Target tracking method and system based on DBSCAN and Kmeans

Abstract

The present invention discloses a target tracking method and system based on DBSCAN and Kmeans, which relates to the field of target tracking technology. The method comprises the following steps: marking all point cloud data sets as unvisited; sorting all point cloud data sets in descending order according to signal-to-noise ratio; selecting a point in sequence as a research object; setting an association radius according to the speed of the research object and a density threshold according to the signal-to-noise ratio of the research object; performing cluster analysis on the research object using the DBSCAN method and outputting clusters; performing an over-clustering judgment on the clusters; if no cluster exists, outputting the clusters; if so, calculating the number of possible targets; performing secondary clustering on the DBSCAN clusters using the Kmeans algorithm and outputting a new cluster with the same number of targets. The method adds an over-clustering judgment after DBSCAN clustering to prevent the same target from being divided into multiple clusters, thereby improving the accuracy of the clustering results and the accuracy of traffic flow statistics.

Description

Technical Field

[0001] The present invention relates to the technical field of target tracking, and in particular to a target tracking method and system based on DBSCAN and Kmeans. Background Art

[0002] Millimeter-wave radar, a common sensor, is increasingly being used in road traffic monitoring due to its wide detection range, high accuracy, and strong anti-interference capabilities. It tracks multiple targets in real time, implements traffic flow statistics, speeding detection, and other functions, and provides technical and data support for intelligent transportation systems. As millimeter-wave radar's carrier frequency gradually shifts from 24 GHz to 77 GHz and higher, its resolution is also increasing. The number of scattering points generated by processing the reflected echoes from the same target is also increasing. Furthermore, in actual application scenarios, multiple targets need to be detected and tracked. Therefore, correctly clustering multiple scattering points from the same target before tracking is a key research topic in the millimeter-wave radar field.

[0003] Clustering algorithm is a solution strategy for unsupervised learning cases. It is applicable to the case where the class information is unknown and the data is divided into clusters based on the similarity of the data.

[0004] Different clustering algorithms are suitable for different data scenarios. According to the characteristics of millimeter wave point cloud radar data, the density-based DBSCAN clustering algorithm is the most common and effective clustering method. At present, in the field of millimeter wave point cloud radar clustering, people mainly optimize based on the DBSCAN algorithm to adapt to different scenarios and different targets. In a certain actual traffic scenario, two cars are moving in parallel. Projecting the point traces into the Cartesian coordinate system, we get the following: Figure 1 The point distribution diagram shown in Figure 1 is a graph of the vehicle's trajectory. The red circles represent scattered points from the vehicle marked with a red box in the actual scene, while the blue stars represent scattered points from the vehicle marked with a blue box in the actual scene. Using the standard DBSCAN algorithm will cluster two vehicles into one cluster, which can easily cause target loss. Furthermore, improperly setting the association radius and density threshold can split the same target into multiple clusters, leading to over-clustering and inaccurate target tracking results. Summary of the Invention

[0005] The present invention aims to provide a target tracking method and system method based on DBSCAN and Kmeans, which solves the problems of target loss and inaccurate target tracking results when multiple vehicles are driving in parallel.

[0006] The present invention is achieved through the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a target tracking method based on DBSCAN and Kmeans, comprising:

[0008] Use millimeter-wave radar to obtain point cloud datasets of moving targets in road scenes;

[0009] Mark all point cloud datasets as unvisited;

[0010] Sort all point cloud datasets in descending order according to the signal-to-noise ratio;

[0011] Select a point as the research object in sequence and mark the research object as visited;

[0012] Set the correlation radius according to the speed of the research object, and set the density threshold according to the signal-to-noise ratio of the research object;

[0013] The DBSCAN clustering method was used to perform cluster analysis on the research objects and output the DBSCAN clusters;

[0014] The a priori target lateral size is used as the threshold to judge the DBSCAN clustering clusters;

[0015] If there is no over-clustering, the DBSCAN cluster is output;

[0016] If there is over-clustering, calculate the number of possible targets;

[0017] The Kmeans algorithm is used to perform secondary clustering on the DBSCAN clusters and output the target number of new clusters.

[0018] Furthermore, the specific method of setting the correlation radius according to the speed of the research object includes:

[0019] The function of the correlation radius set according to the research object changes with the speed is:

[0020] ;

[0021] Where Ɛ is the associated radius in meters, and v is the velocity of the point in meters per second.

[0022] Furthermore, the specific method of setting the density threshold according to the signal-to-noise ratio of the research object includes:

[0023] The density threshold for clustering is set according to the signal-to-noise ratio. The calculation formula for the density threshold is:

[0024] ;

[0025] Where MinPts is the density threshold and snr is the signal-to-noise ratio in dB.

[0026] Furthermore, the specific method of using the a priori target lateral size as a threshold to perform over-clustering judgment on the DBSCAN clusters includes:

[0027] Perform point trace contour analysis on the internal points of the cluster to calculate the vehicle length and width information;

[0028] Set a maximum vehicle width threshold and compare the vehicle width with the maximum vehicle width threshold;

[0029] If the vehicle width information is less than or equal to the maximum vehicle width threshold, it is determined that there is no over-clustering;

[0030] If the vehicle width information is greater than the maximum vehicle width threshold, it is determined that over-clustering exists.

[0031] Furthermore, the specific method of performing cluster analysis on the research objects using the DBSCAN clustering method includes:

[0032] The number of objects within the neighborhood of the research object's correlation radius is compared with the density threshold;

[0033] If the number of objects is greater than the density threshold, the research object is a noise point;

[0034] If the number of objects is less than or equal to the density threshold, the research object is taken as the core point, and the research object and the objects in the neighborhood associated with the research object are added to the newly created cluster;

[0035] Select a point from the newly created cluster. If the point is marked as unvisited, mark it as visited, calculate the number of objects in the neighborhood of the point's associated radius, and compare the number of objects in the neighborhood of the point's associated radius with the density threshold.

[0036] If the number of objects in the neighborhood of the point's associated radius is greater than the density threshold, the point is considered a core point, and the objects in the point's neighborhood are added to the newly created cluster;

[0037] If the number of objects in the neighborhood of the point's associated radius is less than or equal to the density threshold, the point is considered a boundary point;

[0038] Loop through the cluster until all points in the newly created cluster are marked as visited.

[0039] In a second aspect, an embodiment of the present invention provides a target tracking system based on DBSCAN and Kmeans, comprising: a data acquisition module, a first clustering module, an over-clustering judgment module, and a second clustering module.

[0040] The data acquisition module uses millimeter wave radar to acquire a point cloud data set of moving targets in a road scene;

[0041] The first clustering module is used to mark all point cloud data sets as unvisited, sort all point cloud data sets in descending order according to the signal-to-noise ratio, select a point as a research object in order, mark the research object as visited, set the association radius according to the speed of the research object, set the density threshold according to the signal-to-noise ratio of the research object, use the DBSCAN clustering method to perform cluster analysis on the research object, and output the DBSCAN cluster cluster;

[0042] The over-clustering judgment module uses the prior target horizontal size as a threshold to perform over-clustering judgment on the DBSCAN clusters. If there is no over-clustering, the DBSCAN clusters are output. If there is over-clustering, the number of possible targets is calculated.

[0043] The second clustering module is used to perform secondary clustering on the DBSCAN clusters using the Kmeans algorithm and output a target number of new clusters.

[0044] Furthermore, the first clustering module includes an association radius calculation unit, and the association radius calculation unit is used to calculate the association radius according to the speed of the research object.

[0045] Furthermore, the first clustering module includes a density threshold calculation unit, and the density threshold calculation unit is used to calculate the density threshold according to the signal-to-noise ratio of the research object.

[0046] Furthermore, the over-clustering judgment module includes a calculation unit and a comparison unit, wherein the calculation unit is used to perform point trace contour analysis on the internal points of the cluster to calculate the vehicle length and width information;

[0047] The comparison unit is used to compare the vehicle width with a preset maximum vehicle width threshold. If the vehicle width information is less than or equal to the maximum vehicle width threshold, it is determined that there is no over-clustering; if the vehicle width information is greater than the maximum vehicle width threshold, it is determined that there is over-clustering.

[0048] Furthermore, the first clustering module includes a cluster analysis unit, which compares the number of objects in the associated radius neighborhood of the research object with the density threshold. If the number of objects is greater than the density threshold, the research object is a noise point. If the number of objects is less than or equal to the density threshold, the research object is a core point. The research object and the objects in the associated neighborhood of the research object are added to the newly created cluster. A point is selected from the newly created cluster. If the point is marked as unvisited, it is marked as visited, and the number of objects in the associated radius neighborhood of the point is calculated. The number of objects in the associated radius neighborhood of the point is compared with the density threshold. If the number of objects in the associated radius neighborhood of the point is greater than the density threshold, the point is a core point. The objects in the neighborhood of the point are added to the newly created cluster. If the number of objects in the associated radius neighborhood of the point is less than or equal to the density threshold, the point is a boundary point. The loop is traversed until all points in the newly created cluster are marked as visited.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] The present invention provides a target tracking method based on DBSCAN and Kmeans. By adding over-clustering detection to the clusters generated by DBSCAN, this method prevents the same target from being split into multiple clusters, thereby improving the accuracy of the clustering results. Furthermore, the method utilizes basic prior knowledge to solve the problem of obtaining the k parameter in the Kmeans algorithm. This method is suitable for target tracking when two or more vehicles are traveling in parallel, accurately tracking targets and significantly improving the accuracy of traffic flow statistics.

[0051] The present invention provides a target tracking system based on DBSCAN and Kmeans. By adding over-clustering detection to the clusters generated by DBSCAN, this system prevents the same target from being split into multiple clusters, thereby improving the accuracy of clustering results. Furthermore, the system utilizes basic prior knowledge to solve the problem of obtaining the k parameter in the Kmeans algorithm. This system is suitable for tracking targets when two or more vehicles are traveling in parallel, accurately tracking targets and significantly improving the accuracy of traffic flow statistics. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings:

[0053] Figure 1 It is a point distribution map obtained by projecting the points in a real scene into the Cartesian coordinate system;

[0054] Figure 2 A flowchart of a target tracking method based on DBSCAN and Kmeans provided in the first embodiment of the present invention;

[0055] Figure 3 A comparison chart of point cloud clustering effects between the target tracking method based on DBSCAN and Kmeans provided by an embodiment of the present invention and the existing method;

[0056] Figure 4 A structural block diagram of a target tracking system based on DBSCAN and Kmeans is provided in another embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention. Example 1

[0058] like Figure 2 As shown, an embodiment of the present invention provides a target tracking method based on DBSCAN and Kmeans, including:

[0059] Use millimeter-wave radar to obtain a point cloud dataset P of moving targets in road scenes;

[0060] Mark all point cloud datasets P as unvisited;

[0061] Sort all point cloud datasets P in descending order according to the signal-to-noise ratio;

[0062] Select an unvisited object as the research object p in order, and mark the research object p as visited;

[0063] The correlation radius Ɛ is set according to the velocity of the research object p, and the density threshold (MinPts) is set according to the signal-to-noise ratio of the research object p;

[0064] The DBSCAN clustering method is used to perform cluster analysis on the research objects and output the DBSCAN cluster C;

[0065] The a priori target lateral size is used as the threshold to judge the DBSCAN clustering clusters;

[0066] If there is no over-clustering, the DBSCAN cluster is output;

[0067] If there is over-clustering, calculate the number of possible targets;

[0068] The Kmeans algorithm is used to perform secondary clustering on the DBSCAN cluster C and output the target number of new clusters.

[0069] The conventional DBSCAN clustering algorithm does not prioritize the input objects to be clustered, randomly selecting one as a core point for subsequent algorithms. This results in some samples being closer than the correlation radius to both core objects. However, since these two core objects are not directly accessible by density, they do not belong to the same cluster, yet the sample's category needs to be defined. Generally speaking, the DBSCAN clustering algorithm uses a first-come, first-served approach, with the first cluster to cluster the sample assigned to its category. However, radar point cloud data has signal-to-noise ratio information. The higher the signal-to-noise ratio, the more accurate the point's measurement information (distance, velocity, angle), making it more likely to be a target point. The lower the signal-to-noise ratio, the more likely it is a noise point. Therefore, starting clustering with points with high signal-to-noise ratios as core points can improve clustering results and ensure clustering stability. Therefore, in this embodiment, all point cloud datasets are first sorted in descending order of signal-to-noise ratio, and then a point is selected from this order as the target for analysis.

[0070] In actual traffic scenarios, the faster the speed of the vehicle target, the greater the following distance between vehicles. This is reflected in the radar detection points, where the distance between points of different targets is larger. Therefore, when clustering, the correlation radius Ɛ between the points can be appropriately increased without worrying about clustering points of different targets into one cluster. Conversely, when the vehicle speed is lower, the distance between different target points detected by the radar is smaller, and the correlation radius Ɛ needs to be reduced to prevent the points of different targets from being clustered into one cluster. Therefore, in this embodiment, the correlation radius is set according to the speed of the research object, and the function of the correlation radius changing with speed is as follows:

[0071] ,

[0072] Where Ɛ is the associated radius in meters, and v is the velocity of the point in meters per second.

[0073] In traffic application scenarios, there are many types of targets that need to be tracked, including cars, trucks, pedestrians, bicycles, etc. The density of different target traces is inconsistent. Through a lot of data analysis, we can have a simple understanding that the larger the size of the target, the higher the target trace density, which is also in line with the laws of physics. At the same time, through a lot of data analysis, basically the RCS (radar cross-section) of large-sized targets will also be larger, which is reflected in the traces as a larger signal-to-noise ratio. Therefore, the MinPts value during clustering can be adjusted according to the signal-to-noise ratio. The MinPts value is set to 3 levels, namely level 1, level 3, and level 5, as shown in the following formula:

[0074] ,

[0075] Where MinPts is the density threshold, snr is the signal-to-noise ratio, and the unit is dB (decibel).

[0076] In this embodiment, the association radius is set according to the speed of the point trace, and the density threshold during clustering is adjusted according to the signal-to-noise ratio. This can avoid clustering points of different targets into one cluster, and the output cluster is near the center of the actual target, thereby improving the accuracy of DBSCAN clustering.

[0077] In this embodiment, the specific method of using the DBSCAN clustering method to perform cluster analysis on the research objects includes:

[0078] The number of objects N in the neighborhood of the research object p is compared with the density threshold MinPts;

[0079] If the number of objects N is greater than the density threshold MinPts, the research object p is a noise point;

[0080] If the number of objects N is less than or equal to the density threshold MinPts, the research object p is taken as the core point, and the research object p and the objects in the neighborhood of the research object p are added to the newly created cluster c;

[0081] Select a point p2 from the newly created cluster c. If the point is marked as unvisited, mark it as visited, calculate the number of objects N2 in the neighborhood of p2, and compare the number of objects N2 with the density threshold MinPts.

[0082] If N2 is greater than MinPts, the point is a core point, and the objects in the neighborhood of the point are added to the newly created cluster c;

[0083] If N2 is less than or equal to MinPts, then the point is a boundary point;

[0084] Loop through until all points in the newly created cluster c are marked as visited;

[0085] Output DBSCAN cluster C;

[0086] Continue the above operation on the data set P until all points in the data set P have been visited.

[0087] In this embodiment, the specific method of using the a priori target lateral size as a threshold to perform over-clustering judgment on the DBSCAN clusters includes:

[0088] Perform point trace contour analysis on the internal points of the cluster to calculate the vehicle length and width information;

[0089] Set a maximum vehicle width threshold and compare the vehicle width with the maximum vehicle width threshold;

[0090] If the vehicle width information is less than or equal to the maximum vehicle width threshold, it is determined that there is no over-clustering, cluster C is normal, and cluster C is directly output;

[0091] If the vehicle width information is greater than the maximum vehicle width threshold, it is determined that over-clustering exists.

[0092] In real-world scenarios, vehicle lengths vary widely, ranging from a few meters to over ten meters. However, vehicle widths vary more narrowly, typically from 1.6m to 2.5m. This relatively small range can be used as prior knowledge to guide clustering. The specific steps are as follows:

[0093] Perform point trace contour analysis on the internal points of DBSCAN cluster C to calculate the vehicle length ysize and vehicle width xsize. Set the maximum vehicle width indicator Thr = 2.5m and compare xsize with Thr. If xsize ≤ Thr, the width of the clustered vehicle does not exceed the maximum limit. Therefore, it can be determined that the clustering is valid and multiple targets are not clustered together. If xsize > Thr, it is clear that the clustering result no longer meets the actual physical laws and must have clustered two or more targets together. Therefore, cluster C needs to be reallocated into multiple clusters. Specifically, the number of targets clustered together can be determined by rounding xsize / Thr. That is, K = xsize / Thr. The number of targets calculated in this way is basically consistent with the actual number through large-scale data analysis. In this embodiment, by using vehicle width to determine whether the clusters after DBSCAN clustering are over-clustered, it is possible to effectively determine whether the target clustering is accurate and improve the accuracy of the clustering results.

[0094] Clustering the points in cluster C into k targets is obviously easiest to achieve using the Kmeans algorithm. Since the value of k is known, the biggest flaw of the Kmeans algorithm is avoided. Therefore, the subsequent process uses the Kmeans algorithm to redistribute cluster C into K new clusters. The process of the Kmeans algorithm is as follows:

[0095] Randomly select k points in cluster C as the initial mean center , i=1,2…k;

[0096] Calculate the distance from other sample points to each mean center, assign the sample points to k clusters according to the principle of closest distance, and recalculate the mean center for each cluster , i=1,2…k;

[0097] Determine whether the difference between mi and ui is less than the set threshold;

[0098] If so, keep ui unchanged;

[0099] If not, replace ui with mi;

[0100] Determine whether all UIs no longer change;

[0101] If so, output as K clusters;

[0102] If not, return to the step of calculating the distances from other sample points to the mean center of each point.

[0103] like Figure 3 As shown, a comparison diagram of the point cloud clustering effects of the method of this embodiment and the existing DBSCAN algorithm is shown. From the actual results, it can be seen that the present invention can better adapt to target clustering in multi-vehicle parallel scenarios than the existing method.

[0104] The present invention provides a target tracking method based on DBSCAN and Kmeans. By adding over-clustering detection to the clusters generated by DBSCAN, this method prevents the same target from being split into multiple clusters, thereby improving the accuracy of the clustering results. Furthermore, the method utilizes basic prior knowledge to solve the problem of obtaining the k parameter in the Kmeans algorithm. This method is suitable for target tracking when two or more vehicles are traveling in parallel, accurately tracking targets and significantly improving the accuracy of traffic flow statistics. Example 2

[0105] like Figure 4 As shown, another embodiment of the present invention provides a target tracking system based on DBSCAN and Kmeans, including: a data acquisition module, a first clustering module, an over-clustering judgment module and a second clustering module, the data acquisition module uses millimeter wave radar to obtain a point cloud data set of a moving target in a road scene; the first clustering module is used to mark all point cloud data sets as unvisited, sort all point cloud data sets in descending order according to the signal-to-noise ratio, select a point in sequence as a research object, and mark the research object as visited, set the association radius according to the speed of the research object, set the density threshold according to the signal-to-noise ratio of the research object, use the DBSCAN clustering method to perform cluster analysis on the research object, and output DBSCAN cluster clusters; the over-clustering judgment module uses the prior target lateral size as a threshold to perform over-clustering judgment on the DBSCAN cluster clusters, if there is no over-clustering, then output the DBSCAN cluster cluster, if there is over-clustering, then calculate the possible number of targets; the second clustering module is used to use the Kmeans algorithm to perform secondary clustering on the DBSCAN cluster clusters, and output a new cluster of the target number.

[0106] The first clustering module includes a correlation radius calculation unit, which is used to calculate the correlation radius according to the speed of the research object. The first clustering module includes a density threshold calculation unit, which is used to calculate the density threshold according to the signal-to-noise ratio of the research object.

[0107] The over-clustering judgment module includes a calculation unit and a comparison unit. The calculation unit is used to perform point trace contour analysis on the internal points of the cluster to calculate the vehicle length and width information; the comparison unit is used to compare the vehicle width with a preset maximum vehicle width threshold. If the vehicle width information is less than or equal to the maximum vehicle width threshold, it is determined that there is no over-clustering; if the vehicle width information is greater than the maximum vehicle width threshold, it is determined that there is over-clustering.

[0108] The first clustering module includes a cluster analysis unit, which compares the number of objects in the associated radius neighborhood of the research object with the density threshold. If the number of objects is greater than the density threshold, the research object is a noise point. If the number of objects is less than or equal to the density threshold, the research object is a core point. The research object and the objects in the associated neighborhood of the research object are added to a newly created cluster. A point is selected from the newly created cluster. If the point is marked as unvisited, it is marked as visited, and the number of objects in the associated radius neighborhood of the point is calculated. The number of objects in the associated radius neighborhood of the point is compared with the density threshold. If the number of objects in the associated radius neighborhood of the point is greater than the density threshold, the point is a core point. The objects in the neighborhood of the point are added to the newly created cluster. If the number of objects in the associated radius neighborhood of the point is less than or equal to the density threshold, the point is a boundary point. The loop is traversed until all points in the newly created cluster are marked as visited.

[0109] The present invention provides a target tracking system based on DBSCAN and Kmeans. By adding over-clustering detection to the clusters generated by DBSCAN, this system prevents the same target from being split into multiple clusters, thereby improving the accuracy of the clustering results. It also uses basic prior knowledge to solve the problem of obtaining the k parameter in the Kmeans algorithm. This system is suitable for tracking targets when two or more vehicles are traveling in parallel, accurately tracking targets and significantly improving the accuracy of traffic flow statistics.

[0110] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A target tracking method based on DBSCAN and Kmeans, characterized in that: include: Use millimeter-wave radar to obtain point cloud datasets of moving targets in road scenes; Mark all point cloud datasets as unvisited; Sort all point cloud datasets in descending order according to the signal-to-noise ratio; Select a point as the research object in sequence and mark the research object as visited; Set the correlation radius according to the speed of the research object, and set the density threshold according to the signal-to-noise ratio of the research object; The DBSCAN clustering method was used to perform cluster analysis on the research objects and output the DBSCAN clusters; The specific method of using the DBSCAN clustering method to perform cluster analysis on the research object includes: The number of objects within the neighborhood of the research object's correlation radius is compared with the density threshold; If the number of objects is greater than the density threshold, the research object is a noise point; If the number of objects is less than or equal to the density threshold, the research object is taken as the core point, and the research object and the objects in its associated neighborhood are added to the newly created cluster; Select a point from the newly created cluster. If the point is unvisited, mark it as visited, calculate the number of objects in the neighborhood of the point's associated radius, and compare the number of objects in the neighborhood of the point's associated radius with the density threshold. If the number of objects in the neighborhood of the point's associated radius is greater than the density threshold, the point is considered a core point, and the objects in the point's neighborhood are added to the newly created cluster; If the number of objects in the neighborhood of the point's associated radius is less than or equal to the density threshold, the point is considered a boundary point; Loop through the cluster until all points in the newly created cluster are marked as visited; The a priori target lateral size is used as the threshold to judge the DBSCAN clustering clusters; The specific method of using the prior target horizontal size as a threshold to perform over-clustering judgment on the DBSCAN clusters includes: Perform point trace contour analysis on the internal points of the cluster to calculate the vehicle length and width information; Set a maximum vehicle width threshold and compare the vehicle width with the maximum vehicle width threshold; If the vehicle width information is less than or equal to the maximum vehicle width threshold, it is determined that there is no over-clustering; If the vehicle width information is greater than the maximum vehicle width threshold, it is determined that there is over-clustering; if there is no over-clustering, the DBSCAN cluster is output; If there is over-clustering, calculate the number of possible targets; The Kmeans algorithm is used to perform secondary clustering on the DBSCAN clusters and output the target number of new clusters.

2. The target tracking method based on DBSCAN and Kmeans according to claim 1, characterized in that: The specific method of setting the correlation radius according to the speed of the research object includes: The function of the correlation radius set according to the research object changes with the speed is: ； Where Ɛ is the associated radius in meters, and v is the velocity of the point in meters per second.

3. The target tracking method based on DBSCAN and Kmeans according to claim 2, characterized in that: The specific method of setting the density threshold according to the signal-to-noise ratio of the research object includes: The density threshold for clustering is set according to the signal-to-noise ratio. The calculation formula for the density threshold is: Where MinPts is the density threshold and snr is the signal-to-noise ratio in dB.

4. A target tracking system based on DBSCAN and Kmeans, characterized in that: The method for implementing the target tracking method based on DBSCAN and Kmeans according to any one of claims 1 to 3 comprises: a data acquisition module, a first clustering module, an over-clustering judgment module and a second clustering module, The data acquisition module uses millimeter wave radar to acquire a point cloud data set of moving targets in a road scene; The first clustering module is used to mark all point cloud data sets as unvisited, sort all point cloud data sets in descending order according to the signal-to-noise ratio, select a point as a research object in order, mark the research object as visited, set the association radius according to the speed of the research object, set the density threshold according to the signal-to-noise ratio of the research object, use the DBSCAN clustering method to perform cluster analysis on the research object, and output the DBSCAN cluster cluster; The over-clustering judgment module uses the prior target horizontal size as a threshold to perform over-clustering judgment on the DBSCAN clusters. If there is no over-clustering, the DBSCAN clusters are output. If there is over-clustering, the number of possible targets is calculated. The second clustering module is used to perform secondary clustering on the DBSCAN clusters using the Kmeans algorithm and output a target number of new clusters.

5. The target tracking system based on DBSCAN and Kmeans according to claim 4, characterized in that: The first clustering module includes an association radius calculation unit, which is used to calculate the association radius according to the speed of the research object.

6. The target tracking system based on DBSCAN and Kmeans according to claim 5, characterized in that: The first clustering module includes a density threshold calculation unit, which is used to calculate a density threshold according to a signal-to-noise ratio of a research object.

7. The target tracking system based on DBSCAN and Kmeans according to claim 6, characterized in that: The over-clustering judgment module includes a calculation unit and a comparison unit, wherein the calculation unit is used to perform point trace contour analysis on the internal points of the cluster to calculate the vehicle length and width information; The comparison unit is used to compare the vehicle width with a preset maximum vehicle width threshold. If the vehicle width information is less than or equal to the maximum vehicle width threshold, it is determined that there is no over-clustering; if the vehicle width information is greater than the maximum vehicle width threshold, it is determined that there is over-clustering.

8. The target tracking system based on DBSCAN and Kmeans according to claim 4, characterized in that: The first clustering module includes a cluster analysis unit, which compares the number of objects in the associated radius neighborhood of the research object with the density threshold. If the number of objects is greater than the density threshold, the research object is a noise point. If the number of objects is less than or equal to the density threshold, the research object is a core point. The research object and the objects in the associated neighborhood of the research object are added to the newly created cluster. A point is selected from the newly created cluster. If the point is unvisited, it is marked as visited, and the number of objects in the associated radius neighborhood of the point is calculated. The number of objects in the associated radius neighborhood of the point is compared with the density threshold. If the number of objects in the associated radius neighborhood of the point is greater than the density threshold, the point is a core point, and the objects in the neighborhood of the point are added to the newly created cluster. If the number of objects in the associated radius neighborhood of the point is less than or equal to the density threshold, the point is a boundary point. The loop is traversed until all points in the newly created cluster are marked as visited.

Images