Road scene target recognition method based on lidar point cloud
By using quadratic polynomial model and random sampling consistency to determine the ground point cloud in pavement scenarios, combined with statistical filtering and bilateral filtering for denoising, combined with Meanshift clustering and support vector machine models for target recognition, the problem of difficult to balance the efficiency and accuracy of target recognition in the existing technology is solved, and a more efficient and accurate target recognition effect is achieved.
Patent Information
- Application Number
- CN202210203259.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-02
- Publication Date
- 2025-03-25
- Estimated Expiration
- 2042-03-02
AI Technical Summary
The prior art is difficult to take into account both efficiency and accuracy in lidar point cloud target recognition in pavement scenarios, and it is difficult to accurately segment the target.
The ground point cloud is determined based on the quadratic polynomial model and random sampling consistency, and the point cloud denoising is combined with statistical filtering and bilateral filtering. The target point cloud is segmented using Meanshift clustering, and the target classification is performed through the support vector machine model.
Effectively filter out ground point clouds, remove multi-scale noise, improve the accuracy and efficiency of target recognition, and obtain a more accurate classification model.
Smart Images
Figure CN114612795B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target classification, and in particular to a road scene target recognition method based on laser radar point cloud. Background Art
[0002] LiDAR is widely used in intelligent transportation fields such as intelligent traffic monitoring, vehicle-road collaboration and autonomous driving due to its strong anti-interference ability, high spatial resolution and strong environmental adaptability. However, whether it is in pedestrian and vehicle safety warning, pedestrian and vehicle flow statistics, comprehensive traffic monitoring and command, or in the application of autonomous driving, it is very dependent on the precise imaging and target recognition of point cloud targets. At present, the LiDAR point cloud target recognition algorithms for road scenes cannot take into account both efficiency and accuracy at the same time, or it is difficult to accurately segment the target. Summary of the invention
[0003] The purpose of the present invention is to provide a road scene target recognition method based on laser radar point cloud to address the deficiencies of the prior art.
[0004] The technical solution to achieve the purpose of the present invention is: a road scene target recognition method based on laser radar point cloud, comprising the following steps:
[0005] Step 1: Use LiDAR to collect road scene point cloud data;
[0006] Step 2: Determine the ground point cloud according to the quadratic polynomial model and random sampling consistency, and delete the ground point cloud data from the point cloud data;
[0007] Step 3: Filter the remaining point cloud in step 2, including removing isolated outliers using statistical filtering and correcting the position of the target surface point cloud based on bilateral filtering of the point cloud normal vector and Euclidean distance;
[0008] Step 4: Use Meanshift clustering to segment the 3D point cloud obtained in step 3 to obtain the target point cloud set;
[0009] Step 5: Describe the target point cloud set obtained in step 4 and extract structural feature information;
[0010] Step 6: Input the extracted structural feature information into the trained support vector machine model to classify the target point cloud.
[0011] Preferably, the road scene point cloud data includes spatial coordinate information and reflectivity information of the target.
[0012] Preferably, the specific method for determining the ground point cloud based on the quadratic polynomial model and random sampling consistency is:
[0013] Step 2.1: Set the initial ground height threshold d, take z = 0 as the reference of the ground plane, randomly select 6 data points from z∈[-d, d], and fit the initial ground model according to the quadratic polynomial fitting road surface expression as the current optimal model;
[0014] Step 2.2: Randomly select 6 different data points from z∈[-d,d] and fit the new ground model according to step 2.1;
[0015] Step 2.3: Compare the number of inliers between the current optimal model and the new ground model, and update the model with more inliers as the current optimal model;
[0016] Step 2.4: Repeat steps 2.2-2.3 until the maximum number of iterations is reached, the process ends, and the internal points of the current optimal model are recorded as the ground point cloud.
[0017] Preferably, the method for determining the interior points of the model is:
[0018] Traverse the data points in the point cloud and determine whether they are internal points. The specific determination method is: according to the quadratic polynomial fitting road surface expression, determine the corresponding z value of the point (x, y) on the current fitting surface, expressed as z g , the z-axis coordinate value corresponding to the actual coordinates of the point (x, y) is z t , when |z t -z g |≤e, the point is judged as an interior point, where e is the distance error threshold.
[0019] Preferably, the maximum number of iterations is:
[0020]
[0021] Where u is the ratio of the point cloud of ground points to the entire point cloud, N is the number of points in the model, and P is the confidence of the model.
[0022] Preferably, the specific method for filtering the remaining point cloud in step 2 is:
[0023] Step 3.1: Statistical filtering: Take the k points around each point in the point cloud as its neighborhood points, calculate the distance between the point and its neighborhood points respectively, and calculate the average distance d a , make a Gaussian function with d a is the horizontal coordinate corresponding to the peak value, and the distance from this point is d a The number of points is taken as the ordinate corresponding to the peak value. From the Gaussian function formula, we can know that the corresponding Gaussian function parameter mean μ and standard deviation σ can be obtained through these two data. The 90% confidence interval of the Gaussian distribution function is taken, and the corresponding maximum distance in the interval is set to d max , if the average distance d of a pointa >d max , classify the point cloud as a noise point and remove it, otherwise, retain the point and obtain the three-dimensional point cloud after statistical filtering;
[0024] Step 3.2: Correct the normal vector of the points after statistical filtering: set the angle threshold T and use the radius search to obtain the point p to be corrected i k neighborhood points of , and calculate the constraint factor of each point:
[0025]
[0026] Determine point p i The corrected normal vector is expressed as: Where n i is the point to be corrected p i The normal vector, n ij Indicates p i The normal vectors of the j neighboring points of .
[0027] Traverse the points in the point cloud and correct the normal vector of each point;
[0028] Step 3.3: Perform bilateral filtering on the corrected points:
[0029] For the target point p i and its neighboring point p ij , calculate the filter parameters u1,u2, where u1=||p i -p ij || represents the target point p i Distance to neighboring points; u2 = <n i ,p i -p ij > represents the target point p i The inner product of the signed distance to a neighboring point and the normal vector of that point;
[0030] according to and Calculation Function and and bring in the value of Calculate the weight factor α;
[0031] Update target point information: p i :=p i +αn i , complete the filtering of the current point.
[0032] Preferably, the three-dimensional point cloud obtained in step 3 is segmented using Meanshift clustering, and the specific method for obtaining the target point cloud set is:
[0033] Step 4.1: Define point p in the point cloudi and its neighboring point p ij , randomly select a point p in the unlabeled point cloud data i As a starting point;
[0034] Step 4.2: Find the set of neighborhood points centered on the starting point through radius search, treat them as the same category, and mark these points in the record, with the number of times they appear in the category increased by 1;
[0035] Step 4.3: Calculate the mean shift vector with the starting point as the center and pass it through p iter+1 =p iter +M h (p iter ) iteratively updates the starting point, where p iter is the new starting position obtained at the iterth iteration, M h is the mean shift vector;
[0036] Step 4.4: Repeat steps 4.2 and 4.3 until the modulus of the mean shift vector is less than the set threshold, stop the iteration, and record the starting point obtained at this time as the center point;
[0037] Step 4.5: Calculate the distance between the center point obtained in this iteration and the existing center point. If it is less than the distance threshold, the two categories are merged, otherwise it is recorded as a new category.
[0038] Step 4.6: Repeat steps 4.2 to 4.5, traverse the points in the point cloud until all points are visited, count the number of times each point is visited by each category, take the class with the largest number of visits as the class to which the point belongs, and complete the Meanshift clustering operation.
[0039] Compared with the prior art, the present invention has the following significant advantages:
[0040] (1) The present invention adopts a model based on a second-order polynomial to fit the ground, which is more in line with the characteristics of the road being "high in the middle and low on both sides", and can effectively filter out the ground point cloud;
[0041] (2) The present invention uses statistical filtering and bilateral filtering based on normal vector and Euclidean distance to perform joint point cloud denoising, effectively removing multi-scale noise;
[0042] (3) The hyperparameters used in the training phase of the point vector machine in the present invention, namely, the kernel function parameter γ and the penalty coefficient C, are not set manually, but are obtained through training based on the data set, so that the trained classification model is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1It is a flow chart of road scene target recognition based on laser radar point cloud of the present invention.
[0044] Figure 2 LiDAR collects point cloud data of road scenes.
[0045] Figure 3 It is a schematic diagram of point cloud comparison before and after removing ground point cloud data.
[0046] Figure 4 It is a schematic diagram of point cloud comparison before and after multi-scale joint denoising.
[0047] Figure 5 It is a point cloud image obtained by segmenting point cloud data using the Meanshift 3D point cloud clustering algorithm.
[0048] Figure 6 This is an example of some of the processed data set files.
[0049] Figure 7 It is the support vector machine classification process with adaptive parameter selection. DETAILED DESCRIPTION
[0050] In order to make the patent, technical solutions and advantages of this application more clear, the application is further described in detail below in conjunction with the accompanying drawings and specific implementations.
[0051] In a specific embodiment, in combination Figure 1 , provides a road scene target recognition method based on laser radar point cloud, the method comprising the following steps:
[0052] Step 1: Use LiDAR to collect road scene point cloud data, including the target's spatial coordinate information and reflectivity information. The point cloud space visualization is as follows: Figure 2 ,LiDAR obtains information on the road scene and converts it into a point cloud of three-dimensional coordinates.
[0053] Step 2: Filter the ground point cloud from the collected point cloud information, determine the ground point cloud based on the quadratic polynomial model and random sampling consistency, and extract it from the point cloud data to achieve the reduction of the point cloud data size, such as Figure 3 , the specific process is as follows:
[0054] Step 2.1: According to the current scene and the selection of the coordinate system, set the initial ground height threshold d. Take z = 0 as the reference of the ground plane, then randomly select 6 data points from z∈[-d, d] and fit the initial ground model according to the quadratic polynomial fitting road surface expression as the current optimal model. In a specific embodiment, set d = 0.2m;
[0055] The quadratic polynomial fitting road surface expression selected is: z = a0 + a1x + a2y + a3xy + a4x 2 +a5y 2 .
[0056] Step 2.2: Randomly select 6 different data points from z∈[-d,d] again and fit the new ground model according to step 2.1;
[0057] Step 2.3: Compare the number of inliers between the current optimal model and the new model, and update the model with more inliers as the current optimal model;
[0058] Specifically, the method for determining the interior point of the model is: traverse the data points in the point cloud and determine whether they are interior points. The specific determination method is: determine the corresponding z value of the point (x, y) on the current fitting surface according to the quadratic polynomial fitting road surface expression, expressed as z g , the z-axis coordinate value corresponding to the actual coordinates of the point (x, y) is z t , when |z t -z g |≤e, the point is considered as an interior point, otherwise it is considered as an exterior point, and e is the distance error threshold. The distance error threshold e is determined as That is 0.05m.
[0059] Step 2.4: Repeat steps 2.2-2.3 until the maximum number of iterations is reached. When , the process ends and the internal points of the current optimal model are recorded as ground points.
[0060] Where u is the ratio of the ground point cloud to the entire point cloud, N is the number of points in the model, P = 1-(1-u N ) t Represents the confidence of the model. In the calculation of the maximum number of iterations, the confidence P is set to 95%. In the embodiment, the ground point cloud accounts for 19.9% and the number of ground point clouds is 59,800.
[0061] Step 3: Filter the point cloud in step 2. First, use statistical filtering to remove isolated outliers. Then use bilateral filtering based on the point cloud normal vector and Euclidean distance to correct the point cloud position on the target surface to remove noise point clouds. Figure 4 , the specific process is as follows:
[0062] Step 3.1: First, perform statistical filtering: take the k points around each point in the point cloud as its neighborhood points, calculate the distance between the point and its neighborhood points, and calculate the average distance d a ,
[0063] Draw a Gaussian function with d ais the horizontal coordinate corresponding to the peak value, and the distance from this point is d a The number of points is taken as the ordinate corresponding to the peak value. From the Gaussian function formula, we can know that the corresponding Gaussian function parameter mean μ and standard deviation σ can be obtained through these two data. The 90% confidence interval of the Gaussian distribution function is taken, and the corresponding maximum distance in the interval is set to d max , if the average distance d of a point a >d max , classify the point cloud as a noise point and remove it, otherwise, retain the point to obtain a three-dimensional point cloud after statistical filtering. In a specific embodiment, the number of neighborhood points of the point cloud is set to k = 50, and the distance threshold d max =0.5m.
[0064] Step 3.2: Correct the normal vector: The points after the statistical filtering in step 3.1 are the points to be corrected, an angle threshold T is set, and the correction point p is obtained by radius search. i k neighborhood points of , and calculate the constraint factor of each point:
[0065]
[0066] Determine point p i The corrected normal vector is expressed as: Where n i is the point to be corrected p i The normal vector, n ij Indicates p i The normal vectors of the j neighboring points of . Traverse the points in the point cloud and correct the normal vector of each point.
[0067] In a specific embodiment, the number of neighborhood points of the point cloud is set to T=50°.
[0068] Step 3.3: Perform bilateral filtering: For the target point p i and its neighboring point p ij , calculate the filter parameters u1,u2, where u1=||p i -p ij || represents the target point p i Distance to neighboring points; u2 = <n i ,p i -p ij > represents the target point p i The inner product of the directed distance to the neighboring point and the normal vector of this point; according to and Calculation Function and and bring in the value of Obtain the weight factor α; Update the target point information: p i :=pi +αn i , complete the filtering of the current point; when all points in the point set have been calculated, the process ends.
[0069] Step 4: Use Meanshift clustering to segment the 3D point cloud obtained in step 3, as shown in Figure 5 , the process is as follows:
[0070] Step 4.1: Define point p in the point cloud i and its neighboring point p ij , randomly select a point p in the unlabeled point cloud data i As a starting point;
[0071] Step 4.2: Find the set of neighborhood points centered on the starting point through radius search, treat them as the same category, and mark these points in the record, with the number of times they appear in the category increased by 1;
[0072] Step 4.3: Calculate the mean shift vector with the starting point as the center and pass it through p iter+1 =p iter +M h (p iter ) iteratively updates the starting point; where p iter The new starting position obtained at the iterth iteration.
[0073] In Meanshift, the Gaussian kernel function is used as the kernel function. The calculation formula of the mean shift vector after the kernel function is introduced is:
[0074]
[0075] In a specific embodiment, h=0.2m, k=80.
[0076] Step 4.4: Repeat steps 4.2 and 4.3 until the modulus of the mean shift vector is less than the set threshold, stop the iteration, and record the starting point obtained at this time as the center point;
[0077] Step 4.5: Calculate the distance between the center point obtained in this iteration and the existing center point. If it is less than the distance threshold, the two categories are merged, otherwise it is recorded as a new category.
[0078] Step 4.6: Repeat steps 4.2 to 4.5, traverse the points in the point cloud until all points are visited, count the number of times each point is visited by each category, take the class with the largest number of visits as the class to which the point belongs, and finally complete the Meanshift clustering operation.
[0079] Step 5: Describe the features of all target point clouds obtained in step 4 from 20 dimensions respectively, and use these 20 feature descriptors as the feature information of the target to form a data set. Some data sets are as follows: Figure 6 .
[0080] Table 1 Eigenvalues and their calculation formulas
[0081]
[0082]
[0083]
[0084] Step 6: For the data set in step 5, organize it to obtain the data shown in Table 1 below:
[0085] Table 2 Example of collected data
[0086] category total car 880 people 807 Bicyclists 225 other 52
[0087] The data set is divided into training set and test set according to a certain ratio.
[0088] Next, we train the two hyperparameters in the support vector machine, the kernel function and the penalty coefficient. The main process is to first determine the parameter range globally, and then further confirm it locally. After determining the hyperparameters, we train the classification model based on the training set. The prediction stage relies on the model obtained through training to confirm the target classification in the test set, such as Figure 7 The specific process is as follows:
[0089] Step 6.1: Randomly initialize the search space to obtain the initial parameters;
[0090] Step 6.2: During the parameter determination process, when the number of iterations is t n , the maximum number of iterations is T max ,set up: C(t n )=2·r, where r∈[0,1] is a random number, A(t n ) is the learning rate in the parameter determination process. When the parameter -1<A<1, a local search is performed in the algorithm, otherwise a global search is performed;
[0091] In the global search phase, after a set of particles are initialized, a search particle is selected and represented as X. r , while the search particle is updating data, other particles perform global search operations and move away from the search particle, so that the limited particles cover a larger search area; in the process, the distance between other particles and the search particle is expressed as:
[0092]
[0093]
[0094] In the local search phase, the optimal particle in each iteration is considered to be the solution closest to the target. In this phase, all other particles move toward the optimal particle and update their position information. In the particle movement mechanism, the optimal solution is obtained by continuously reducing the movement radius through a spiral ascending trajectory. * is the optimal target position, and the distance between the search particle and the target position is:
[0095]
[0096] The position of the search particle is updated iteratively by the spiral equation:
[0097]
[0098] Where b is the shape parameter, which is a constant, l∈[-1,1] and is a random number;
[0099] During the entire iterative process, according to the change of parameter A, the search mode switches between local search and global search, and the parameters are continuously updated. The iteration is stopped when the termination condition is met. The solution at this time is considered to be the final parameter.
[0100] Here, the classification rate is used as the fitness for calculation.
[0101] Step 6.3: Set the number of iterations when the fitness tends to be stable as the maximum number of iterations, stop the iteration when the maximum number of iterations is reached and obtain the optimal parameters. In this embodiment, the maximum number of iterations is 28 times, the adaptively determined kernel function parameter γ value is 5.5, and the penalty coefficient C is 5.9.
[0102] Step 6.4: Use the calculated optimal parameters as the initial parameters of the SVM multi-classifier to establish the model. The multi-classifier uses a one-to-one support vector machine classifier. In this embodiment, 4 categories need to be classified. According to the calculation, 6 binary classifiers are required. In each binary classifier, the sample to be tested calculates the feature values of two different categories and outputs the results. Finally, the classification result is generated through the "voting" mechanism of the category.
[0103] Step 6.5: Test the data in the test set with the trained model to obtain the target classification prediction results.
[0104] The classification results are shown in Table 2 below:
[0105]
[0106]
[0107] This embodiment is based on the point cloud data processing of Livox Mid-40 laser radar, and the classification targets are only cars, pedestrians, cyclists and other four categories. The detection accuracy and efficiency are improved compared with other point cloud target detection methods. At the same time, the present invention is not limited to this embodiment and is universal for other models of laser radars or more classification standards.
Claims
1. A road scene target recognition method based on laser radar point cloud, characterized in that: The steps include: Step 1: Use LiDAR to collect road scene point cloud data; Step 2: Determine the ground point cloud based on the quadratic polynomial model and random sampling consistency, and delete the ground point cloud data from the point cloud data, specifically: The specific method is: Step 2.1: Set the initial ground height threshold d, take z = 0 as the reference of the ground plane, randomly select 6 data points from z∈[-d, d], and fit the initial ground model according to the quadratic polynomial fitting road surface expression as the current optimal model; Step 2.2: Randomly select 6 different data points from z∈[-d,d] and fit the new ground model according to step 2.1; Step 2.3: Compare the number of inliers between the current optimal model and the new ground model, and update the model with more inliers as the current optimal model; Step 2.4: Repeat steps 2.2-2.3 until the maximum number of iterations is reached, the process ends, and the internal points of the current optimal model are recorded as the ground point cloud; Step 3: Filter the remaining point cloud in step 2, including removing isolated outliers using statistical filtering and correcting the point cloud position on the target surface based on bilateral filtering of the point cloud normal vector and Euclidean distance; Step 4: Use Meanshift clustering to segment the 3D point cloud obtained in step 3 to obtain the target point cloud set; Step 5: Describe the target point cloud set obtained in step 4 and extract structural feature information; Step 6: Input the extracted structural feature information into the trained support vector machine model to classify the target point cloud.
2. The road scene target recognition method based on laser radar point cloud according to claim 1 is characterized in that: The road scene point cloud data includes spatial coordinate information and reflectivity information of the target.
3. The road scene target recognition method based on laser radar point cloud according to claim 1 is characterized in that: The method to determine the interior points of the model is: Traverse the data points in the point cloud and determine whether they are internal points. The specific determination method is: according to the quadratic polynomial fitting road surface expression, determine the corresponding z value of the point (x, y) on the current fitting surface, expressed as z g , the z-axis coordinate value corresponding to the actual coordinates of the point (x, y) is z t , when |z t -z g |≤e, the point is judged as an interior point, where e is the distance error threshold.
4. The road scene target recognition method based on laser radar point cloud according to claim 1 is characterized in that: The maximum number of iterations is: Where u is the ratio of the point cloud of ground points to the entire point cloud, N is the number of points in the model, and P is the confidence of the model.
5. The method for road scene target recognition based on laser radar point cloud according to claim 1, characterized in that: The specific method for filtering the remaining point cloud in step 2 is: Step 3.1: Statistical filtering: Take the k points around each point in the point cloud as its neighborhood points, calculate the distance between the point and its neighborhood points respectively, and calculate the average distance d a , make a Gaussian function with d a is the horizontal coordinate corresponding to the peak value, and the distance from this point is d a The number of points is taken as the ordinate corresponding to the peak value. From the Gaussian function formula, we can know that the corresponding Gaussian function parameter mean μ and standard deviation σ can be obtained through these two data. The 90% confidence interval of the Gaussian distribution function is taken, and the corresponding maximum distance in the interval is set to d max , if the average distance d of a point a >d max , classify the point cloud as a noise point and remove it, otherwise, retain the point and obtain the three-dimensional point cloud after statistical filtering; Step 3.2: Correct the normal vector of the points after statistical filtering: set the angle threshold T and use the radius search to obtain the point p to be corrected i k neighborhood points of , and calculate the constraint factor of each point: Determine point p i The corrected normal vector is expressed as: where n i is the point to be corrected p i The normal vector, n ij Indicates p i Normal vectors of the j neighboring points of ; Traverse the points in the point cloud and correct the normal vector of each point; Step 3.3: Perform bilateral filtering on the corrected points: For the target point p i and its neighboring point p ij , calculate the filter parameters u1,u2, where u1=||p i -p ij || represents the target point p i Distance to neighboring points; u2 = <n i ,p i -p ij > represents the target point p i The inner product of the signed distance to a neighboring point and the normal vector of that point; according to and Calculation Function and and bring in the value of Calculate the weight factor α; Update target point information: p i :=p i +αn i , complete the filtering of the current point.
6. The road scene target recognition method based on laser radar point cloud according to claim 1 is characterized in that: The specific method of using Meanshift clustering to segment the three-dimensional point cloud obtained in step 3 and obtain the target point cloud set is as follows: Step 4.1: Define point p in the point cloud i and its neighboring point p ij , randomly select a point p in the unlabeled point cloud data i As a starting point; Step 4.2: Find the set of neighborhood points centered on the starting point through radius search, treat them as the same category, and mark these points in the record, with the number of times they appear in the category increased by 1; Step 4.3: Calculate the mean shift vector with the starting point as the center and pass it through p iter+1 =p iter +M h (p iter ) iteratively updates the starting point, where p iter is the new starting position obtained at the iterth iteration, M h is the mean shift vector; Step 4.4: Repeat steps 4.2 and 4.3 until the modulus of the mean shift vector is less than the set threshold, stop the iteration, and record the starting point obtained at this time as the center point; Step 4.5: Calculate the distance between the center point obtained in this iteration and the existing center point. If it is less than the distance threshold, the two categories are merged, otherwise it is recorded as a new category. Step 4.6: Repeat steps 4.2 to 4.5, traverse the points in the point cloud until all points are visited, count the number of times each point is visited by each category, take the class with the largest number of visits as the class to which the point belongs, and complete the Meanshift clustering operation.
7. The method for road scene target recognition based on laser radar point cloud according to claim 6, characterized in that: The calculation formula of the mean shift vector is: Where h is the search radius of the target point neighborhood, k is the number of point clouds within the search radius, and the Gaussian kernel function G used is specifically defined as:
8. The method for road scene target recognition based on laser radar point cloud according to claim 1, characterized in that: Characterization is performed from 20 dimensions, including:
9. The method for road scene target recognition based on laser radar point cloud according to claim 1, characterized in that: The training method of the support vector machine model is: Step 6.1: Randomly initialize the search space to obtain the initial parameters; Step 6.2: During the parameter determination process, when the number of iterations is t n , the maximum number of iterations is T max ,set up: Among them, r∈[0,1] is a random number, A(t n ) is the learning rate in the parameter determination process. When the parameter -1<A<1, a local search is performed in the algorithm, otherwise a global search is performed; In the global search phase, after a set of particles are initialized, a search particle is selected and represented as X. r , while the search particle is updating data, other particles perform global search operations and move away from the search particle, so that the limited particles cover a larger search area; in the process, the distance between other particles and the search particle is expressed as: In the local search phase, the optimal particle in each iteration is considered to be the solution closest to the target. In this phase, all other particles move toward the optimal particle and update their position information. In the particle movement mechanism, the optimal solution is obtained by continuously reducing the movement radius through a spiral ascending trajectory. * is the optimal target position, and the distance between the search particle and the target position is: The position of the search particle is updated iteratively by the spiral equation: Where b is the shape parameter, which is a constant, l∈[-1,1] and is a random number; During the entire iteration process, according to the change of parameter A, the search mode switches between local search and global search, and the parameters are continuously updated. The iteration stops when the termination condition is met. The solution at this time is considered to be the final parameter. Step 6.3: Use the calculated optimal parameters as the initial parameters of the SVM multi-classifier to establish the model; the multi-classifier uses a one-to-one support vector machine classifier.