A density-based fuzzy clustering method and device

Abstract

The present invention discloses a density-based fuzzy clustering method, comprising the following steps: obtaining current frame data, obtaining the number of clusters k, calculating the maximum difference in lateral distances between each cluster, updating the number of clusters, constructing an objective function, solving the objective function, obtaining a primary fuzzy clustering result, repeating until a termination condition is met and updating the membership matrix k and cluster centers, using the updated primary clustering result as input initial values to construct a new objective function, solving the objective function again to obtain a secondary fuzzy clustering result, iteratively updating the membership matrix and location center points, continuing if the secondary fuzzy clustering meets the iteration condition; otherwise, repeating, outputting the updated membership matrix and location center points to obtain a secondary fuzzy clustering result. The updated membership matrix is the final membership fuzzy matrix. The present invention utilizes this density-based fuzzy clustering method to optimize clustering results and effectively address the problem of adjacent vehicles of the same speed being clustered together.

Description

Technical Field

[0001] The present invention relates to the field of millimeter-wave radar point cloud data processing in traffic scenarios, and in particular to a density-based fuzzy clustering method and device. Background Art

[0002] In traffic scenarios, millimeter-wave radar point cloud data is sparse. Within the radar's monitoring range, data density is high in areas with targets and low in areas without targets. Furthermore, the number of detected targets in traffic scenarios is uncertain. Therefore, density-based clustering algorithms, such as DBSCAN, are well-suited for use in traffic scenarios. These algorithms can effectively cluster point cloud data in high-density areas into clusters, without requiring a predetermined number of clusters.

[0003] Furthermore, due to the complexity of traffic scenarios, the data density of the same vehicle varies at different distances from the radar, and different vehicle models also have different data densities. Point cloud data from large vehicles often has a wide scattering area. To ensure that the points are not fragmented, the parameter ε_d (distance neighborhood threshold) in the DBSCAN algorithm needs to be set large. In this case, the only way to distinguish adjacent vehicles is by comparing their speed differences.

[0004] However, when two adjacent vehicles are traveling at the same speed or the speed difference between them is within a threshold, the DBSCAN algorithm can easily cluster the two adjacent vehicles into one target.

[0005] In order to obtain better clustering effects in traffic scenarios, many researchers in this field are developing new clustering algorithms to obtain more accurate clustering results.

[0006] In April 2019, Su Tao et al. proposed a high-resolution millimeter-wave radar multi-target clustering method. Before clustering the point cloud data, this method first sets the signal-to-noise ratio detection threshold, filters the points detected by the radar, obtains valid points, and then clusters the valid points, reducing the amount of computation.

[0007] In December 2019, Zhang Chenyong proposed a density clustering method and device for radar target detection. This method statistically analyzes the distribution status of each target point detected. If the distribution is directional, an elliptical wave gate is used; otherwise, a circular wave gate is used.

[0008] In June 2020, Xiaoyun Liu proposed a principal axis-based density peak fuzzy clustering (SDPFC) algorithm. This algorithm uses the idea of secondary clustering to correct clustering results. It first performs clustering using the fast search and density peak (CFSFDP) algorithm to obtain the cluster centers of the initial clustering. This initial clustering result is then corrected using a principal axis-based fuzzy clustering algorithm to obtain the final clustering result.

[0009] In April 2021, Zhang Xinyi proposed the Adaptive Ellipse Distance Density Peak Fuzzy (AEDDPF) algorithm, which processes point cloud data through adaptive ellipse distance, and then uses the exponential function curve to quickly select the density peak points of the decision graph generated by the CFSFDP algorithm to obtain the cluster center and the number of clusters, and then uses the fuzzy clustering algorithm to iterate the initial clustering results to obtain the final clustering results.

[0010] During the same period, Liu Jiahuan proposed a method and system for automotive millimeter-wave radar track aggregation. This method prioritizes clustering tracks and track points based on the predicted target position, speed, and target size from the previous frame when clustering the current frame's track data. This effectively addresses the problem of large target fragmentation in automotive millimeter-wave radar.

[0011] In September 2021, He Yuyi and others proposed the elliptical DBSCAN clustering algorithm for automotive millimeter-wave radar. This algorithm converts the circular neighborhood in the DBSCAN clustering algorithm into an elliptical neighborhood, and can adaptively calculate the neighborhood parameters based on the characteristic information of the target data, thereby obtaining correct clustering results in the case of uneven data density.

[0012] In November 2021, Zhang Zhengwen et al. proposed an improved fuzzy clustering algorithm based on millimeter-wave radar. The algorithm combines the DPC algorithm and the FCM algorithm, and uses target position information and speed information for secondary FCM clustering respectively. It has high accuracy, good robustness, and can correctly cluster adjacent vehicle targets.

[0013] During the same period, Lu Hailing et al. proposed the inter-frame DBSCAN clustering algorithm, which used multi-frame merging processing and frame sequence features to solve the multipath noise problem, and used the spatial longitudinal segmentation method to improve the original algorithm's shortcomings in insufficient target discrimination in the longitudinal direction.

[0014] In March 2022, Xu Xiaoyong and others provided a millimeter-wave radar point cloud clustering method and device, which obtained the current frame radar point cloud information of the target to be detected and the track information output by the previous frame tracking; then, according to a preset algorithm, the current frame radar point cloud information was associated with the track information output by the previous frame tracking to obtain the track coordinates, heading angle and vehicle type classification of the target to be detected.

[0015] However, the above clustering methods all require large amounts of computation. Summary of the Invention

[0016] The purpose of the present invention is to provide a density-based fuzzy clustering method to optimize clustering results and effectively solve the problem of adjacent vehicles with the same speed being clustered into one cluster.

[0017] To achieve the above object, the present invention provides a density-based fuzzy clustering method, comprising the following steps:

[0018] S1: Obtain the current frame data based on the received point cloud data, perform DBSCAN pre-clustering first, and obtain the number of clusters k;

[0019] S2: Calculate the maximum difference in the horizontal distance of each cluster after the initial clustering, max_y, and update the number of clusters k new :

[0020]

[0021] Where d is the lane width;

[0022] S3: If the number of clusters remains unchanged after the update, perform point condensation on each cluster data point to obtain the initial cluster center v i , calculate the distance from all points to the center of each cluster, and the membership of all sample points to the center of each cluster μ ij , μ ij Represents the membership of the jth sample to the i-th cluster center and constructs the objective function J;

[0023] If the number of clusters changes after the update, only the number of clusters k is input new , randomly generate membership matrix μ ij , construct the objective function J;

[0024] S4: Use the Lagrange multiplier method to solve the objective function and obtain a fuzzy clustering result;

[0025] S5: Repeat step S4 until the fuzzy clustering iteration termination condition is met and the membership matrix is updated to μ' ij and cluster center v' i ;

[0026] S6: The clustering result u' updated by step S5 ij As the input initial value, substitute it into the FCM algorithm to construct a new objective function Q;

[0027] S7: Use the Lagrange multiplier formula again to solve the objective function Q, obtain the secondary fuzzy clustering result, and iteratively update the membership matrix μ' ij and position center point I i ;

[0028] S8: If the secondary fuzzy clustering satisfies the iteration condition, continue to execute S9, otherwise repeat step S8;

[0029] S9: Output updated membership matrix μ″ ij and position center point I' i, get the secondary fuzzy clustering result, and get the updated membership matrix μ″ ij This is the final membership fuzzy matrix.

[0030] Preferably, step S1 specifically includes the following sub-steps:

[0031] S11: Get current frame data

[0032] S111: Get the original target point data of the current frame point cloud {P j}, j = 1, 2, 3, ..., n, where n is the number of original target points, and the information of each target point includes P j ={r j ,v j ,θ j},r j is the distance information of the target, v j is the target’s velocity information, θ j is the angle information of the target;

[0033] S112: Convert the above point trace data into a rectangular coordinate system to obtain: P′ j ={r j ,x j ,y j ,v j ,θ j}

[0034] Among them, x j =r j *cosθ j ,y j =r j *sinθ j , x i and y i They are the position information of the original point trace in the rectangular coordinate system;

[0035] S12: Pre-clustering using the DBSCAN algorithm

[0036] S121: Determine the parameters in the pre-clustering: ellipse neighborhood parameters eps_d, a, b, and the minimum number of points in each cluster Minpts = 2;

[0037] S122: Initialize the number of clusters k = 0, the unvisited sample set Γ1 = P', the cluster partition set

[0038] S123: The trace data are processed by DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C1, C2, ..., C k}.

[0039] Preferably, in step S121, the eps_d neighborhood calculation is improved by the DBSCAN clustering algorithm:

[0040]

[0041] Among them, x=r*cos(θ); y=r*sin(θ);

[0042] Where x is the longitudinal distance, y is the transverse distance, a is the weight for balancing the longitudinal distance, and b is the weight for balancing the transverse distance.

[0043] Preferably, in step S3, if the number of clusters after update remains unchanged, the calculation formula of the objective function J is as follows:

[0044]

[0045] Among them, m represents the weighted value of each fuzzy degree of the sample to the cluster center;

[0046] The constraints are:

[0047]

[0048] Preferably, in step S4, the constraints are incorporated into the objective function using the Lagrange multiplier method, and the Lagrange function can be obtained as follows:

[0049]

[0050] Where λ is the Lagrange multiplier;

[0051] The membership degree μ ij and the initial cluster center v i Find the partial derivative and set it to 0, and we get:

[0052]

[0053] Preferably, the termination condition of one fuzzy clustering iteration in step S5 is:

[0054] ε=J new -J pre <ε0

[0055] J new Indicates the currently calculated objective function value, J pre represents the objective function value calculated in the last iteration, ε0 is the minimum update amount, and ε represents J new and J pre The difference between

[0056] ε0=0.0001, the number of calculation iterations is within 100.

[0057] Preferably, the termination condition of the secondary fuzzy clustering in step S8 is:

[0058] η=Q new -Q pre <η0

[0059] Q new Indicates the currently calculated objective function value, Q pre represents the objective function value calculated in the last iteration, η0 is the minimum update amount, and η represents Q new and Q pre The difference between

[0060] η0=0.0001, the number of calculation iterations is within 100.

[0061] A device for density-based fuzzy clustering method, comprising:

[0062] Radar signal processing module, used to perform 2D-FFT, CFAR, and DOA processing on radar signals and output radar trace information;

[0063] Density clustering module, used to cluster point cloud data and ultimately output the distance, speed and angle information of each cluster center;

[0064] Program storage chip, used to burn the firmware into the flash and start the program from the external flash;

[0065] The radar is connected to the density clustering module via the radar signal processing module, and the density clustering module is also connected to the program storage chip.

[0066] Preferably, the radar trace information includes distance, speed, and angle information of each point.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] 1. In step S1, the DBSCAN algorithm with improved eps neighborhood range is more suitable for use in traffic scenarios, and the DBSCAN algorithm can automatically obtain the number of clusters.

[0069] 2. In step S2, the maximum difference in the y direction of each cluster of data is compared with the lane width, which can effectively detect whether two adjacent vehicles are clustered into one cluster.

[0070] 3. After obtaining the number of clusters through the DBSCAN algorithm, the number of clusters is updated based on the ratio of the maximum y-direction difference in each cluster to the lane width, solving the problem of poor clustering of adjacent vehicles of the same speed on urban roads. In addition, when no adjacent targets are clustered together, the cluster center is input to initialize a more accurate membership matrix and reduce the number of iterations. The fuzzy clustering algorithm can also be used to optimize the cluster center and obtain more accurate clustering results.

[0071] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a flow chart of the method of the present invention;

[0073] Figure 2 This is the effect diagram of clustering adjacent vehicles using only DBSCAN;

[0074] Figure 3 This is a diagram showing the clustering effect of adjacent vehicles according to the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0076] Figure 1 It is a flow chart of the method of the present invention, as shown in Figure 1 As shown, the present invention includes the following steps:

[0077] S1: Obtain the current frame data based on the received point cloud data, perform DBSCAN pre-clustering first, and obtain the number of clusters k;

[0078] Preferably, step S1 specifically includes the following sub-steps:

[0079] S11: Get current frame data

[0080] S111: Get the original target point data of the current frame point cloud {P j}, j = 1, 2, 3, ..., n, where n is the number of original target points, and the information of each target point includes P j ={r j ,v j ,θ j},r j is the distance information of the target, v j is the target’s velocity information, θ j is the angle information of the target;

[0081] S112: Convert the above point trace data into a rectangular coordinate system to obtain: P′ j ={r j ,x j ,y j ,v j ,θ j}

[0082] Among them, x j =r j *cosθ j ,y j =r j *sinθ j , x i and y i They are the position information of the original point trace in the rectangular coordinate system;

[0083] S12: Pre-clustering using the DBSCAN algorithm

[0084] S121: Determine the parameters in the pre-clustering: ellipse neighborhood parameters eps_d, a, b, and the minimum number of points in each cluster Minpts = 2;

[0085] Preferably, in step S121, the eps_d neighborhood calculation is improved by the DBSCAN clustering algorithm:

[0086]

[0087] Among them, x=r*cos(θ); y=r*sin(θ);

[0088] Where x is the longitudinal distance, y is the lateral distance, a is the weight for balancing the longitudinal distance, and b is the weight for balancing the lateral distance. a and b are set according to the actual road conditions.

[0089] S122: Initialize the number of clusters k = 0, the unvisited sample set Γ1 = P′, the cluster partition set

[0090] S123: The trace data are processed by DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C1, C2, ..., C k}.

[0091] S2: Calculate the maximum difference in the horizontal distance of each cluster after the initial clustering, max_y, and update the number of clusters k new :

[0092]

[0093] Where d is the lane width;

[0094] S3: If the number of clusters after the update remains unchanged, that is, the width of each cluster is less than the lane width, then the data points of each cluster are condensed to obtain the initial cluster center v i , calculate the distance from all points to the center of each cluster, and the membership of all sample points to the center of each cluster μ ij , μ ij Represents the membership of the jth sample to the i-th cluster center and constructs the objective function J;

[0095] If the number of clusters changes after the update, only the number of clusters k is input new , randomly generate membership matrix μ ij , construct the objective function J;

[0096] Preferably, in step S3, if the number of clusters after update remains unchanged, the calculation formula of the objective function J is as follows:

[0097]

[0098] Among them, m represents the weighted value of each fuzziness of the sample to the cluster center, which is generally set to 2;

[0099] The constraints are:

[0100]

[0101] S4: Use the Lagrange multiplier method to solve the objective function and obtain a fuzzy clustering result;

[0102] Preferably, in step S4, the constraints are incorporated into the objective function using the Lagrange multiplier method, and the Lagrange function can be obtained as follows:

[0103]

[0104] Where λ is the Lagrange multiplier;

[0105] The membership degree μ ij and the initial cluster center v i Find the partial derivative and set it to 0, and we get:

[0106]

[0107] S5: From the above formula, we can see that the membership matrix μ ij and v i There is a certain correlation. Through continuous iteration, the optimal membership matrix and cluster center are obtained, and finally the objective function J tends to be stable. Therefore, step S4 is repeated until the iteration termination condition of the fuzzy clustering is met, and the membership matrix is updated to μ′ ij and cluster center v′ i ;

[0108] Preferably, the termination condition of one fuzzy clustering iteration in step S5 is:

[0109] ε=J new -J pre <ε0

[0110] J new Indicates the currently calculated objective function value, J pre represents the objective function value calculated in the last iteration, ε0 is the minimum update amount, and ε represents J new and J pre The difference between

[0111] ε0=0.0001, the number of calculation iterations is within 100.

[0112] S6: The clustering result u′ updated by step S5 ij As the input initial value, substitute it into the FCM algorithm to construct a new objective function Q;

[0113] S7: Use the Lagrange multiplier formula again to solve the objective function Q, obtain the secondary fuzzy clustering result, and iteratively update the membership matrix μ' ij and position center I i ;

[0114] S8: If the secondary fuzzy clustering satisfies the iteration condition, continue to execute S9, otherwise repeat step S8; preferably, the termination condition of the secondary fuzzy clustering in step S8 is:

[0115] η=Q new -Q pre <η0

[0116] Q new Indicates the currently calculated objective function value, Q pre represents the objective function value calculated in the last iteration, η0 is the minimum update amount, and η represents Q new and Q pre The difference between

[0117] η0=0.0001, the number of calculation iterations is within 100.

[0118] S9: Output updated membership matrix μ″ ij and position center point I' i , get the secondary fuzzy clustering result, and get the updated membership matrix μ″ ij This is the final membership fuzzy matrix.

[0119] A device for density-based fuzzy clustering method, comprising:

[0120] The radar signal processing module is used to perform 2D-FFT, CFAR, and DOA processing on the radar signal and output radar trace information; the radar trace information includes the distance, speed, and angle information of each point.

[0121] Density clustering module, used to cluster point cloud data and ultimately output the distance, speed and angle information of each cluster center;

[0122] Program storage chip, used to burn the firmware into the flash and start the program from the external flash;

[0123] The radar is connected to the density clustering module via the radar signal processing module, and the density clustering module is also connected to the program storage chip.

[0124] Figure 2 This is the effect diagram of clustering adjacent vehicles using only DBSCAN. Figure 3 This is the clustering effect diagram of adjacent vehicles in the present invention, as shown in Figure 2 and Figure 3 As shown in the figure, the present invention uses the DBSCAN algorithm to perform initial clustering on the point cloud data, obtain clustering results, and determine the number of targets. The maximum difference in the y direction of each cluster is calculated. If it is greater than the width of a lane, it can be determined that the cluster contains multiple targets. The maximum difference in the y direction divided by the lane width can be used to determine how many targets are clustered together at the same distance and speed.

[0125] If the width of each cluster is smaller than the lane width, the cluster center of each cluster (including the distance center and the speed center) can be calculated, and the speed center point can be used as the input condition of the algorithm to perform a fuzzy clustering to obtain a more accurate membership function μ ij and the corrected velocity center, and then μ ij The distance center point is used as the input condition of the algorithm to perform secondary fuzzy clustering and obtain the corrected distance center point. This makes the clustering result more accurate and effectively reduces the number of iterations of the fuzzy clustering algorithm.

[0126] If the width of a cluster is greater than the lane width, the number of clusters is updated proportionally, and the number of clusters is used as an input condition for fuzzy clustering, which can effectively separate adjacent vehicles that are clustered into one target in the initial DBSCAN clustering.

[0127] Therefore, the present invention adopts the above-mentioned density-based fuzzy clustering method to optimize the clustering results and effectively solve the problem of adjacent vehicles with the same speed being clustered into one cluster.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A density-based fuzzy clustering method, characterized by: The following steps are involved: S1: Obtain the current frame data based on the received point cloud data, perform DBSCAN pre-clustering first, and obtain the number of clusters k; S2: Calculate the maximum difference in the horizontal distance of each cluster after the initial clustering, max_y, and update the number of clusters k new : Where d is the lane width; S3: If the number of clusters remains unchanged after the update, perform point condensation on each cluster data point to obtain the initial cluster center v i , calculate the distance from all points to the center of each cluster, and the membership of all sample points to the center of each cluster μ ij , μ ij Represents the membership of the jth sample to the i-th cluster center and constructs the objective function J; If the number of clusters changes after the update, only the number of clusters k is input new , randomly generate membership matrix μ ij , construct the objective function J; S4: Use the Lagrange multiplier method to solve the objective function and obtain a fuzzy clustering result; S5: Repeat step S4 until the fuzzy clustering iteration termination condition is met and the membership matrix is updated to μ' ij and cluster center v' i ; S6: The clustering result u' updated by step S5 ij As the input initial value, substitute it into the FCM algorithm to construct a new objective function Q; S7: Use the Lagrange multiplier formula again to solve the objective function Q, obtain the secondary fuzzy clustering result, and iteratively update the membership matrix μ' ij and position center I i ; S8: If the secondary fuzzy clustering satisfies the iteration condition, continue to execute S9, otherwise repeat step S8; S9: Output updated membership matrix μ″ ij and position center point I' i , get the secondary fuzzy clustering result, and get the updated membership matrix μ″ ij This is the final membership fuzzy matrix.

2. The density-based fuzzy clustering method according to claim 1, characterized in that: Step S1 specifically includes the following sub-steps: S11: Get current frame data S111: Get the original target point data of the current frame point cloud {P j }, j = 1, 2, 3, ..., n, where n is the number of original target points, and the information of each target point includes P j ={r j ,v j ,θ j },r j is the distance information of the target, v j is the target’s velocity information, θ j is the angle information of the target; S112: Convert the above point trace data into a rectangular coordinate system to obtain: P′ j ={r j ,x j ,y j ,v j ,θ j } Among them, x j =r j *cosθ j ,y j =r j *sinθ j , x i and y i They are the position information of the original point trace in the rectangular coordinate system; S12: Pre-clustering using the DBSCAN algorithm S121: Determine the parameters in the pre-clustering: ellipse neighborhood parameters eps_d, a, b, and the minimum number of points in each cluster Minpts = 2; S122: Initialize the number of clusters k = 0, the unvisited sample set Γ1 = P', the cluster partition set S123: The trace data are processed by DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C1, C2, ..., C k }.

3. The density-based fuzzy clustering method according to claim 2, characterized in that: In step S121, the eps_d neighborhood calculation is improved by the DBSCAN clustering algorithm: Among them, x=r*cos(θ); y=r*sin(θ); Where x is the longitudinal distance, y is the transverse distance, a is the weight for balancing the longitudinal distance, and b is the weight for balancing the transverse distance.

4. The density-based fuzzy clustering method according to claim 3, characterized in that: In step S3, if the number of clusters after update remains unchanged, the calculation formula of the objective function J is as follows: Among them, m represents the weighted value of each fuzzy degree of the sample to the cluster center; The constraints are:

5. The density-based fuzzy clustering method according to claim 4, characterized in that: In step S4, the Lagrange multiplier method is used to bring the constraint conditions into the objective function, and the Lagrange function can be obtained as follows: Where λ is the Lagrange multiplier; The membership degree μ ij and the initial cluster center v i Find the partial derivative and set it to 0, and we get:

6. The density-based fuzzy clustering method according to claim 5, characterized in that: The termination condition of one fuzzy clustering iteration in step S5 is: e=J new -J pre <e0 J new Indicates the currently calculated objective function value, J pre represents the objective function value calculated in the last iteration, ε0 is the minimum update amount, and ε represents J new and J pre The difference between ε0=0.0001, the number of calculation iterations is within 100.

7. The density-based fuzzy clustering method according to claim 6, characterized in that: Termination condition of secondary fuzzy clustering in step S8: η=Q new -Q pre <η0 Q new Indicates the currently calculated objective function value, Q pre represents the objective function value calculated in the last iteration, η0 is the minimum update amount, and η represents Q new and Q pre The difference between η0=0.0001, the number of calculation iterations is within 100.

8. A device based on the density-based fuzzy clustering method according to any one of claims 1 to 7, characterized in that: include: Radar signal processing module, used to perform 2D-FFT, CFAR, and DOA processing on radar signals and output radar trace information; Density clustering module, used to cluster point cloud data and ultimately output the distance, speed and angle information of each cluster center; Program storage chip, used to burn the firmware into the flash and start the program from the external flash; The radar is connected to the density clustering module via the radar signal processing module, and the density clustering module is also connected to the program storage chip.

9. The apparatus for the density-based fuzzy clustering method according to claim 8, characterized in that: Radar point trace information includes the distance, speed, and angle information of each point.

Images