle model. The point cloud data of each vehicle model will independently train the specific model. By specifically training the samples of each category, each model can focus on specific types of data, and the specific model corresponds one-to-one with the vehicle type in the category label. The generalization model is a drag coefficient prediction model obtained by training and detecting for multiple vehicle models, which can process data from all vehicle models and learn a more extensive drag coefficient prediction method, thereby improving the robustness and generalization ability of the model. These models work together to select the most suitable model for drag coefficient prediction according to the classification result or generalization determination of the point cloud data.

[0035] S400. Determine whether the maximum probability value in the vehicle model classification label is higher than a preset threshold. If so, select the specific model corresponding to the vehicle model category for drag coefficient prediction. If not, select the generalization model for drag coefficient prediction.

[0036] Specifically, the classification network model will output a probability distribution, indicating the possibility that the input point cloud data belongs to each vehicle category; if the maximum probability value is greater than the preset probability threshold, the classification result is used as the type of the point cloud; if the maximum probability value is less than the preset probability threshold, it is considered that the characteristics of the point cloud data are too complex to be clearly classified, and thus it is assigned to the generalization model for processing.

[0037] The input point cloud data is classified by the classification network model. Once the vehicle type is identified, the system will select the corresponding specific prediction model for drag coefficient calculation, thereby minimizing the prediction error between different vehicle models. In addition, if some input data cannot be classified as a certain vehicle model with high confidence, it will be automatically assigned to the generalization model for prediction to ensure the applicability and robustness of the system.

[0038] It can be understood that in this application, a classification network model is first constructed, and the preprocessed point cloud data is input into the classification network model for vehicle type classification, and a vehicle type classification label is output. Then, a wind resistance coefficient prediction model library is constructed, which includes a number of dedicated models corresponding one-to-one to the vehicle type categories and a generalization model. Finally, a dedicated model or a generalization model is selected for matching through the probability value in the classification label, and then the wind resistance coefficient is predicted. The present invention avoids classifying vehicle types and matching suitable wind resistance coefficient prediction models. Compared with the traditional method that requires a large amount of calculation and has a large calculation error when predicting through a unified model, the prediction speed and the accuracy of the prediction result are greatly improved. The present invention can be widely applied to the evaluation and optimization of the wind resistance coefficient in vehicle design, shortening the development cycle and reducing the R & D cost. Moreover, compared with the existing wind resistance coefficient prediction methods, by using dedicated models trained for different vehicle types, this application can effectively solve the difference problem in wind resistance coefficient prediction for different vehicle types, thereby improving the accuracy and reliability of the prediction. This application fully combines the high accuracy of the dedicated model and the versatility of the generalization model, not only improving the accuracy of the prediction, but also enhancing the adaptability and robustness of the system when facing diverse vehicles. Overall, this application not only improves the accuracy of wind resistance coefficient prediction, but also improves the automation level of the system, and can better support the aerodynamic performance optimization in the whole vehicle design.

[0039] Please refer to Figure 2 , in some embodiments, the steps of the preprocessing specifically include: performing regularization processing on the point cloud data; performing enhancement processing on the regularized point cloud data; and performing maximum-minimum normalization processing on the enhanced point cloud data.

[0040] Specifically, the quality and robustness of the point cloud data are improved through the preprocessing. Point cloud regularization includes: normalizing the point cloud data through rotation, scaling, and displacement operations; data enhancement includes: random sampling, interpolation, and adding noise points to improve the adaptability of the model to different point cloud distributions; normalization includes: unifying the scale of the point cloud data to avoid the influence of scale differences on the classification model.

[0041] More specifically, the input point cloud data is a three-dimensional matrix. Assuming there are N points in the input point cloud, the feature of each point is a 3D vector (x, y, z); therefore, the input data can be represented as a tensor with a shape of (N, 3), where N is the number of points and 3 represents the coordinates (x, y, z) of each point. Then, data augmentation techniques are applied, including translation (moving the point cloud by a random factor), adding Gaussian noise (jitter), and randomly discarding some points in the point cloud. These augmentation operations help the model learn more spatial transformations and noise robustness. Finally, the code normalizes the point cloud data, mapping the coordinate values of each point cloud to the range of [0, 1] for subsequent deep learning training. Specifically, the maximum-minimum normalization method is used to normalize the 3D point cloud data. By traversing each three-dimensional point in the data, max and min are set for data normalization, and its calculation formula is as follows: , where is the value after maximum-minimum normalization processing, x is the value of the original data, is the minimum value in the original data, is the maximum value in the original data.

[0042] A series of standard processing transformations are performed on the image to transform it into a fixed standard form. The entire data processing process ensures the diversity and standardization of the point cloud data, providing more representative and generalizable data for model training.

[0043] Please refer to Figure 4 , in some embodiments, the step of constructing the classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting the vehicle type classification label specifically includes: constructing a PointNet classification network model, processing the point cloud data through several MLP layers with the same structure, generating a global feature vector through a max pooling layer, and then performing feature extraction and classification through a fully connected layer, and finally outputting the probability value corresponding to each vehicle type category through a softmax layer.

[0044] Specifically, this application processes the point cloud data, improving the robustness of the network model and ensuring the stability of the system under complex data distributions. Since the point cloud data has no clear structured order, the PointNet classification network model with several MLP layers with the same structure is used to process the unordered input point cloud data; during the training of the PointNet classification grid, a labeled point cloud data set is constructed, which includes three-dimensional point cloud samples of fastback vehicle type sample data, station wagon vehicle type sample data, and notchback vehicle type sample data; the supervised learning method is adopted, and the model is optimized through the cross-entropy loss function; through multiple rounds of iterative training, the classification model is enabled to have the generalization ability for new point cloud data.

[0045] Furthermore, the PointNet classification network model extracts the local features of each point through several multi-layer perceptron (MLP) networks with the same structure. The features of each point are processed by the MLP to obtain a new feature representation, and its calculation formula is as follows: , where is the result obtained after being processed by the multi-layer perceptron. MLP is a combination of a series of linear transformations and activation functions, and x is the input data.

[0046] After each point passes through the MLP, a high-dimensional feature vector is generated, which reflects the local geometric information of the point, that is, the local feature. In order to perform the same processing on each point, the PointNet classification network model adopts a strategy of sharing weights, that is, the same MLP structure is used for all points, avoiding the influence of the order of points on the network.

[0047] Next, through the max pooling operation, the local features of each point are aggregated into a global feature vector. This operation can retain the most significant feature information in the point cloud. Even if the order of the point cloud is disordered, it can ensure that the network outputs a stable global feature representation. The formula for the max pooling operation is: , where is the feature vector of each point, max performs the max pooling operation on each feature dimension, is the global feature vector.

[0048] Global feature is fed into the fully connected layer for further processing, and finally the probability distribution of each category is output. Specifically, the softmax function will convert the scores of all categories into a normalized probability distribution, ensuring that the probability of each category is within the range of [0, 1], and the sum of the probabilities of all categories is 1. The category corresponding to the maximum probability is the prediction result of the network. If the maximum probability does not exceed 0.8, it is considered that the data belongs to the generalization model data.

[0049] In some embodiments, before the step of constructing the PointNet classification network model, it further includes: constructing a spatial transformation network model to perform a global spatial transformation on the point cloud data.

[0050] Specifically, the spatial transformation network STN3d will perform a global spatial transformation on the entire point cloud data; the spatial transformation network gradually extracts the features of the point cloud through three 1D convolutional layers and performs max pooling to extract the global features. Then, three fully connected layers will further process the features and output a 9-dimensional transformation matrix (3x3), which is used to align the point cloud data so that the input point cloud data can be processed under the same reference framework.

[0051] In some embodiments, the vehicle type classification labels include: a fastback vehicle type and its corresponding probability value ; a wagon vehicle type and its corresponding probability value ; a notchback vehicle type and its corresponding probability value ; wherein , , , and .

[0052] It should be understood that, to achieve the above object, the present invention adopts the following solution: Different from using a single generalization model to predict the drag coefficient of multiple vehicle types, the present invention classifies the input point cloud data based on the external shape characteristics of the vehicle, and selects the most suitable dedicated prediction model for different body designs, thereby improving the prediction accuracy.

[0053] Specifically, the present invention divides automobiles into three major categories: fastback, wagon, and notchback according to the differences in the roof line and the tail structure. These three designs cover most of the passenger car structures on the market and can effectively represent the aerodynamic characteristics of various sedans and some sports vehicles.

[0054] The fastback vehicle type is characterized in that the roof line smoothly transitions from the front windshield to the rear of the vehicle, presenting an overall streamlined shape, such as some sports cars and sport sedans. For this type of vehicle, due to the relatively small tail inclination angle, the airflow separation occurs relatively late, and the drag coefficient is usually low.

[0055] The wagon corresponds to a station wagon or a hatchback. Its roof is relatively long and only forms an obvious downward trend when approaching the rear to provide a larger trunk space. Since the rear of the vehicle is relatively vertical, an airflow separation area is easily formed, so the aerodynamic characteristics are different from those of the fastback or notchback vehicle types.

[0056] The notchback refers to a traditional sedan with a clear three-box structure. The trunk part forms an obvious fold angle with the rear windshield, making the air flow pattern at the rear significantly different from that of the fastback and wagon, and usually forming a strong vortex structure in the trunk area.

[0057] Specifically, the point cloud data will be divided into three categories: fastback vehicle type, wagon vehicle type, and notchback vehicle type. The entire system has strong adaptability and supports multi-vehicle type expansion. The models of each category are trained according to the geometric characteristics of different vehicle types to better capture these characteristics and improve the prediction accuracy; the classification module will output a probability distribution indicating the possibility that the input point cloud data belongs to each category. The softmax function will convert the scores of all categories into a normalized probability distribution, ensuring that the probability of each category is within the range of [0, 1], and the sum of the probabilities of all categories is 1.

[0058] In some embodiments, the prediction threshold is and .

[0059] Specifically, the prediction threshold is set to and which can ensure the accuracy of the classification model, thereby reducing the computing resources and computing time of the wind resistance coefficient prediction model, and improving the accuracy of the wind resistance coefficient prediction. If the threshold is too low, it will instead affect the accuracy of the dedicated model. Exemplarily, in this application, the threshold is preset to 0.8, and the classification module will output a probability distribution indicating the possibility that the input point cloud data belongs to each category. If the maximum probability value is greater than 0.8, the classification result is used as the type of the point cloud; if the maximum probability value is less than 0.8, it is considered that the features of the point cloud data are too complex to be clearly classified, and thus it is assigned to the generalization model for processing.

[0060] It can be understood that by introducing a threshold of 0.8 in this application, a balance can be achieved between classification accuracy and generalization ability: when the classification result is relatively clear, the system will use the dedicated model to improve the prediction accuracy; when the classification confidence is low, the system will use the generalization model to ensure the stability of the prediction. This mechanism effectively avoids the influence of classification errors on the accuracy of the wind resistance coefficient prediction, and at the same time ensures that the system can handle a wider range of vehicle designs, making it more reliable and adaptable in practical applications. By classifying the point cloud data and matching the dedicated regression model, the wind resistance coefficient is predicted according to the characteristics of different vehicle models, solving the problems of poor generalization ability of traditional single models and insufficient prediction accuracy of generalization models.

[0061] Specifically, if the highest category probability of the point cloud data is greater than the threshold of 0.8, it is considered that the classification result is reliable, and the dedicated model corresponding to this category is directly used for wind resistance coefficient prediction. The dedicated model is trained for the geometric features of this category and can more accurately capture the aerodynamic characteristics of this vehicle model, thereby reducing the prediction error and improving the accuracy. However, in practical applications, the shapes of some cars may be between different categories. For example, for some models with special roof designs, their geometric features may simultaneously have the characteristics of a fastback and a wagonback. In this case, the classification network may be difficult to make a clear judgment, resulting in the probabilities of all categories being relatively close, and the confidence of no category exceeding 0.8. If the maximum category probability is lower than 0.8, it is considered that the features of the point cloud data are too complex or ambiguous to be accurately classified. At this time, the system will automatically assign it to the generalization model for processing. The generalization model has strong adaptability and can predict the point cloud data that has not been clearly classified, ensuring the robustness and applicability of the entire system.

[0062] Please refer to Figure 3, in some embodiments, the dedicated model is trained with point cloud data of corresponding vehicle model categories, and the generalization model is trained with point cloud data of all vehicle model categories.

[0063] Specifically, the sample data division is to randomly divide the data samples of each category, with 80% of the data for training and 20% for testing. The sample data of the notchback model, notchback model, and notchback model are used to train Model1, Model2, and Model3 respectively, to obtain three dedicated models. In addition to the dedicated models, a generalization model is also needed to process the point cloud data that cannot be clearly classified. When the classification module determines that the maximum classification probability of the point cloud data is less than 0.8, the data is sent to the generalization model. To train the generalization model Model4, the samples of all the data (notchback model sample data, notchback model sample data, notchback model sample data) are combined and randomly divided into 80% training data and 20% test data. In this way, Model4 can process data from all three vehicle models and learn a more extensive method for predicting the drag coefficient, thereby improving the robustness and generalization ability of the model. Finally, the model library will contain four models, Model1, Model2, Model3, and Model4, where three are dedicated models and one is a generalization model. These models work together to select the most suitable model for drag coefficient prediction according to the classification result or generalization determination of the point cloud data. And the system uses feature recognition and can dynamically expand the dedicated model library according to new vehicle model categories, with good adaptability and scalability.

[0064] Please refer to Figure 5 , in some embodiments, each model in the drag coefficient prediction model library is based on a dynamic graph edge convolution network. The KNN algorithm is used to calculate the distance between the input point cloud data and its nearest neighbor points and construct an adjacency graph to capture the local features of the point cloud data; the relative position vectors of each point are calculated through the EdgeConv operation of edge convolution, and the local features are transformed and updated in combination with a multi-layer perceptron (MLP) to extract the global features of the vehicle point cloud; the local features and the global features are aggregated, and the drag coefficient prediction value is output through a fully connected layer.

[0065] Specifically, after classification, the point cloud data will be further sent to the corresponding dedicated model in the model library for processing. These dedicated models use the method of graph convolutional neural network to extract the local and global features of the point cloud data. In the graph convolutional network, first, the KNN (K-Nearest Neighbors) algorithm needs to be used to find the k nearest neighbors for each point. This process first involves calculating the distance between points, usually using the Euclidean distance as the metric standard. Set the feature of each point as , where N is the number of points in the point cloud, and the features of the points can be their spatial coordinates or other additional descriptive information. The calculation formula for the Euclidean distance is as follows: , where is the point and The Euclidean distance between two points is often used for neighbor search in point cloud data. For each pair of points, the model calculates the inner product between points through matrix multiplication and combines the sum of squares of points to obtain the pairwise Euclidean distance. Specifically, assuming is the matrix of point cloud data, N is the number of points, and D is the dimension (three-dimensional coordinates) of each point, then the pairwise distance matrix D can be calculated as follows: , where is the Euclidean distance between the i-th and j-th points, and the and are the feature vectors of point i and point j, respectively.

[0066] For the calculated distance matrix D, we need to find the k nearest neighbors for each point: , where , is the index of the k nearest neighbors of point i, and the constraint ensures that itself will not be selected; this optimization objective ensures that the selected points are the ones with the smallest distance. This method calculates the pairwise distances of all points and selects the nearest neighbor points for each point from the distance matrix to obtain the indices of the neighbors .

[0067] Once the k nearest neighbors of each point are found, the next step is to calculate the relative position vectors and construct the graph feature matrix. The relative position vector of each point is the offset of point relative to point , and its calculation formula is: , where is the relative position vector of each point, is the i-th point, is the j-th point.

[0068] In this way, the relative position features between each point and its neighbors are extracted. The graph feature matrix G is formed by concatenating the features of each point with the relative position features of its k neighbors. Assuming the feature dimension of each point is D, the shape of the graph feature matrix of each point and its neighbors is , where includes the features of each point and the relative positions with neighbor points.

[0069] Graph Convolution (GCN) is a convolution operation based on graph-structured data. It extracts local and global features by aggregating the neighbor information of nodes. Among them, the EdgeConv operation is used to update the features of each point through the information of neighboring points. Each layer of graph convolution uses relative position vectors and original point features to calculate the convolution. Specifically, the EdgeConv operation can be expressed as: First, the nearest neighbors of each point are found through the above KNN algorithm and a local adjacency relationship is established; for each point i and its neighbor j, the relative position vector is calculated.

[0070] After that, a density weight mechanism is introduced, and combined with the features of the point itself , the edge features are constructed: , where is the density weight coefficient, is the Euclidean distance between point and point . Using the Euclidean distance calculated by KNN as the weight can reduce the waste of computing power. Introducing the density weight coefficient can enhance the model's attention to edge-sensitive parts such as rearview mirrors, door handles, and wheels: Feature transformation is performed using an MLP: where is the convolution feature of point i, is the neighbor set of point i, is a feature transformation function based on the features of point and its neighbor point , usually implemented by a multi-layer perceptron MLP. In each layer of convolution operation, through multiple EdgeConv operations, local geometric information and global information are continuously combined to extract more abstract and rich features. After the convolution layer, in order to enhance the expression of global information, the model adopts a pooling operation. Common pooling operations include Max Pooling and Average Pooling. Through pooling, the model can aggregate features at the global level and obtain a more comprehensive representation.

[0071] The features after Max Pooling and Average Pooling are sent to multiple fully connected layers for processing. By gradually reducing the feature dimension, the final output of the model is obtained. In a regression task, the final output is usually a scalar, such as the predicted value of the drag coefficient. Through the end-to-end training method, the model continuously optimizes the parameters during the training process to make the final prediction result as close as possible to the true value.

[0072] For the regression task, the final output of the network is the drag coefficient, which is a scalar. The network optimizes the output through a regression loss function, such as the mean squared error (MSE), and finally obtains the best prediction result through training.

[0073] Specifically, the regression loss function can be expressed as: , where is the predicted value of the i-th sample, is the true value, and N is the number of samples.

[0074] Generally speaking, after classification by the PointNet network, graph convolution operations are combined to predict the drag coefficient of point cloud data. The nearest neighbors of each point in the point cloud data are calculated through the KNN algorithm, and local and global features are extracted using graph convolution operations. The model gradually learns the spatial structure of the point cloud and finally outputs the prediction result through the regression layer. When dealing with point cloud data, this method can effectively capture the geometric features of the point cloud and provide accurate predictions for the regression task.

[0075] It can be understood that the prediction model is based on the dynamic graph edge convolution network DGCNN (EdgeConv). By using the KNN (K-Nearest Neighbor) algorithm to calculate the distance between the input point cloud data and its nearest neighbors, an adjacency graph is further constructed. Then, the information between adjacent points is propagated through graph edge convolution operations to update the feature representation of each point. The dynamic graph edge convolution layer can effectively learn the local geometric structure of the point cloud data, enabling the model to better capture the relationships between different points in the point cloud. Then, max pooling and average pooling operations are used to aggregate the features. These pooling operations can help the model extract global information and reduce the dimension of the data to a certain extent, improving the computational efficiency and stability. Max pooling usually retains the most significant features in the input data, while average pooling helps to capture the overall features. Finally, the pooled features are processed through a fully connected layer, and the final predicted value of the drag coefficient is generated. The fully connected layer learns the global features of the point cloud data, fuses multi-level information, and outputs a regression value, which is the predicted drag coefficient. The entire process is trained through backpropagation and gradient descent optimization algorithms to minimize the error between the predicted value and the true drag coefficient.

[0076] In some embodiments, the step of aggregating the local features and the global features includes: aggregating the local features and the global features through max pooling and / or average pooling.

[0077] Please refer to Figure 6, on the other hand, the present invention also provides an integrated prediction system for the vehicle point cloud drag coefficient based on feature recognition, including: a preprocessing module, configured to obtain the point cloud data of the vehicle model and preprocess the point cloud data; a classification module, configured to construct a classification network model, input the preprocessed point cloud data into the classification network model for vehicle type classification, and output a vehicle type classification label, wherein the vehicle type classification label includes a vehicle type category and a probability value corresponding to the vehicle type category; a prediction module, configured to construct a drag coefficient prediction model library, wherein the drag coefficient prediction model library includes a number of dedicated models corresponding one-to-one to the vehicle type categories and a generalization model; a matching module, configured to determine whether the maximum probability value in the vehicle type classification label is higher than a preset threshold. If so, select the dedicated model corresponding to the vehicle type category for drag coefficient prediction. If not, select the generalization model for drag coefficient prediction.

[0078] Among the 4000 data samples adopted by the present invention, there are 34% fastback vehicle type data samples, 36% wagonback vehicle type data samples and 30% notchback vehicle type data samples. During the construction of the model library, first, the data of each category are randomly divided into 80% training data and 20% test data. The sample data of fastback, wagonback and notchback are used to train Model1, Model2 and Model3 respectively. Then, all 4000 data samples are also randomly divided into 80% training data and 20% test data for training the generalization model Model 4. The model library includes a total of 3 dedicated models, namely Model 1, Model2, Model3 and a generalization model Model4.

[0079] Further, when inputting a point cloud data, the Spatial Transformer Network STN3d will perform a global spatial transformation on the entire point cloud data. The Spatial Transformer Network gradually extracts the features of the point cloud through three 1D convolutional layers and performs max pooling to extract global features. Then, three fully connected layers will further process the features and output a 9-dimensional transformation matrix (3x3), which is used to align the point cloud data so that the input point cloud data can be processed under the same reference framework. The point cloud data processed by the transformation matrix will enter the PointNet encoder. First, a convolutional layer (conv1, conv2, conv3) is used to extract local features, and then max pooling is used to extract global features, and finally a feature vector is output. Each time a point cloud data is input, the PointNet network will output a feature vector to reflect the probability distribution of the data in each category. The category corresponding to the maximum probability is the result of network classification. If the maximum probability does not exceed 0.8, the data will be classified into the generalization model category.

[0080] Furthermore, after the point cloud data is classified by the PointNet network, it will be input into the corresponding model in the model library for further processing. First, the model calculates the k (k = 20 in the figure) nearest neighbors of each point through the KNN algorithm. To find the nearest neighbors of each point, the model calculates the pairwise distances between points through transposition and matrix multiplication. Specifically, the Euclidean distance between each pair of points is calculated through the inner product and the sum of the squares of the points, and finally a pairwise distance matrix is obtained. Based on this distance matrix, the model selects the k points with the smallest distance and returns the indices of the k nearest neighbors of each point. Subsequently, using these found nearest neighbors, the model calculates the relative position vectors and constructs a graph feature matrix to form a local graph feature. Next, the graph convolution EdgeConv operation extracts local and global features through multiple convolutional layers, continuously learning the spatial structure of the point cloud. In each convolutional operation, the original point features and the relative position features of the neighboring points are jointly used to calculate the convolution. In this way, the network can effectively capture the geometric features of the point cloud data and gradually refine higher-order features in multiple convolutional layers. After the graph convolutional layer, the model uses max pooling and average pooling to globally aggregate the features of each layer, further enhancing the expression of global information. Subsequently, these features are processed through multiple fully connected layers, gradually reducing the feature dimension, and finally the prediction result is output through the regression layer. In the regression task, the final output is usually a scalar, such as the predicted value of the drag coefficient. The entire model is trained in an end-to-end manner to optimize the network parameters so that, given the point cloud data, the target value can be accurately predicted.

[0081] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0082] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0083] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in the relevant technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. An integrated prediction method for the vehicle point cloud drag coefficient based on feature recognition, characterized in that Including: Obtain the point cloud data of the vehicle model and preprocess the point cloud data; Construct a classification network model, input the preprocessed point cloud data into the classification network model for vehicle type classification, and output a vehicle type classification label, where the vehicle type classification label includes a vehicle type category and a probability value corresponding to the vehicle type category; Construct a wind resistance coefficient prediction model library, where the wind resistance coefficient prediction model library includes a number of dedicated models corresponding one-to-one to the vehicle type categories and a generalization model; Judge whether the maximum probability value in the vehicle type classification label is higher than a preset threshold. If so, select the dedicated model corresponding to the vehicle type category for wind resistance coefficient prediction. If not, select the generalization model for wind resistance coefficient prediction.

2. The integrated prediction method for the vehicle point cloud wind resistance coefficient based on feature recognition according to claim 1, wherein The steps of the preprocessing specifically include: Perform regularization processing on the point cloud data; Perform enhancement processing on the point cloud data after regularization processing; Perform maximum-minimum normalization processing on the point cloud data after enhancement processing.

3. The integrated prediction method for the vehicle point cloud aerodynamic drag coefficient based on feature recognition according to claim 1, wherein The steps of constructing the classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting a vehicle type classification label specifically include: Construct a PointNet classification network model, process the point cloud data through a number of MLP layers with the same structure, generate a global feature vector through a max pooling layer, and then perform feature extraction and classification through a fully connected layer, and finally output each vehicle type category and the corresponding probability value through a softmax layer.

4. The integrated prediction method for the vehicle point cloud wind resistance coefficient based on feature recognition according to claim 3, wherein Before the step of constructing the PointNet classification network model, it also includes: Construct a spatial transformation network model to perform global spatial transformation on the point cloud data.

5. The integrated prediction method of the vehicle point cloud wind resistance coefficient based on feature recognition according to claim 1, wherein The vehicle type classification label includes: Fastback models and their corresponding probability values ; Travel backpack models and their corresponding probability values ; Hatchback models and their corresponding probability values ; Among them, , , , and .

6. The integrated prediction method of the vehicle point cloud aerodynamic drag coefficient based on feature recognition according to claim 1, wherein The prediction threshold is , and .

7. The integrated prediction method of the vehicle point cloud aerodynamic drag coefficient based on feature recognition according to claim 1, characterized in that The dedicated model is trained through the point cloud data of the corresponding vehicle type category, and the generalization model is trained through the point cloud data of all vehicle type categories.

8. The integrated prediction method for the vehicle point cloud aerodynamic drag coefficient based on feature recognition according to claim 1, wherein Each model in the wind resistance coefficient prediction model library is based on a dynamic graph edge convolution network, uses the KNN algorithm to calculate the distance between the input point cloud data and its nearest neighbor points and constructs an adjacency graph to capture the local features of the point cloud data; Calculate the relative position vectors of each point through an EdgeConv operation of the edge convolution, and combine a multi-layer perceptron MLP to transform and update the local features to extract the global features of the vehicle point cloud; Converge the local features and the global features, and output the wind resistance coefficient prediction value through a fully connected layer.

9. The integrated prediction method of the vehicle point cloud wind resistance coefficient based on feature recognition according to claim 7, wherein The step of converging the local features and the global features includes: Converge the local features and the global features through max pooling and / or average pooling.

10. An integrated prediction system for the vehicle point cloud wind resistance coefficient based on feature recognition, characterized in that, Including: A preprocessing module for obtaining the point cloud data of the vehicle model and preprocessing the point cloud data; A classification module for constructing a classification network model, inputting the preprocessed point cloud data into the classification network model for vehicle type classification, and outputting a vehicle type classification label, where the vehicle type classification label includes a vehicle type category and a probability value corresponding to the vehicle type category; A prediction module for constructing a wind resistance coefficient prediction model library, wherein the wind resistance coefficient prediction model library includes a number of dedicated models corresponding one by one to the vehicle type categories and a generalization model; A matching module for determining whether the maximum probability value in the vehicle type classification label is higher than a preset threshold. If so, a dedicated model corresponding to the vehicle type category is selected for wind resistance coefficient prediction. If not, the generalization model is selected for wind resistance coefficient prediction.

Images