Multi-object point cloud clustering method

Abstract

A multi-target point cloud clustering method is provided. Based on the existing DBSCAN clustering algorithm, the method first improves the single clustering parameter of the algorithm into a multi-parameter including distance, speed and angle. Secondly, the cloud point range of the target at the previous moment and the state prediction value of the target at the next moment are used to predict the clustering area range of the target cloud point at the next moment. In addition, according to the state prediction value of the target at the next moment, the distance, speed and angle of the clustering algorithm and the minimum clustering condition parameters are adjusted. In this way, in the case of multi-target recognition, the recognition accuracy of the millimeter-wave radar for the target is significantly improved, and the situation of the target being merged and abandoned is greatly reduced.

Description

Technical Field

[0001] The invention relates to the field of radar signal target detection technology, and in particular to a multi-target point cloud clustering method used in multi-target discrimination of unknown clutter. Background Art

[0002] Point cloud refers to points distributed in three-dimensional space, which gather and present a cloud-like shape. Therefore, point cloud is a collection of massive points, and "point" is the basic unit of point cloud. Points obtained under various scene systems or measurement principles contain information such as three-dimensional space coordinates (XYZ), color information (RGB), and laser reflection intensity (intensity).

[0003] For example, due to its wide frequency band and high resolution, millimeter wave radar is widely used in target identification needs in technical fields such as intelligent driving, tracking, and monitoring. In these scenarios, the radar wave signal emitted by the millimeter wave radar is reflected by the target (fixed target, mobile target) in the covered airspace to obtain an echo. The target's echo signal is a point cloud composed of a large number of cloud points with distance, speed, and angle information.

[0004] Due to the advantages of wide bandwidth and high resolution mentioned above, the echo point cloud can effectively help the radar to identify the size and posture of the target, and these identification results can be used to make further judgments and predictions on the position and movement trend of the target. On the other hand, affected by unknown clutter and uncertain noise in the scene, and considering the differences in the reflection characteristics of electromagnetic waves of the target in different states, the number of target echo cloud points, the distribution density and distribution range of the cloud points in three-dimensional space will show uncertain changes with the scene changes and target movement, and the high-resolution characteristics of millimeter-wave radar may amplify the uncertain changes in the distribution density range of cloud points, thus requiring the processing of radar echo signal point cloud with higher accuracy.

[0005] One of the important bases that affects the processing accuracy of the echo signal point cloud and the accuracy of target discrimination is how to classify the cloud points that make up the point cloud. This classification process is usually called clustering. The purpose of clustering is to distinguish the cloud points in the echo point cloud according to the different targets they belong to, and to classify the cloud points belonging to the same target into one cluster. In this way, in the radar echo signal, the cloud points of different targets are presented in the form of different clusters, that is, each cluster in the space will correspond to a target in the covered airspace. Since all the cloud points in a cluster belong to and are used to represent the same target, in order to save memory and computing resources, after the clustering is completed, the cloud points in each cluster will be synthesized to obtain points in the three-dimensional space, and the coordinates of these points in the three-dimensional space are used to represent the status (distance, speed, angle) of the corresponding targets of each cluster in the covered airspace.

[0006] The clustering of cloud points needs to be realized with the help of clustering algorithms. The basic purpose of clustering algorithms is to group data with the greatest possible similarity in a data set into one group, and make the differences between each group of data as significant as possible. In the applications of autonomous driving and personnel tracking involving millimeter-wave radar, after completing the clustering of cloud points, the clustering algorithm will also provide the clustering results to the target tracking module for track association and target tracking. Among them, track association is to attribute the targets in the covered airspace obtained by clustering to the track of the tracked targets, and target tracking is to establish and update the target track and provide a predicted value of the target's state at the next moment.

[0007] After years of development, the existing technology has known a variety of clustering algorithms. These clustering algorithms have their own advantages and disadvantages, and are therefore used in different specific scenarios. Among them, the clustering algorithm mainly used for echo signal clustering processing is the DBSCAN (Density-Based Spatial Clustering of Application with Noise) algorithm. This density-based clustering algorithm can identify clusters of any shape and has good robustness to noise points. However, since the algorithm identifies clusters through customized core points and density thresholds, it is more sensitive to parameters. Summary of the invention

[0008] In view of this, the present invention provides a method for solving at least one of the above problems.

[0009] To solve the above technical problems, the multi-target point cloud clustering method provided by the present invention includes the following steps: constructing a spatial coordinate system based on the distance, speed and angle of the target relative to the radar, and aggregating the echo signal of the target on the radar at the current time t into a point cloud within the spatial coordinate system; configuring clustering algorithm parameters according to the distance, speed and angle respectively to determine whether each cloud point constituting the point cloud is a core cloud point, so as to obtain at least one core cloud point and multiple non-core cloud points; linking each of the core cloud points, and then linking each core cloud point and the non-core cloud points that can be linked to it to obtain at least one cluster; obtaining the distribution range of the cloud points of each cluster in terms of distance, speed and angle at the current time t respectively; obtaining the state prediction value of the target at the next time (t+1), and configuring the clustering area of ​​the cluster at the next time according to the state prediction value, as well as the clustering algorithm parameters of the clustering area at the next time.

[0010] As a more preferred scheme of the preferred embodiment of the present scheme, the step of configuring the clustering area of ​​the cluster at the next moment according to the predicted value specifically includes: configuring the change in distance and speed of the clustering area of ​​the cluster at the next moment according to the speed change of the target, and configuring the change in angle of the clustering area according to the angle change of the target.

[0011] As a further preferred embodiment of the present scheme, it also includes the step of uniquely classifying cloud points, specifically: when the echo point clouds of at least two clusters overlap, the step of obtaining the center points of each overlapping cluster; for any cloud point in the overlapping area, the step of obtaining its distance from the center point of each cluster; and the step of classifying the cloud point to the cluster with the shortest distance and the largest compactness among the clusters.

[0012] As a further preferred embodiment of the present invention, after dividing the cloud points into the nearest cluster, the method further includes: a step of judging whether the remaining points in other overlapping clusters satisfy the minimum clustering condition; if not, the cloud points in the cluster are judged as noise points.

[0013] As another preferred embodiment of the present scheme, the steps of configuring clustering algorithm parameters according to distance, speed and angle are specifically: the steps of setting distance parameter preset value, speed parameter preset value and angle parameter preset value: when any two points in the space are judged to be linkable in terms of distance, speed and angle, then the two points are determined to be linkable, otherwise the two points are determined to be unlinkable; the step of configuring according to the predicted value of the state of the cluster at the next moment to determine the distance parameter preset value, speed parameter preset value, angle parameter preset value and the minimum clustering condition at the next moment.

[0014] Further preferably, the step of determining whether two points can be linked is specifically as follows: for any two cloud points in the space, p=( , , ) and q=( , , ),in For distance, For distance, The distance is if it satisfies both:

[0015] ;

[0016] Then the cloud points p and q are judged to be linkable, where A preset value for the distance parameter, is a preset value for the speed parameter, A preset value for the angle parameter.

[0017] As a further preferred embodiment of this scheme, in the step of setting the preset value of the distance parameter, the preset value of the speed parameter and the preset value of the angle parameter, the preset value of the distance parameter and the preset value of the speed parameter are configured to increase as the distance between the target and the radar increases, and the preset value of the angle parameter is configured to decrease as the distance between the target and the radar increases.

[0018] As a further preferred embodiment of the present invention, the step of obtaining the distribution range of the cloud points of each cluster at time t in terms of distance, speed and angle is specifically as follows: K clusters are obtained by clustering at time t, and the kth cluster contains Cloud points, , (n = 1, 2, …, ), then the distribution range of the kth cluster in distance, speed and angle ( )satisfy

[0019] ;

[0020] ;

[0021] ;

[0022] in,

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] , and Take the maximum and minimum values ​​respectively.

[0029] As a further preferred embodiment of the present invention, the step of configuring the clustering area of ​​the cluster at the next moment according to the state prediction value is specifically: obtaining the speed prediction value of the target at the next moment (t+1) and the predicted value of the angle change ( ) step according to the distribution range of the k-th cluster cloud points at time t ( ), set the distance prediction range, speed prediction range and angle prediction area, where the clustering area model parameters are set as , and is the clustering region model parameter, then,

[0030] Distance area meets: ;

[0031] The speed area meets: ;

[0032] The angle area meets: ;

[0033] As a preferred embodiment of the present invention, the step of configuring the clustering algorithm parameters of the clustering area at the next moment is specifically: configuring the clustering model to obtain the model parameters of the distance, speed, angle and minimum clustering condition, then the clustering algorithm parameters of the target at the moment (t+1) can be expressed as:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] in, is the distance model parameter, is the velocity model parameter, is the angle model parameter, is the minimum clustering condition model parameter.

[0039] Compared with the prior art, the present invention, based on the existing DBSCAN clustering algorithm, firstly improves the single clustering parameter of the algorithm into a multi-parameter including distance, speed and angle; secondly, uses the cloud point range of the target at the previous moment and the state prediction value of the target at the next moment to predict the clustering area range of the target cloud point at the next moment; and, according to the state prediction value of the target at the next moment, adjusts the distance, speed and angle of the clustering algorithm and the minimum clustering condition parameters, thereby significantly improving the recognition accuracy of the millimeter-wave radar for the target in the multi-target recognition scenario and greatly reducing the situation where the target is merged and abandoned. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic diagram showing the target state (distance, speed, angle) and clustering areas;

[0041] Figure 2 is a schematic diagram showing the trend of the default parameters of the clustering algorithm in Example 1 changing with distance;

[0042] Figure 3 is a simulation result diagram, which shows the clustering effect of the existing DBSCAN clustering algorithm and the multi-target point cloud clustering algorithm of the present invention in scenario 1;

[0043] Figure 4 It is a simulation result diagram, which shows the clustering effect of using the existing DBSCAN clustering algorithm and the multi-target point cloud clustering algorithm described in the present invention in scenario 2. DETAILED DESCRIPTION

[0044] The working principle of the DBSCAN (Density-Based Spatial Clustering of Application with Noise) algorithm is to classify points that meet certain density requirements into a cluster, and judge points that are not clustered as noise points. In the millimeter-wave radar echo processing scenario of autonomous driving, the essence is to find as many similar cloud points as possible in the covered airspace. These cloud points are close enough to each other and maintain the same speed and angle. Then, these points are classified together to form a cluster.

[0045] The DBSCAN algorithm is developed based on relatively uniformly distributed big data. Therefore, the DBSCAN clustering method usually uses the same preset parameters to process all echo cloud points. When the data in the processing scene is uniform, the algorithm can remove noise points when the number of targets (number of clusters) in the scene is unknown. However, the algorithm is not suitable for scenes with multiple targets and changing target states. This is because when the state of the target changes in the covered airspace (for example, the distance of the target relative to the radar changes, or the target's own posture changes), the number and density of its echo cloud points will also change accordingly. Especially in multiple scenes, the states of multiple objects change dynamically at the same time or non-simultaneously, and the number and density of echo cloud points also show their own different characteristics. In dynamic scenes, there will be obvious differences between different clusters corresponding to multiple targets. Therefore, if the DBSCAN algorithm that uses a single parameter to cluster the entire point cloud is applied to target discrimination in dynamic scenes, the two most common situations are:

[0046] 1) The echo cloud points of the real target are judged as noise points;

[0047] 2) The cloud points belonging to different targets are grouped into one cluster (or in other words, multiple targets are mixed into one target).

[0048] Obviously, both of these situations are undesirable because they affect the accuracy of judgment.

[0049] In another attempt of the prior art, the point track aggregation method directly associates the point cloud with the track, clusters the cloud points close to the track, and then clusters the cloud points that have not been clustered. The essence of the point track aggregation scheme is to make a judgment based on the distance between a single point and the existing cluster, but it does not take the number and density of multiple points around the target into consideration. Therefore, although this method can more quickly group points that are close enough to the track and trajectory into a cluster in a dynamic scene, thereby predicting the movement trend of the target more quickly, since the clustered points have not been clustered, it is easy to misjudge noise points as part of the target point cloud, and there is also the possibility of classifying the target point cloud as noise points.

[0050] In view of the above defects of the existing algorithms, it is easy to think of using an adaptive method to solve the problem of a single parameter. For example, in the DBSCAN algorithm, the same parameters are no longer used to cluster the entire cloud points, but instead the corresponding parameters are adaptively used for discrimination according to different targets or different states of the same target. However, it is obvious that in a multi-target dynamic scene, it is impossible to predict the target type, random clutter and noise in the scene, so it is impossible to simply adapt the parameters according to the target. Even for the same target, when its state changes, its adapted parameters are not the same. Therefore, even if the method of simply obtaining the target type and state and then adapting can achieve a certain recognition accuracy, it cannot meet the requirements of response accuracy and response rate.

[0051] Let's first look at the general logic of the millimeter-wave radar clustering algorithm. The radar sends a detection signal, the signal is reflected by the surface of the target object, and the millimeter-wave radar receives the echo signal to obtain point cloud data, which is also the input of the millimeter-wave radar clustering algorithm. Define each cloud point that makes up the point cloud, and the point coordinates in the three-dimensional space coordinate system are expressed as (A, B, C). Then a value "A, B, C" in the three-phase coordinates of the point can be used to represent one of the distance, speed and angle measurement values ​​of the point. In this way, the points with distance, speed and angle values ​​in these three-dimensional spaces are gathered to form a point cloud. When the millimeter-wave radar performs discrimination, it will detect the airspace range it covers at a preset time interval, and each detection will obtain a frame of point cloud data at the current moment. The clustering algorithm module analyzes and processes several frames of point cloud data, and classifies the cloud points that meet the requirements into a cluster, so that each cluster can represent a detected target. Then, the cloud points in each cluster are synthesized. In this process, the speed, distance, and angle of all cloud points are usually averaged to obtain a coordinate in the three-dimensional space that reflects the current target state. After processing several frames of point cloud data, the change trajectory of the target cloud point in continuous time in the three-dimensional space is obtained.

[0052] While the clustering algorithm module is executing, the tracking algorithm module is also executing. The tracking algorithm module first performs track association, that is, based on the first frame of point cloud data obtained by the clustering algorithm, a new track is created for each cluster in the data, and then, based on the subsequent continuous frame point cloud data obtained by the clustering algorithm, the clusters in the subsequent frames are attributed to the track of the tracked target, and a new track is created for the newly appeared target (cluster). After the target is associated, the tracking module will predict the next state of the target based on the changes in the cloud point coordinates.

[0053] The DBSCAN algorithm is a density-based algorithm that has two main parameters:

[0054] 1) Neighborhood radius Eps, which is used to set the proximity requirement between cloud points in a cluster (the degree of proximity between two points);

[0055] 2) The minimum number of points MinPts within the neighborhood radius to become a core object. When the number of sample points in the neighborhood of a given object Eps is greater than or equal to MinPts, the given object is called a core object. Generally, for two selected cloud points p and q, if the proximity between them is less than the neighborhood radius Eps, the two selected cloud points p and q are defined as linkable.

[0056] Set the two cloud points p and q in the three-dimensional coordinate system as follows:

[0057] p=( , , )、q=( , , ),

[0058] The distance between two cloud points in three-dimensional space can be expressed according to the definition of linkable relationship and Euclidean distance formula. When judging that two cloud points p and q in space are linkable, we have , the judgment formula can also be interpreted as that according to the definition of the neighborhood radius, the two cloud points p and q are each within the neighborhood range of each other. When the neighborhood of a certain cloud point contains the number of neighboring cloud points of MinPts, or in other words, there are MinPts cloud points that are judged to be linkable with the cloud point, then the cloud point is identified as a core cloud point. After completing the identification of the core cloud points, the DBSCAN algorithm will first link all the core cloud points, and then link the core cloud points with those non-core cloud points that can be linked to them. In this way, the linked core cloud points and the non-core cloud points that can be linked to them form a cluster. Correspondingly, those cloud points that are not linked to any core cloud points are judged to be noise points. It should be noted that since the cluster will contain at least one core cloud point, MinPts should also be the minimum number of cloud points required to form a cluster. In the subsequent preferred embodiments of the present invention, MinPts is also called the minimum clustering condition, which will not be repeated in the following text.

[0059] Compared with spectral clustering and hierarchical clustering algorithms, DBSCAN algorithm has a small amount of computation and is more suitable for the discrimination scenario of unknown number of targets. However, as discussed above, since only the aforementioned neighborhood radius and the minimum number of neighboring points are used as discrimination parameters, and the values ​​of both need to be set in advance, the reliability of the algorithm is difficult to maintain when the target state and environmental state change in a complex manner.

[0060] In order to improve the reliability of the clustering algorithm in an environment with multiple targets and unknown clutter characteristics, in view of the above-mentioned defects of the existing algorithms, the preferred embodiments of the present invention recognize that: although in a multi-target scenario, the echo cloud points of the target object will present a random distribution according to its displacement, relative displacement, or state change, the distribution range of the target cloud points is always within a certain range (such as the target size), and to a certain extent, since the state (distance, speed, angle) of the target object at the next moment can be predicted according to its motion trajectory, the distribution range of the target point cloud should also be predictable. Accordingly, the preferred embodiments of the present invention attempt to divide the cloud points in the space at the next moment into regions according to the possible distribution ranges of different target point clouds, and then use the predicted values ​​of the distance, speed and angle of different targets to correct and adjust the parameters of the clustering algorithm, thereby achieving the purpose of adaptively distinguishing the cloud points in different regions according to different clustering parameters.

[0061] The distribution range of the target object's echo point cloud is determined by the target object's reflection area relative to the radar and the target object's own state changes. So, since the distribution of the target cloud points is always within a certain range, then according to the target object's current state (distance, speed, angle), the range of its echo cloud point distribution at the next moment should also be predictable. On this basis, the clustering algorithm's division of clustering areas at the next moment can also be predicted according to the tracking algorithm module's prediction of the target object's cloud point distribution range at the next moment based on the current moment's target cloud point distribution, and the clustering area of ​​the clustering algorithm at the next moment is set accordingly. It should be understood that, compared with the existing algorithms, the preferred embodiment of the present invention determines the clustering range based on the cloud point distribution range, which adds the prediction of the cloud point distribution range on the basis of the existing state prediction, thereby using the cloud point distribution range to confirm the division range of the clustering area.

[0062] Figure 1 is a schematic diagram showing a schematic diagram of the target state (distance, speed, angle) and the clustering area. It is particularly pointed out that the distance, speed and angle pointing to the target state should construct a three-dimensional coordinate system, but it is difficult to represent the cloud points and the distribution range of the cloud points in a plane after constructing the coordinate system in this way. Therefore, in this preferred embodiment, for the convenience of explanation and display, Figure 1 In the above, a target in the covered airspace is always set to maintain a constant speed, that is, the target angle state value is the horizontal axis, the target distance state value is the vertical axis, and the direction of the Y axis is consistent with the normal direction of the radar. In this way, the plane coordinate system is established, which simplifies the three-dimensional space into a two-dimensional space. However, it should be understood that the simplified coordinate system does not affect the accuracy of the implementation of this solution. It is only a schematic explanation of this solution. Obviously, in actual judgment, when the target speed state changes and the Z axis is introduced, the solution can still be implemented with the same principle.

[0063] Continue to view Figure 1 , the position of the millimeter wave radar is located at the origin of the coordinate system. After synthesizing the target point cloud using the average method, a point is formed in the spatial coordinates. Then, relative to the origin of the coordinate system, the state measurement value of the target at the current time t is 『 , , ”, the state prediction value at the next (t+1) moment is 『 , , ', where r is the straight-line distance from the target to the radar, is the speed of the target relative to the radar (set to be constant in this example), is the angle between the target and the normal direction. As shown in the figure, since the distribution of the target point cloud is always within a certain range, the clustering area can be predicted based on the state value of the target at the next moment. The clustering area is centered on the predicted state value of the target at the moment (t+1). and The three-dimensional space defined by Figure 1 In this way, at time (t+1), the clustering algorithm can Figure 1 The clustering is performed in the area defined by the bold frame. Since the target type and the distribution range of the target cloud points are determined, the corresponding parameters can be used to complete the clustering in the corresponding area according to the different types of targets.

[0064] Embodiments of the present invention will be described below with reference to the accompanying drawings. One of ordinary skill in the art will appreciate that the described embodiments may be modified in various ways without departing from the spirit and scope of the present invention. Therefore, the drawings and description are illustrative in nature and are not intended to limit the scope of the claims. In addition, in this specification, the drawings are not drawn to scale and the same reference numerals represent the same parts.

[0065] It should be noted that the expressions “first” and “second” used in the embodiments of the present invention are intended to distinguish two non-identical entities with the same name or non-identical parameters. It can be seen that “first” and “second” are only for the convenience of expression and should not be understood as limitations on the embodiments of the invention. The subsequent embodiments will not explain this one by one.

[0066] Embodiment 1

[0067] The preferred embodiment of the present invention first performs clustering processing on the radar echo signal, and the clustering process is generally consistent with the existing DBSCAN algorithm. Specifically, the current time is set to t, and the current time is the initial time. The tracked target module does not include the tracked target, and the default clustering parameters are used to classify the global cloud points. As mentioned above, the default clustering parameters in the DBSCAN algorithm only include the neighborhood radius Eps and the minimum number of points MinPts within the neighborhood radius of the core object. Since the neighborhood radius only considers the cloud point state in terms of distance, and since the preferred embodiment of the present invention attempts to consider the distribution range of the target cloud points, the preferred embodiment of the present invention attempts to introduce more consideration dimensions in the core cloud point clustering parameters.

[0068] Since the distribution range of the target cloud points is determined by their distance, speed and angle relative to the radar, the judgment parameters for the core cloud points should include speed and angle in addition to distance. Since there are no tracked targets at the initial moment, all targets are clustered using the default clustering parameters, and in subsequent moments, clustering is performed based on the clustering parameters determined at the previous moment.

[0069] In the first embodiment of the present invention, for any two cloud points in the space at the current time t, p=( , , ) and q=( , , ), defined when two cloud points simultaneously satisfy: | - 、| - 、| - When , the cloud points p and q are judged to be linkable, where the initial moment , , The default clustering parameters are used. A preset value for the distance parameter, is a preset value for the speed parameter, is the preset value of the angle parameter. The default clustering parameters also include the minimum clustering condition P, that is, the minimum number of points that constitute the core point. According to the minimum clustering condition P, all cloud points in the entire domain are judged as core points to obtain a number of core points and non-core points. Multiple clusters can be obtained by linking each core point and the non-core points that can be linked to the core point. At this time, the point cloud also includes a part of points. These points are first judged as non-core points because they do not meet the judgment criteria of the core point, and then they are judged as being unlinkable to any core point. In other words, during the clustering process, these points are not clustered into any cluster, so these points will be regarded as noise points and discarded.

[0070] Set the initial time t to get K clusters, and the kth cluster contains Cloud points, , (n = 1,2, …, ), then the distribution range of the kth cluster in distance, speed and angle ( )satisfy

[0071] ;

[0072] ;

[0073] ;

[0074] in,

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] , and Take the maximum and minimum values ​​respectively.

[0081] The clustering result at the current time t will be sent to the tracking algorithm module, which will create a new route track for the identified new target and predict the state of the target at the next time (t+1). In the first embodiment of the present invention, the state prediction value is expressed as follows based on the same distance, speed and angle considerations: ),in, is the predicted distance value of the current target at the next moment, is the predicted value of its speed at the next moment, The predicted angle value at the next moment is then used to determine the clustering area of ​​the target at the next moment based on the predicted values ​​of distance, speed and angle.

[0082] The range of the clustering area will take two aspects into consideration: one is the range of the target cloud point distribution at the current moment, and the other is the predicted state of the target at the next moment. This is because the target cloud points should always be distributed within a certain range, but changes in the target state will also affect the specific distribution of the cloud points.

[0083] In the first embodiment, according to the result output by the tracking algorithm module at the current time t, it is known that:

[0084] 1) The distribution range of the target cloud points at the next moment is still ( ) within the scope determined by

[0085] 2) The predicted value of the target state at the next moment is ( );

[0086] However, since there are three different parameters (distance, speed and angle), and the three parameters reflect the state changes of the target at the next moment from different angles, selecting a single parameter to determine the clustering area cannot be absolutely accurate. Therefore, we should try to consider the three different parameters when determining the clustering area at the next moment.

[0087] Considering that the distribution range of the target point cloud will change with the state of the target, for example, the change of the target in distance and speed depends on the speed of the target, therefore, the preferred embodiment of the present invention tries to use the speed as the increment of the speed and distance change of the clustering area when predicting the clustering area range. function; accordingly, the target point cloud angle range changes and the predicted angle changes , so for the cluster area angle increment, the angle is used Based on this, when determining the range of the clustering area, the clustering area at time (t+1) is based on the target point cloud distribution area at time t ( , and ) and the predicted value of the target speed at time (t+1) and the predicted value of the angle change ( ) to set up:

[0088] Distance area meets: ;

[0089] The speed area meets: ;

[0090] The angle area meets: .in, , and are the clustering region model parameters.

[0091] After determining the clustering area at the next moment, the clustering parameters at the next moment need to be adapted and adjusted. Looking back at the above, after improvement, the default parameters of the clustering algorithm in the preferred embodiment of the present invention include the preset value of the distance parameter , speed parameter preset value , angle parameter preset value And the minimum clustering condition P. Theoretically, the adjustment of the above clustering algorithm parameters should also be considered from the perspectives of distance, speed and angle as much as possible, just like when determining the clustering area, but obviously, this will increase the complexity of the clustering algorithm setting. The preferred embodiment of the present invention recognizes that the number and discreteness of the target cloud points are strongly correlated with the target distance. This is because the electromagnetic wave energy attenuates with increasing distance, resulting in a decrease in the signal-to-noise ratio and the signal-to-clutter ratio. Therefore, the preferred embodiment of the present invention adopts clustering parameters that vary with the target distance. Due to the increase in missed detections when the distance increases, the restrictions on the distribution of the target point cloud in terms of distance and speed should be relaxed accordingly, that is, and As the distance to the target increases, the target's angular coverage area decreases. The overall effect is It should be appropriately reduced as the target distance increases. Similarly, due to the increase in missed detections, the minimum number of linkable cloud points required for core points (P) should be reduced as the distance increases. Taking into account performance and complexity, the mathematical model of clustering parameters of the present invention uses a quadratic function with the target distance as a variable. In this way, based on the above results, at the next moment (t+1), according to the state prediction value and the model parameters, the preset parameters of the next moment (t+1) can be determined, which are:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] in, is the distance model parameter, is the velocity model parameter, is the angle model parameter, is the minimum clustering condition model parameter.

[0097] Based on the preferred embodiments of the present invention, those skilled in the art know that in actual scenarios, the determination of clustering areas and the selection of model parameters can be adjusted according to factors such as radar waveform parameters, application scenarios, and system performance. For example, in Example 1, a 77GHz band vehicle-mounted frequency-modulated continuous wave radar (frequency modulation bandwidth 165MHz, sweep time 18.8us, 128 sweeps per frame, effective range 10m-200m, maximum unambiguous speed measurement 40m / s) is selected. Similar systems can refer to the following model parameters:

[0098] At the initial moment, the default parameters of the clustering algorithm are set as:

[0099] (rice)

[0100] = 2.5 (m / s)

[0101] = 1 (degree)

[0102] = 8;

[0103] When determining the extent of the clustering area, the model parameters

[0104] ;

[0105] ;

[0106] ,

[0107] When adjusting the clustering algorithm model parameters,

[0108] = 0.0; = -0.001; = 2.0518;

[0109] = 0.0; = 0.0149; = 0.8115;

[0110] = -0.0001; = 0.0023; = 2.9014;

[0111] = 0.0002; = -0.1026; = 12.5060. Substituting the above parameters into the cluster parameter model, we get Figure 2 The graph shown, Figure 2 The schematic diagram shows the trend of the default parameters of the clustering algorithm in Example 1 changing with distance. As shown in the figure, as the target state changes, the clustering parameters , , And P will change accordingly.

[0112] The preferred embodiment of the present invention can also apply the multi-objective point cloud clustering algorithm described herein together with the existing DBSCAN algorithm to different specific scenarios to compare the performance of different algorithms in the same scenario.

[0113] Scenario 1: Two moving targets of the same size and speed travel to positions 50 meters and 80 meters away from the radar respectively;

[0114] Scenario 2: A movable target moves from the left lane at a speed of 27 m / s, overtaking a movable target in the right lane at a speed of 25 m / s. Figure 3 and Figure 4 The simulation result diagram shows the clustering effect of the existing DBSCAN clustering algorithm and the multi-target point cloud clustering algorithm of the preferred embodiment in the above two scenarios. The asterisk in the figure is used to identify the point cloud of cluster output target one, the plus sign is used to identify the point cloud of cluster output target two, and the circle is used to identify the cluster output noise point cloud. Figure 3 , due to the decrease in the number of distant target echo point clouds, the minimum number of points in the neighborhood of the DBSCAN algorithm with fixed clustering parameters cannot be met, so the cloud points of the target at 80 meters are misjudged as noise points by the DBSCAN algorithm, resulting in target loss. The preferred embodiment of the present invention uses clustering parameters that vary with target distance, so that two targets can be correctly clustered. Figure 4 , the point cloud when the two moving targets are driving in the left and right lanes respectively and the driving track trajectories are roughly parallel is as follows Figure 4 As shown in FIG. 1 , since the point clouds of the two targets are too close to each other, DBSCAN clusters the point clouds of the two targets into one group and misjudges them as the same target. However, since the algorithm of the present invention divides the clustering area according to the prediction of the target state, the two targets can be accurately clustered.

[0115] Embodiment 2

[0116] In the first embodiment of the present invention, according to the technical guidance of the present invention, the technical means adopted mainly include two aspects:

[0117] 1) In the clustering process, in addition to the distance parameter, speed and angle parameters are introduced;

[0118] 2) Based on the target point cloud distribution at the previous moment and the predicted data, predict the range of the target point cloud distribution at the next moment and the range of the clustering area.

[0119] Through the above-mentioned technical means, clustering processing in multi-target scenarios has been achieved, and the accuracy of target discrimination has been significantly improved compared to the existing DBSCAN algorithm. However, in long-term experiments and practices, the preferred embodiment of the present invention realizes that in the echo signal of the millimeter-wave radar, the point clouds between adjacent targets sometimes overlap, and this overlapping phenomenon will further lead to overlapping in the division of clustering areas in the clustering processing of the clusters corresponding to these targets. In this way, the target discrimination result occasionally merges different targets into one target, and when two targets with a large volume difference overlap, the smaller target may be discarded by the algorithm due to occlusion when the point cloud overlaps. In view of this, the preferred embodiment of the present invention proposes a complete solution to the problem of overlapping clustering areas caused by adjacent targets.

[0120] Specifically. In fact, the overlap of point clouds is inevitable, because in the global space, it is impossible for multiple moving targets to always be unobstructed relative to the radar position. When clustering, each cloud point contained in the overlapping area point cloud is considered to belong to multiple different clusters. It is conceivable that when two targets of similar volume overlap more in the field of view, since the cloud points in the overlapping area are judged to belong to two targets, when linking core points and non-core points, cloud points that do not belong to the same target will be linked, and finally the system will judge the two targets as one target. Another conceivable situation is that when two targets of different volumes overlap in the field of view, since the cloud points of the smaller target that are not overlapped are judged to be unlinkable with the cloud points of the larger target, they will be judged as noise points and abandoned. Then, the smaller target will be abandoned by the system at this time. Then one way to try is to make another judgment on the cloud points of the overlapping point cloud part so that it uniquely belongs to a certain target (cluster) to solve the overlapping problem.

[0121] Obviously, in order to accurately determine the attribution of the cloud points in the overlapping point cloud part, it should be attributed to the target or cluster to which it originally belongs as much as possible. However, in theory, the distribution of cloud points of different targets will change with the change of their state, and when there are obvious differences between the target volumes, the attribution of the points will become more and more complicated with the different locations of the overlapping parts. The second embodiment of the present invention realizes that two aspects can be considered for the judgment of points:

[0122] 1) When two targets overlap, the points can be assigned to which target (cluster) they are closer to in terms of relative distance;

[0123] 2) Considering the impact of volume difference and overlapping position on attribution judgment, the concept of compactness is introduced while judging the relative distance, that is, judging whether the target to which the cloud point belongs has a higher compactness at the overlapping position.

[0124] This is because, as the overlap occurs at different positions, the point attribution judgment not only considers which target or cluster center point it is closer to in relative distance, but also, in theory, the cloud point should be closer to the target or the core point of the cluster. For example, in the set scene, target one and target two are included, and the volume of target one is significantly larger than the volume of target two. In this way, in non-extreme scenes, no matter where the overlap of the two target point clouds occurs, the points in the overlapping area will be closer to target two in terms of the distance relative to the center point. However, when the overlapping position appears at a surface or structural feature position belonging to target one, the possibility that the point should belong to target one cannot be ruled out. However, the preferred embodiment of the present invention realizes that when judging the attribution of the cloud points in the overlapping area in combination with the above two situations, the cluster or target to which the cloud points belong should have a higher cloud point density at the location where the overlapping area occurs, that is, the target or cluster to which it is attributed should have a denser cloud point distribution at the cloud point position. In this way, through the two parameters of relative distance and cluster density, the cloud points in the overlapping part can be attributed to the target or cluster to which they originally belonged as much as possible.

[0125] In Example 2, the clustering area of ​​Target 1 is set And the clustering area of ​​target 2 There is overlap, and cloud points In the overlapping area and in the clusters of the two areas are respectively classified as areas Cluster (Assume that the cluster has Cloud points: , , ), n = 1, …, ) and region Cluster (Assume that the cluster has Cloud points: , , ), n = 1, …, ). According to the technical guide 1 of the second embodiment), it is necessary to determine the relative cluster of a certain point and The distance of the clusters needs to be calculated first to determine the relative distance. and In the preferred embodiment of the present invention, the center position of the cluster also needs to consider the three aspects of distance, speed and angle. Taking distance as an example, the center distance of the cluster is the sum of the distances of N cloud points relative to the radar origin and then taking the average value to obtain the distance of the cluster relative to the radar, and similarly obtain the relative average speed and relative average angle, so as to obtain the cluster respectively. and The coordinates of the center point are shown in the following formula.

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] .

[0132] Next, define the closeness between cloud points and clusters as the inverse of the relative distance between cloud points and cluster centers:

[0133]

[0134]

[0135] if , then it is confirmed that the cloud point p belongs to the cluster , otherwise, it is confirmed that the cloud point p belongs to the cluster In the case of overlap, when a cloud point is assigned to a cluster, it will be deleted from the overlapping cluster accordingly, so that the cloud point can be uniquely assigned. However, the extreme cases that still need to be considered are:

[0136] 1) For the cluster of removed cloud points, a core point contained in it may no longer meet the minimum point number condition MinPts as a core point;

[0137] 2) When a cluster contains only one core point, and the number of the core point and the non-core points that can be linked to the core point is exactly equal to the minimum number of core points condition MinPts, or in other words, the number of cloud points in the cluster is exactly MinPts before the cloud point is removed, it can be imagined that when the cloud point is removed, not only the core point no longer meets the core point condition, but more importantly, the cluster no longer meets the clustering condition.

[0138] In view of the above situation, in theory, when processing overlapping point cloud scenes and uniquely assigning a core point to one of the overlapping clusters, it is necessary to re-judge the core points in other overlapping clusters. However, usually, the number of cloud points contained in the point cloud and the number of non-core points that can be linked to the core points far exceed the number of points specified by the minimum clustering condition. The removal of one or several points will not cause a core point or a cluster to no longer meet its establishment conditions. Therefore, when considering the above two factors, the preferred embodiment of the present invention realizes that it is only necessary to judge whether the cloud points in the clustering area after the determination of attribution meet the minimum clustering condition. If a cluster no longer meets the minimum clustering condition after some cloud points are removed, all the remaining cloud points of the cluster are set as noise points.

[0139] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A multi-target point cloud clustering method, which is used to cluster radar echo signals, comprises the following steps: The step of constructing a spatial coordinate system based on the distance, speed and angle of the target relative to the radar, and aggregating the echo signal of the target on the radar at the current time t into a point cloud in the spatial coordinate system; The clustering algorithm parameters are configured according to the distance, speed and angle to determine whether each cloud point constituting the point cloud is a core cloud point, so as to obtain at least one core cloud point and a plurality of non-core cloud points; A step of linking each of the core cloud points, and then linking each of the core cloud points and the non-core cloud points that can be linked thereto, to obtain at least one cluster; The steps of respectively obtaining the distribution range of the cloud points of each cluster at the current time t in terms of distance, speed and angle; The step of obtaining the state prediction value of the target at the next moment (t+1), and configuring the clustering area of ​​the cluster at the next moment and the clustering algorithm parameters of the clustering area at the next moment according to the state prediction value, specifically configuring the change in distance and speed of the clustering area of ​​the cluster at the next moment according to the speed change of the target, and configuring the change in angle of the clustering area according to the angle change of the target.

2. The multi-target point cloud clustering method according to claim 1, wherein: It also includes the steps of uniquely classifying the cloud points, specifically: When the echo point clouds of at least two clusters overlap, a step of obtaining the center point of each overlapping cluster; For any cloud point in the overlapping area, obtain the distance between it and the center point of each cluster; The cloud point is assigned to the cluster with the shortest distance and the largest density.

3. The multi-target point cloud clustering method according to claim 2, wherein: After the cloud points are divided into the nearest cluster, the method further includes: determining whether the remaining points in other overlapping clusters meet the minimum clustering condition. If not, the cloud points in the cluster are determined to be noise points.

4. The multi-target point cloud clustering method according to claim 1, wherein: The steps to configure clustering algorithm parameters according to distance, speed and angle are as follows: To set the distance parameter preset value, speed parameter preset value, and angle parameter preset value: When any two points in the space are judged to be linkable in terms of distance, speed and angle, the two points are deemed to be linkable, otherwise the two points are deemed to be unlinkable; The step of configuring according to the predicted value of the state of the cluster at the next moment to determine the preset value of the distance parameter, the preset value of the speed parameter, the preset value of the angle parameter, and the minimum clustering condition at the next moment.

5. The multi-target point cloud clustering method according to claim 4, wherein: The specific steps to determine whether two points can be linked are: For any two cloud points in space, p = (r i ,v i ,θ i ) and q=(r j ,v j ,θ j ), where r is the distance, v is the distance, and θ is the distance if both of them are satisfied: |r i -r j |<E r |v i -v j | <E v |θ i -θ j | <E θ ; Then the cloud points p and q are judged to be linkable, where E r is the preset value of the distance parameter, E v is the preset value of the speed parameter, E θ A preset value for the angle parameter.

6. The multi-target point cloud clustering method according to claim 4 or 5, wherein: In the step of setting the distance parameter preset value, the speed parameter preset value and the angle parameter preset value, The preset value of the distance parameter and the preset value of the speed parameter are configured to increase as the distance between the target and the radar increases, and the preset value of the angle parameter is configured to decrease as the distance between the target and the radar increases.

7. The multi-target point cloud clustering method according to claim 6, wherein: The specific steps for obtaining the distribution range of cloud points of each cluster in terms of distance, speed and angle at time t are: Suppose that at time t, K clusters are obtained by clustering, and the kth cluster contains N k Cloud points, Then the distribution range of the kth cluster in distance, speed and angle is satisfy in, max() and min() are used to get the maximum and minimum values ​​respectively.

8. The multi-target cloud point clustering method according to claim 7, wherein: The step of configuring the clustering area of ​​the cluster at the next moment according to the state prediction value is specifically as follows: Get the target's predicted speed value v at the next moment (t+1) (k) (t+1) and the predicted value of the angle change (θ (k) (t+1)-θ (k) Steps (t)) According to the distribution range of the k-th cluster cloud points at time t Set the distance prediction range, speed prediction range and angle prediction area. Among them, the clustering region model parameters are set as α, β and γ respectively, then, Distance area meets: The speed area meets: The angle area meets:

9. The multi-target cloud point clustering method according to claim 8, wherein: The steps of configuring the clustering algorithm parameters of the clustering region at the next moment are specifically as follows: The clustering model is configured to obtain the model parameters of distance, speed, angle and minimum clustering condition. Then the clustering algorithm parameters of the target at time (t+1) can be expressed as: P (k) (t+1)=d0+d1r (k) (t+1)+d2(r (k) (t+1)) 2 ; Among them, a0, a1, a2 are distance model parameters, b0, b1, b2 are velocity model parameters, c0, c1, c2 are angle model parameters, P (k) is the minimum clustering condition at time (t+1), where d0, d1, and d2 are the minimum clustering condition model parameters.

Images