Millimeter wave radar multi-feature multi-core SVM-based human-vehicle distinguishing method

Abstract

The invention relates to a millimeter-wave radar multi-feature multi-kernel SVM-based human-vehicle distinguishing method, which comprises the following steps of: acquiring one-dimensional range profile data, RD diagram data and target point cloud data of a millimeter-wave radar, respectively extracting one-dimensional range profile features, RD diagram features and target point cloud features, and then selecting a proper kernel function according to feature types, the features are fused through a multi-core SVM for classification, so that accurate distinguishing of people and vehicles is realized; according to the method, through the advantages of feature fusion and the multi-core SVM, the limitation of a traditional single-core SVM in processing complex data is overcome, the classification precision of different targets such as pedestrians and vehicles is improved, and efficient and accurate target distinguishing can be carried out in a complex scene.

Description

Technical Field

[0001] The present invention relates to the field of automatic driving technology for automobiles, and in particular to a method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM. Background Art

[0002] The rapid development of autonomous driving technology is placing increasing demands on environmental perception capabilities. As one of the key sensors for environmental perception, millimeter-wave radar plays an irreplaceable and important role in target detection and collision warning due to its excellent all-day and all-weather operation capability, high range resolution, and high radial velocity measurement accuracy. By emitting electromagnetic waves in the millimeter-wave frequency band and receiving the echo signals reflected by the target, millimeter-wave radar can obtain key information such as the target's distance, speed, and angle in real time, providing key decision-making basis for autonomous driving systems. However, in complex and changing real-world traffic scenarios, existing millimeter-wave radar has significant limitations in target recognition capabilities. Simply detecting the presence of targets cannot meet the accuracy required for autonomous driving. Autonomous driving systems must be able to accurately distinguish different types of targets, such as people, vehicles, and obstacles, in order to make appropriate response decisions. Traditional radar signal processing methods usually use a single-class feature and a single kernel function SVM algorithm to classify targets. When faced with complex and diverse target types, the classification effect of this method is often unsatisfactory. For example, CN110427986B is a kernel support vector machine target classification method based on millimeter-wave radar point cloud features. By preprocessing the original radar point cloud data, the point cloud outside the radar detection area is eliminated, and then the target point cloud is clustered into one category through a clustering algorithm, thereby eliminating the noise point cloud. Combining the point cloud characteristics of the target, a feature vector consisting of 11 features is constructed, and a kernel support vector machine classifier is used for training and testing to achieve target classification. However, the extraction of a single feature only uses the point cloud. The distance or speed information in the data, but different types of targets may show similar single features in complex scenarios. For example, some small vehicles may be highly similar to pedestrians in distance features at specific angles, leading to misclassification. In addition, data of different feature types have different distribution characteristics and internal structures. When processing complex data, the SVM with a single kernel function can often only adapt to a specific data distribution pattern. It is difficult to effectively process data of multiple feature types such as one-dimensional range images, RD graph data, point cloud data, etc., resulting in limited target classification accuracy. Although CN110427986B eliminates some noisy point clouds through preprocessing and clustering algorithms, it still has the problem of insufficient adaptability to complex scenarios. Summary of the Invention

[0003] In view of this, the present invention provides a method for distinguishing people and vehicles based on millimeter-wave radar multi-feature multi-core SVM, which can solve the problem that the existing radar system is difficult to effectively distinguish different types of targets in complex scenarios.

[0004] To achieve the above objectives, the present invention provides a method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM, comprising the following steps: S1. Obtain millimeter-wave radar data to obtain one-dimensional range image, RD map and target point cloud data, and perform preprocessing; S2: Input the one-dimensional range image, RD map and target point cloud into the feature extractor and extract the corresponding features, respectively constructing the feature vectors of the one-dimensional range image, RD map and target point cloud. 、 、 , divide the feature set into training set and test set according to the ratio of 7:3; S3. Construct a set of kernel functions and use each kernel function to capture different features of the data in the training set; S301, using radial basis function kernel RBF function to process the feature vector of one-dimensional range image ; S302, using a polynomial kernel function to process the eigenvector of the RD graph ; S303, using the linear kernel function to process the feature vector of the target point cloud ; S4, perform multi-core SVM fusion and build a multi-core SVM classifier; S401, calculate the kernel matrix by combining the one-dimensional range image feature, RD map feature and target point cloud feature in the training set with their respective assigned kernel functions, and add the weight matrix to the kernel matrix. The weights of the initialization matrix are all set to , get the synthesized kernel matrix; S402, using the convex optimization problem of SVM to reversely adjust the weight matrix until the SVM iteration stop condition is met to obtain the optimal weight matrix, and calculate the fused kernel matrix, which is expressed as: ; ; in, 、 、 Represent the kernel functions corresponding to the one-dimensional range image feature, RD graph feature and target point cloud feature, respectively. and 、 and 、 and Respectively represent One-dimensional range image features, RD graph features and target point cloud features of frame data, 、 、 Represents kernel functions 、 、 The weight of S403: Construct a multi-core SVM classifier, process the training set and the test set respectively through the fused kernel matrix, input the training set to train the multi-core SVM classifier, and input the test set to obtain the recognition result of distinguishing between people and vehicles.

[0005] Preferably, obtaining millimeter wave radar data includes the following steps: S101, performing ADC sampling on the acquired millimeter-wave radar signal, and converting the continuous analog beat frequency signal into digital IQ data through a high-speed analog-to-digital converter; S102, one-dimensional FFT processing, mapping the time domain waveform of the digital IQ data to the frequency domain, separating different range units RB, achieving range resolution, and obtaining a one-dimensional range image; S103, two-dimensional FFT processing to achieve velocity resolution, distinguishing stationary and moving targets on the range-velocity two-dimensional plane, and obtaining an RD map; S104, incoherent accumulation, improves the signal-to-noise ratio (SNR) and suppresses random noise and scattered wave scintillation effects; S105, CFAR detection, extracting valid target points; S106, DOA estimation, performing FFT on the complex signals received by each channel of the array antenna after weighted delay according to different angles to obtain the incident direction corresponding to the spatial frequency domain.

[0006] Preferably, the characteristics of the one-dimensional range profile include peak energy, peak width, peak envelope area and peak standard deviation; The peak energy is the maximum value of the intensity of the radar echo signal at each position in the one-dimensional range image, and is expressed as: ; in, represents the peak energy, Represents the first Signal strength at each location; The peak width is expressed as peak energy The distance range occupied by 50% of the threshold in the one-dimensional distance image is expressed as: ; ; in, 、 Indicates that the energy in the one-dimensional range image is the peak energy half of the point; The peak envelope area is the area between the envelope of the peak signal and the range axis, indicating the distribution of the signal strength at the target. The expression is: ; in, represents the peak envelope area; The expression of the peak standard deviation is: ; in, represents the peak standard deviation, represents the average value of radar signal strength, Indicates the number of radar signal points; The features extracted from the one-dimensional range image are used to form a one-dimensional range image feature vector, which is expressed as: ; in, Represents the one-dimensional distance image feature vector.

[0007] Preferably, the features of the RD graph include HOG features, and extracting the HOG features includes the following steps: 1) Use the Sobel operator to calculate the magnitude and direction of the gradient of each pixel in the image. The expression is: ; ; ; ; in, represents the horizontal gradient, Represents RD graph pixel The pixel value of represents the vertical gradient, Represents pixel points Gradient size, Represents pixel points The gradient direction of 2) Divide the RD graph into Cells are counted, and the gradient directions of the pixels in each cell are divided into multiple bins according to the preset angle interval to obtain the gradient direction histogram, which is expressed as: ; in, Indicates the The gradient direction of each pixel, 、 They represent the gradient direction histogram 、 bin angle interval; 3) Place adjacent The cells are combined into a block, and the gradient histogram of each block is normalized, and the expression is: ; in, represents the normalized HOG features of the block, Indicates adjacent of the cell sum; 4) Construct the HOG feature vector, the expression is: ; in, Indicates the concatenation of the normalized HOG features of each block. Indicates the number of blocks.

[0008] Preferably, the expression of the target point cloud is: ; in, 、 、 、 Respectively represent The horizontal coordinate, vertical coordinate, velocity and energy value of each point cloud, Represents the total number of target point clouds, as the first feature of the target point cloud feature vector ; A rectangular box is proposed to represent the shape of the object, and all target point clouds are The maximum and minimum values on the coordinates, that is, the length of the rectangular box Defined as the second feature of the target point cloud feature vector , the expression is: ; in, Represents all target point clouds Coordinate array; All target point clouds The maximum and minimum values on the coordinates, that is, the width of the rectangular box Defined as the third feature of the target point cloud feature vector , the expression is: ; in, Represents all target point clouds Coordinate array; The area of the rectangular box Defined as the fourth feature of the target point cloud feature vector , the expression is: ; The density of the target point cloud Defined as the fifth feature of the target point cloud feature vector , the expression is: ; The average radial velocity of the target point cloud Defined as the sixth object feature of the target point cloud feature vector , the expression is: ; Set the target point cloud velocity range Defined as the seventh object feature of the target point cloud feature vector , the expression is: ; in, An array representing the radial velocities of all target point clouds; The peak value of the echo intensity of the target point cloud Defined as the eighth object feature of the target point cloud feature vector , the expression is: ; in, An array representing the echo intensities of all target point clouds; The 2D area of the geometric convex hull of the target point cloud The ninth object feature defined as the target point cloud feature vector , the expression is: ; ; in, Represents the geometric convex hull area of the target point cloud, Represents the number of points in the geometric convex hull arranged in counterclockwise order, let Make the geometric convex hull closed end to end; The 2D perimeter of the geometric convex hull of the target point cloud Defined as the tenth object feature of the target point cloud feature vector , the expression is: ; The isoperimetric ratio of the geometric convex hull of the target point cloud The eleventh object feature defined as the target point cloud feature vector , the expression is: ; Isoperimetric The closer it is to 1, the closer the geometric convex hull of the target point cloud is to a circle; Feature vector of target point cloud The expression is: .

[0009] Preferably, the expression of the radial basis function kernel RBF function is: ; in, 、 Represents the one-dimensional distance image feature sample of two inputs, Represents the parameters of the radial basis function kernel RBF, which controls the similarity calculation of the data; The expression of the polynomial kernel function is: ; in, 、 Represents the input RD graph feature sample, represents a constant term, represents the order of the polynomial; The expression of the linear kernel function is: ; in, 、 Represents the input target point cloud feature sample.

[0010] Compared with the prior art, the present invention has the following beneficial effects: The present invention integrates multiple features such as one-dimensional range images, RD maps and point clouds, and combines them with a multi-core SVM algorithm. This method can effectively overcome the limitations of traditional single-core SVM in processing complex data, and greatly improves the classification accuracy of different types of targets (such as pedestrians and vehicles); the multi-core SVM dynamically adjusts the weights of multiple kernel functions and performs adaptive optimization according to the different characteristic characteristics of the target, thereby more accurately capturing the multi-dimensional information of the target, including size, shape, motion state and reflection characteristics. This fusion and optimization strategy not only enhances the expressive ability of features, but also enables the system to exhibit stronger adaptability and robustness when facing complex and changeable traffic scenes; the method provided by the present invention provides the autonomous driving system with a more reliable, efficient and accurate human-vehicle differentiation capability, greatly improving the safety of autonomous driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] FIG1 is a flow chart of millimeter wave radar signal processing according to the present invention; Figure 2 Schematic diagram of the multi-feature multi-core SVM method for distinguishing people and vehicles in the present invention; Figure 3 Schematic diagram of the multi-feature multi-core SVM hot car differentiation strategy and classification process of the present invention. DETAILED DESCRIPTION

[0012] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0013] To achieve accurate human-vehicle differentiation in complex scenarios, this embodiment provides a human-vehicle differentiation method based on millimeter-wave radar multi-feature multi-core SVM, including the following steps: S1. Obtain millimeter-wave radar data to obtain one-dimensional range image, RD map and target point cloud data, and perform preprocessing; S101, performing ADC sampling on the acquired millimeter-wave radar signal, and converting the continuous analog beat frequency signal into digital IQ data through a high-speed analog-to-digital converter, laying the foundation for subsequent digital processing; S102, one-dimensional FFT processing (Range FFT), mapping the time domain waveform of the digital IQ data to the frequency domain, separating different range units RB, achieving range resolution, and obtaining a one-dimensional range image; S103, two-dimensional FFT processing (Doppler FFT) to achieve velocity resolution, distinguish between stationary and moving targets on the range-velocity two-dimensional plane, and obtain an RD map; S104, Magnitude Integration, improves the signal-to-noise ratio (SNR), suppresses random noise and scattered wave flicker effects, simplifies subsequent detection algorithms, and is insensitive to phase; S105, CFAR detection (Constant False Alarm Rate), maintains a constant false alarm rate in complex and variable noise environments (ground clutter, rain and snow scattering) and extracts valid target points; S106, DOA estimation (Direction-of-Arriva), the complex signal received by each channel of the array antenna is weighted and delayed according to different angles, and then FFT is performed to obtain the corresponding incident direction in the spatial frequency domain.

[0014] like Figure 1As shown in the figure, in the signal processing process, the output of the one-dimensional FFT is extracted to obtain the one-dimensional range image; the output of the two-dimensional FFT is extracted to obtain the RD map, as well as the output target point cloud data of the entire process. The one-dimensional range image features, RD map features and point cloud features have different expression forms and data distribution characteristics, which play an important role in distinguishing people from vehicles. Therefore, these three are used as the input of the multi-feature multi-core SVM human-vehicle distinction algorithm.

[0015] S2: Input the one-dimensional range image, RD map and target point cloud into the feature extractor and extract the corresponding features, respectively constructing the feature vectors of the one-dimensional range image, RD map and target point cloud. 、 、 , divide the feature set into training set and test set according to the ratio of 7:3; In a millimeter-wave radar's one-dimensional range profile, the target's echo signal typically manifests as a series of peaks and varying characteristics. These peaks and their corresponding characteristics can reflect the target's properties, such as size, position, and shape. The characteristics of a one-dimensional range profile include peak energy, peak width, peak envelope area, and peak standard deviation. Peak energy is the maximum value of the intensity of the radar echo signal at each position in the one-dimensional range image, and its expression is: ; in, represents the peak energy, Represents the first Signal strength at each location; Peak width refers to the distance range occupied when the signal drops from the peak intensity to a certain threshold (usually a certain percentage of the peak value). The commonly used threshold ratio is 50% of the peak value, which is called the "half power point" or "half maximum width". In this embodiment, the peak energy The distance range occupied by 50% of the threshold in the one-dimensional distance image is expressed as: ; ; in, 、 Indicates that the energy in the one-dimensional range image is the peak energy half of the point; The peak envelope area is the area between the envelope of the peak signal and the range axis. It is an overall feature extracted from the one-dimensional range image and represents the distribution of signal strength at the target. It is usually obtained by numerical integration. Since the one-dimensional range image is discrete, it can be approximated by summation. The expression is: ; in, Indicates the peak envelope area, and the summation interval is the distance between the starting point and the end point of the envelope line. and are the starting and ending positions, and The position where the signal strength first drops to 50% of the peak value (i.e., and This feature helps describe the target's reflection intensity distribution and further infer the target's size and reflection properties; The peak standard deviation describes the degree of dispersion of the signal near the peak value and can be used to measure the concentration of the target reflection characteristics. If the standard deviation is large, it means that the reflection intensity of the target varies greatly at different locations, which usually means that the target has a more complex shape. The standard deviation indicates the degree of fluctuation of the signal intensity relative to its average value. The expression for the signal standard deviation in the one-dimensional range image is: ; in, represents the peak standard deviation, represents the average value of radar signal strength, Represents the number of radar signal points. The standard deviation can reflect the consistency of the target reflection characteristics. For static targets, the standard deviation is small, while for dynamic targets (such as pedestrians or wheels), the standard deviation is large. The features extracted from the one-dimensional range image are used to form a one-dimensional range image feature vector, which is expressed as: ; in, Represents a one-dimensional distance image feature vector. These features are important for target recognition and classification and can help distinguish different types of targets, such as pedestrians and vehicles. In practical applications, these features can be combined to improve the accuracy of target detection and classification.

[0016] The features of the RD image include HOG features. The basic idea of HOG features is to describe the shape characteristics of the target by calculating the gradient direction histogram in the local area of the image. HOG features can effectively capture the edge information in the image, which is particularly important for target recognition and classification. In the RD image, HOG feature extraction can effectively capture the edge and direction characteristics of the target. Since the RD image is a two-dimensional image (a function of distance and Doppler frequency), HOG features can be used to describe the local gradient and direction information in these images. Extracting the HOG features includes the following steps: 1) Use the Sobel operator to calculate the magnitude and direction of the gradient of each pixel in the image. The expression is: ; ; ; ; in, represents the horizontal gradient, Represents RD graph pixel The pixel value of represents the vertical gradient, Represents pixel points Gradient size, Represents pixel points The gradient direction of 2) Divide the RD graph into (For example ) cells, count the gradient directions of the pixels in each cell respectively, divide the gradient directions into multiple bins according to the preset angle interval (usually 0-180 degrees or 0-360 degrees), and obtain the gradient direction histogram, which is expressed as: ; in, Indicates the The gradient direction of each pixel, 、 They represent the gradient direction histogram 、 bin angle interval; 3) Place adjacent The cells are combined into a block, and the gradient histogram of each block is normalized, and the expression is: ; in, represents the normalized HOG features of the block, Indicates adjacent of the cell sum; 4) Construct HOG feature vector for target classification, the expression is: ; in, Indicates the concatenation of the normalized HOG features of each block. Indicates the number of blocks.

[0017] In the RD map of millimeter-wave radar, the shape, edge and structural features of the target are crucial for target classification. The HOG feature can capture the edge and directional features of the target by extracting the gradient information of the target. In this way, different types of targets, such as vehicles and pedestrians, can be effectively distinguished, especially in motion detection in complex environments. The frequency distribution of pedestrians in the RD map is relatively complex, and the HOG feature can effectively capture this complex edge and directional information. Vehicles usually show relatively regular shapes and a large speed range in the RD map, and the HOG feature can help distinguish these regular edge structures. HOG feature extraction is to extract the shape and edge information of the image by calculating the gradient information of the image and converting it into a directional histogram. For the RD map of millimeter-wave radar, the HOG feature can effectively capture the shape characteristics, edge information and motion direction of the target. These features are very useful for target classification tasks (such as pedestrian and vehicle recognition). After target detection, the reflection points corresponding to the target of interest are obtained, which is called the target point cloud. However, further processing is required to obtain the type information of each target. The 11 point cloud features extracted from the shape, speed and intensity of the target point cloud are used to describe the target type, and the target is classified based on these features. The expression of the target point cloud is: ; in, 、 、 、 Respectively represent The horizontal coordinate, vertical coordinate, velocity and energy value of each point cloud, Indicates the total number of target point clouds. The number of object point clouds is determined by the shape, size, and surface material of the object. Pedestrians and vehicles vary greatly in shape, size, and surface material. The total number of target point clouds is used as the first feature of the target point cloud feature vector. ; A rectangular box is proposed to represent the shape of the object, and all target point clouds are The maximum and minimum values on the coordinates, that is, the length of the rectangular box Defined as the second feature of the target point cloud feature vector , the expression is: ; in, Represents all target point clouds Coordinate array; All target point clouds The maximum and minimum values on the coordinates, that is, the width of the rectangular box Defined as the third feature of the target point cloud feature vector , the expression is: ; in, Represents all target point clouds Coordinate array; The area of the rectangular box Defined as the fourth feature of the target point cloud feature vector , the expression is: ; Compared with pedestrians, the appearance and materials of vehicles are more uneven, which makes the distribution of vehicle point clouds in the rectangular box more sparse. Therefore, the sparsity of point clouds in the rectangular box is a good feature for identifying targets. Defined as the fifth feature of the target point cloud feature vector , the expression is: ; When an object moves, the radar sensor can measure the radial velocity of the object through the Doppler frequency shift and calculate the average radial velocity of the target point cloud. Defined as the sixth object feature of the target point cloud feature vector , the expression is: ; Different objects have different motion patterns, resulting in different velocity distributions of point clouds: When a pedestrian walks, each part of the body has complex movements. The pedestrian's head and torso translate along the walking path at the pedestrian's speed, and the arms and legs swing slightly back and forth at a speed different from the head and torso. The velocity components generated by the arms and legs cause the velocity distribution of the pedestrian point cloud to be in a wide range. In contrast, the velocity distribution of the vehicle point cloud is relatively narrow because the vehicle's motion consists only of the translation of its rigid body. The range of the target point cloud velocity is set to Defined as the seventh object feature of the target point cloud feature vector , the expression is: ; in, An array representing the radial velocities of all target point clouds; The reflection intensity of the target point cloud is determined by the amplitude of the reflected radar echo signal. RCS is usually used to characterize the echo intensity generated by an object under radar wave irradiation. The RCS value is affected by the target size, material, radar signal incident angle and frequency. The RCS of a pedestrian is approximately , the RCS of a vehicle is usually and The reflection intensity of the signals reflected by pedestrians and vehicles for the same incident wave is different; the peak value of the echo intensity of the target point cloud is Defined as the eighth object feature of the target point cloud feature vector , the expression is: ; in, An array representing the echo intensities of all target point clouds.

[0018] In the human-vehicle distinction of millimeter-wave radar point clouds, the geometric convex hull is mainly used to characterize the target point cloud in space (usually taking ) on the shape and distribution characteristics, the geometric feature extraction of the point cloud convex hull has the following functions: 1. Shape approximation: The convex hull connects the outermost points of the point cloud into a convex polygon, which approximately captures the geometric shape of the target. Human point clouds are often linearly distributed, slender and irregular; vehicle point clouds are densely distributed and have regular shapes (rectangular or elliptical). 2. Unaffected by concavities and noise: The convex hull relies only on extreme points, ignoring internal concavities and scattered outliers, and has good robustness; 3. Feature quantification: The convex hull can be used to calculate numerical features such as area, perimeter, and isoperimetry, providing the discriminant dimension for the subsequent kernel SVM.

[0019] The 2D area of the geometric convex hull of the target point cloud The ninth object feature defined as the target point cloud feature vector , the expression is: ; ; in, Represents the geometric convex hull area of the target point cloud, Represents the number of points in the geometric convex hull arranged in counterclockwise order, let Make the geometric convex hull closed end to end; The 2D perimeter of the geometric convex hull of the target point cloud Defined as the tenth object feature of the target point cloud feature vector , the expression is: ; The isoperimetric ratio of the geometric convex hull of the target point cloud The eleventh object feature defined as the target point cloud feature vector , the expression is: ; Isoperimetric Used to measure the compactness of the shape. The closer it is to 1, the closer the geometric convex hull of the target point cloud is to a circle. Feature vector of target point cloud The expression is: ; S3. Construct a set of kernel functions and use each kernel function to capture different features of the data in the training set.

[0020] In the process of millimeter-wave radar data processing, range profile features, RD graph features, and point cloud features have different representations and data distribution characteristics. Therefore, when applying support vector machines (SVM) for classification, the appropriate kernel function should be selected based on the respective feature types and data characteristics. S301, the range image feature reflects the distance distribution between the object and the radar by converting the distance information of the radar echo signal into an image form. This type of feature can be regarded as a two-dimensional image data, which usually contains a large number of pixel values. Therefore, the radial basis function kernel RBF function is used to process the feature vector of the one-dimensional range image. , each pixel value of the range image feature represents the position or distance information of an object in space, and has a relatively complex nonlinear distribution. The RBF kernel can effectively handle this nonlinear relationship and is particularly suitable for capturing local patterns and complex structures in the data. In addition, the RBF kernel can maintain good classification ability in high-dimensional space and is suitable for classification tasks of high-dimensional and sparse data. The expression of the radial basis function kernel RBF function is: ; in, 、 Represents the one-dimensional distance image feature sample of two inputs, Represents the parameters of the radial basis function kernel RBF, which controls the similarity calculation of the data; S302, RD map features combine distance and Doppler frequency information to reflect the target's motion characteristics; RD map is usually a two-dimensional matrix that contains the target's distance and relative speed information. This type of feature has a strong spatiotemporal correlation and can reflect the object's motion state in the radar field of view. Therefore, a polynomial kernel function is used to process the eigenvectors of the RD map. , RD graph features contain the target's speed and distance information, which may show a certain polynomial relationship; the polynomial kernel can capture the interaction effect between different features and is particularly suitable for processing data with obvious polynomial relationships. It has a good classification effect on the target's motion characteristics (such as the speed difference between a moving vehicle and a pedestrian). The expression of the polynomial kernel function is: ; in, 、 Represents the input RD graph feature sample, represents a constant term, Represents the order of the polynomial, and appropriate parameters can usually be selected through cross-validation; S303. Point cloud features usually represent the shape and structure of the target in three-dimensional space. The distribution of point clouds reflects the geometric shape of the object. Each point cloud can be regarded as a collection of several points. The shape, size and spatial distribution information of the target can be obtained through these points. Therefore, the linear kernel function is used to process the feature vector of the target point cloud. Point cloud features usually have good geometric structure and high separability. In many point cloud classification problems, the shape and spatial distribution of point clouds can often be well distinguished by linear methods, especially in low-dimensional feature spaces. For point cloud data, linear kernels can effectively capture the overall shape and structure of the target and avoid overly complex nonlinear transformations. The expression of the linear kernel function is: ; in, 、 Represents the input target point cloud feature samples, and their inner product is calculated; By selecting an appropriate kernel SVM, the accuracy and robustness of millimeter-wave radar data classification can be improved, helping to achieve more accurate human-vehicle differentiation tasks.

[0021] S4, perform multi-core SVM fusion and build a multi-core SVM classifier; The basic idea of multi-core SVM is to combine multiple kernel functions, use each kernel function to capture different features of the data, and improve the classification accuracy by weighted combination of the results of these kernels, which can be expressed as: ; in, and is Two input feature samples, Indicates the kernel function; It is The weight of a kernel function reflects the importance of the kernel in the final classification, and ; in, is the number of kernel functions; By optimizing the weight ,Multi-core SVM can dynamically select the kernel function combination that best suits the problem, and perform classification by finding the best hyperplane in multiple feature spaces; S401, calculate the kernel matrix by combining the one-dimensional range image feature, RD map feature and target point cloud feature in the training set with their respective assigned kernel functions, and add the weight matrix to the kernel matrix. The weights of the initialization matrix are all set to , get the synthesized kernel matrix; S402, using the convex optimization problem of SVM to reversely adjust the weight matrix until the SVM iteration stop condition is met to obtain the optimal weight matrix, and calculate the fused kernel matrix, which is expressed as: ; ; in, 、 、 Represent the kernel functions corresponding to the one-dimensional range image feature, RD graph feature and target point cloud feature, respectively. and 、 and 、 and Respectively represent One-dimensional range image features, RD graph features and target point cloud features of frame data, 、 、 Represents kernel functions 、 、 The weight of S403: Construct a multi-core SVM classifier, process the training set and the test set respectively through the fused kernel matrix, input the training set to train the multi-core SVM classifier, and input the test set to obtain the recognition result of distinguishing between people and vehicles.

[0022] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for distinguishing people from vehicles based on millimeter-wave radar multi-feature multi-core SVM, characterized by: The following steps are involved: S1. Obtain millimeter-wave radar data to obtain one-dimensional range image, RD map and target point cloud data, and perform preprocessing; S2: Input the one-dimensional range image, RD map and target point cloud into the feature extractor and extract the corresponding features, respectively constructing the feature vectors of the one-dimensional range image, RD map and target point cloud. 、 、 , divide the feature set into training set and test set according to the ratio of 7:3; S3. Construct a set of kernel functions and use each kernel function to capture different features of the data in the training set; S301, using radial basis function kernel RBF function to process the feature vector of one-dimensional range image ; S302, using a polynomial kernel function to process the eigenvector of the RD graph ; S303, using the linear kernel function to process the feature vector of the target point cloud ; S4, perform multi-core SVM fusion and build a multi-core SVM classifier; S401, calculate the kernel matrix by combining the one-dimensional range image feature, RD map feature and target point cloud feature in the training set with their respective assigned kernel functions, and add the weight matrix to the kernel matrix. The weights of the initialization matrix are all set to , get the synthesized kernel matrix; S402, using the convex optimization problem of SVM to reversely adjust the weight matrix until the SVM iteration stop condition is met to obtain the optimal weight matrix, and calculate the fused kernel matrix, which is expressed as: ； ； in, 、 、 Represent the kernel functions corresponding to the one-dimensional range image feature, RD graph feature and target point cloud feature, respectively. and 、 and 、 and Respectively represent One-dimensional range image features, RD graph features and target point cloud features of frame data, 、 、 Represents kernel functions 、 、 The weight of S403: Construct a multi-core SVM classifier, process the training set and the test set respectively through the fused kernel matrix, input the training set to train the multi-core SVM classifier, and input the test set to obtain the recognition result of distinguishing between people and vehicles.

2. The method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM according to claim 1 is characterized in that: Acquiring millimeter-wave radar data involves the following steps: S101, performing ADC sampling on the acquired millimeter-wave radar signal, and converting the continuous analog beat frequency signal into digital IQ data through a high-speed analog-to-digital converter; S102, one-dimensional FFT processing, mapping the time domain waveform of the digital IQ data to the frequency domain, separating different range units RB, achieving range resolution, and obtaining a one-dimensional range image; S103, two-dimensional FFT processing to achieve velocity resolution, distinguishing stationary and moving targets on the range-velocity two-dimensional plane, and obtaining an RD map; S104, incoherent accumulation, improves the signal-to-noise ratio (SNR) and suppresses random noise and scattered wave scintillation effects; S105, CFAR detection, extracting valid target points; S106, DOA estimation, performing FFT on the complex signals received by each channel of the array antenna after weighted delay according to different angles to obtain the incident direction corresponding to the spatial frequency domain.

3. The method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM according to claim 1 is characterized in that: The characteristics of the one-dimensional range image include peak energy, peak width, peak envelope area and peak standard deviation; The peak energy is the maximum value of the intensity of the radar echo signal at each position in the one-dimensional range image, and is expressed as: ； in, represents the peak energy, Represents the first Signal strength at each location; The peak width is expressed as peak energy The distance range occupied by 50% of the threshold in the one-dimensional distance image is expressed as: ； ； in, 、 Indicates that the energy in the one-dimensional range image is the peak energy half of the point; The peak envelope area is the area between the envelope of the peak signal and the range axis, indicating the distribution of the signal strength at the target. The expression is: ； in, represents the peak envelope area; The expression of the peak standard deviation is: ； in, represents the peak standard deviation, represents the average value of radar signal strength, Indicates the number of radar signal points; The features extracted from the one-dimensional range image are used to form a one-dimensional range image feature vector, which is expressed as: ； in, Represents the one-dimensional distance image feature vector.

4. The method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM according to claim 3 is characterized in that: The features of the RD graph include HOG features, and extracting the HOG features includes the following steps: 1) Use the Sobel operator to calculate the magnitude and direction of the gradient of each pixel in the image. The expression is: ； ； ； ； in, represents the horizontal gradient, Represents RD graph pixel The pixel value of represents the vertical gradient, Represents pixel points Gradient size, Represents pixel points The gradient direction of 2) Divide the RD graph into Cells are counted, and the gradient directions of the pixels in each cell are divided into multiple bins according to the preset angle interval to obtain the gradient direction histogram, which is expressed as: ； in, Indicates the The gradient direction of each pixel, 、 They represent the gradient direction histogram 、 bin angle interval; 3) Place adjacent The cells are combined into a block, and the gradient histogram of each block is normalized, and the expression is: ； in, represents the normalized HOG features of the block, Indicates adjacent of the cell sum; 4) Construct the HOG feature vector, the expression is: ； in, Indicates the concatenation of the normalized HOG features of each block. Indicates the number of blocks.

5. The method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM according to claim 4 is characterized in that: The expression of the target point cloud is: ； in, 、 、 、 Respectively represent The horizontal coordinate, vertical coordinate, velocity and energy value of each point cloud, Represents the total number of target point clouds, as the first feature of the target point cloud feature vector ; A rectangular box is proposed to represent the shape of the object, and all target point clouds are The maximum and minimum values on the coordinates, that is, the length of the rectangular box Defined as the second feature of the target point cloud feature vector , the expression is: ； in, Represents all target point clouds Coordinate array; All target point clouds The maximum and minimum values on the coordinates, that is, the width of the rectangular box Defined as the third feature of the target point cloud feature vector , the expression is: ； in, Represents all target point clouds Coordinate array; The area of the rectangular box Defined as the fourth feature of the target point cloud feature vector , the expression is: ； The density of the target point cloud Defined as the fifth feature of the target point cloud feature vector , the expression is: ； The average radial velocity of the target point cloud Defined as the sixth object feature of the target point cloud feature vector , the expression is: ； Set the target point cloud velocity range Defined as the seventh object feature of the target point cloud feature vector , the expression is: ； in, An array representing the radial velocities of all target point clouds; The peak value of the echo intensity of the target point cloud Defined as the eighth object feature of the target point cloud feature vector , the expression is: ； in, An array representing the echo intensities of all target point clouds; The 2D area of the geometric convex hull of the target point cloud The ninth object feature defined as the target point cloud feature vector , the expression is: ； ； in, Represents the geometric convex hull area of the target point cloud, Represents the number of points in the geometric convex hull arranged in counterclockwise order, let Make the geometric convex hull closed end to end; The 2D perimeter of the geometric convex hull of the target point cloud Defined as the tenth object feature of the target point cloud feature vector , the expression is: ； The isoperimetric ratio of the geometric convex hull of the target point cloud The eleventh object feature defined as the target point cloud feature vector , the expression is: ； Isoperimetric The closer it is to 1, the closer the geometric convex hull of the target point cloud is to a circle; Feature vector of target point cloud The expression is: 。 6. The method for distinguishing between people and vehicles based on millimeter-wave radar multi-feature multi-core SVM according to claim 1 is characterized in that: The expression of the radial basis function kernel RBF function is: ； in, 、 Represents the one-dimensional distance image feature sample of two inputs, Represents the parameters of the radial basis function kernel RBF, which controls the similarity calculation of the data; The expression of the polynomial kernel function is: ； in, 、 Represents the input RD graph feature sample, represents a constant term, represents the order of the polynomial; The expression of the linear kernel function is: ； in, 、 Represents the input target point cloud feature sample.

Images