Point cloud data processing and target classification method based on millimeter wave radar
By introducing Mahalanobis distance and DBSCAN clustering algorithm combined with target velocity information, the problem of inaccurate target clustering in traditional millimeter-wave radar point cloud data processing is solved, and high-precision target classification and real-time trajectory output are achieved in complex traffic environments.
Patent Information
- Application Number
- CN202411041609.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-19
AI Technical Summary
Traditional millimeter-wave radar point cloud data processing algorithms tend to cluster multiple targets into one cluster or a large target into multiple clusters in complex traffic environments, affecting target tracking quality and failing to effectively utilize target speed information.
By combining Mahalanobis distance and DBSCAN clustering algorithm, point cloud data preprocessing and clustering are performed through Cartesian coordinate system transformation and target velocity information. Combined with Kalman filter optimization, target type identification and real-time trajectory output are achieved.
It improves the accuracy and real-time performance of target classification, and can accurately output target trajectories in complex traffic environments, possessing highly intelligent, comprehensive, and accurate characteristics.
Smart Images

Figure CN118981010A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of traffic radar data processing, and in particular to a method for processing point cloud data and classifying targets based on millimeter wave radar. Background Art
[0002] The increasingly complex road traffic environment has prompted my country to develop intelligent traffic management systems towards intelligence, comprehensiveness, accuracy and real-time. To achieve this requirement, it is necessary to rely on reliable, real-time and accurate detection data. The technical advantages of millimeter-wave radar can have extremely important applications in the field of intelligent transportation.
[0003] At present, the processing of millimeter-wave radar in the field of intelligent transportation is mainly divided into two steps: point clustering and target tracking. Among them, the output of point clustering is used as the input of target tracking. Therefore, the quality of point clustering directly affects the quality of target tracking. Because, in a complex traffic environment, the point cloud data of millimeter-wave radar is characterized by: small targets and distant targets have a small number of points or even only one point, large targets (such as large trucks) will have many point cloud data but large dispersion, and the point cloud data of the same target is denser when it is close to the radar and sparser when it is far away from the radar. The traditional clustering algorithm divides the point cloud based on density. It requires that the number of points contained in a certain Euclidean distance range is not less than the set number threshold. By classifying closely connected point clouds into one category, a target cluster is obtained. However, the traditional clustering algorithm relies too much on Euclidean distance information and does not use the target speed information measured by the millimeter-wave radar. Therefore, when multiple targets are very close, the point clouds of multiple targets will be clustered into one target cluster or the point cloud of a large target will be clustered into multiple target clusters, thereby affecting the quality of subsequent target tracking. Summary of the invention
[0004] The purpose of the present invention is to provide a method for processing millimeter wave radar point cloud data and classifying targets, which has the advantage of high accuracy.
[0005] To achieve the above-mentioned object, the present invention provides a method for processing and classifying target points based on millimeter wave radar point cloud data, the method comprising: step S1, obtaining target movement data of a total of M targets in a frame through a millimeter wave radar, t=1, 2...M, and labeling them one by one; the target movement data of the M targets have N radar target points, i=1, 2..., N, and labeling them one by one, N≥M; obtaining target movement data corresponding to the radar target points; step S2, obtaining a point cloud data set P={P1...P i |i=1,2…,N}; and preprocess the point cloud dataset P; where P irepresents the point cloud data of the i-th radar target point; step S3, introduces the Mahalanobis distance and uses the DBSCAN clustering algorithm to perform point cloud clustering processing on the point cloud data in the preprocessed point cloud data set P one by one to obtain the point cloud cluster S of each target t ; Step S4, point cloud clusters S1~S M Transform to obtain intuitive point cloud clusters S1(direct)~S M (direct), and discriminate the target type to obtain the target classification result in the current frame; step S5, calculate the target classification results in other consecutive frames with the frame as the center frame, and perform linkage analysis to output the real-time trajectory.
[0006] Preferably, the step S2 comprises: step S21, performing algorithm operation on the target movement data in the current frame in sequence; obtaining a point cloud data set P = {P1 ... P i |i=1,2…,N}; Step S22, the point cloud data P in the point cloud data set P i Carte-Cartesian coordinate system transformation is performed one by one to obtain the point cloud dataset P in the Cartesian coordinate system. Cartesian ={P1′……P i ′|i=1,2…,N}; where P i ′=(x i ,y i ,v i ,σ i ), P i ′ represents the point cloud data of the i-th radar target point in the Cartesian coordinate system, x i is the lateral distance of the ith radar target point in the Cartesian coordinate system, y i is the longitudinal distance of the ith radar target point in the Cartesian coordinate system, v i is the velocity of the ith radar target point in the Cartesian coordinate system, σ i is the motion angle of the i-th radar target point in the Cartesian coordinate system; Step S23: with the velocity v of the i-th radar target point in the Cartesian coordinate system i To determine the object, classify the point cloud dataset P in the Cartesian coordinate system Cartesian ={P′1……P′ i |i=1,2…,N}, and obtain the approaching radar target point cloud set P containing A radar target points Cartesian (close) = {P′ 1- ...P' k- |k=1,2...,A}; where P k ' - =(x i ,y i ,v i ,σi |v i <0); and a cloud of B radar target points far from the radar target point cloud P Cartesian (far) = {P1′ + ……P q ′ + |q = 1, 2…, B}; where P q ′ + = (x i , y i , v i , σ i |v i > 0).
[0007] Preferably, the step S3 includes: step S31, calculating one by one the Mahalanobis distance D Cartesian (close) = {P1′ - ……P k ′ - |k = 1, 2…, A} between different two points in the approaching radar target point cloud, and then obtaining the approaching radar target point cloud Mahalanobis distance set D(close) = {D kk’ ……D 12 丨k = 1, 2…, A, k′ = 1, 2…, A, k ≠ k’}; where, kk’ Σ Σ kk’D- represents the covariance between the k-th radar target point and the k’-th radar target point in the approaching radar target point cloud; step S32, calculating one by one the Mahalanobis distance D Cartwsian (far) = {P′ 1+ ……P′ q+ |q = 1, 2…, B} between different two radar target points in the far radar target point cloud, and then obtaining the far radar target point cloud Mahalanobis distance set D(far) = {D qq’ ……D 12 ……D qq’ 丨q = 1, 2…, B, q′ = 1, 2…, B, q ≠ q’}; Σ kk’D- represents the covariance between the q-th radar target point and the q’-th radar target point in the far radar target point cloud; step S33, setting the Mahalanobis distance threshold ε m and the spatial range radar point number threshold MinP, and marking the core points and noise points of the radar target points in the approaching radar target point cloud P Cartesian (close) and the far radar target point cloud P Cartesian (far) one by one; by judging any radar target point within ε mThe number of remaining radar target points in the range is compared with the spatial range radar point number threshold MinP. If it is greater than MinP, the point is marked as a core point; if it is less than MinP, the point is marked as a noise point. Step S34, the marked core points and noise points are classified according to the corresponding targets to form M point cloud clusters S1~S M .
[0008] Preferably, the step S4 comprises: step S41, calculating the target motion heading angle corresponding to each point cloud cluster one by one Among them, x center Represents the lateral distance of the radar target point as the core point in the point cloud cluster in the Cartesian coordinate system, y i is the longitudinal distance of the radar target point as the core point in the point cloud cluster in the Cartesian coordinate system; and the point cloud data of each radar target point in the point cloud cluster is calculated according to the heading angle Rotate clockwise to turn the target's motion direction to the horizontal direction, and obtain the corresponding intuitive point cloud clusters S1 (direct) ~ S M (direct); Step S42, through the intuitive point cloud clusters S1 (direct) ~ S M (direct) Calculate the length l of each point cloud cluster t , width w t ; where w t =x max -x min , l t =y max -y min ;x max Indicates the maximum lateral distance among all radar target points in the point cloud cluster; y max Indicates the maximum longitudinal distance among all radar target points in the point cloud cluster; x min Indicates the minimum lateral distance among all radar target points in the point cloud cluster; y min represents the minimum longitudinal distance among all radar target points in the point cloud cluster; Step S43: according to the length l of each point cloud cluster t , width w t and the number of radar target points N t ', get the target classification result outcome(t) in the current frame.
[0009] Preferably, the step S5 comprises: step S51, taking the current frame as the center frame, selecting a total of U frames of data of consecutive frames, u=1, 2...U, and obtaining all point cloud clusters under the 1st frame to the Uth frame one by one according to steps S1 to S4, and the length l corresponding to each point cloud cluster ur , width w ur and the number of radar target points N ur' and the corresponding target classification result outcome(ur); r represents the number of the point cloud clusters in different frames; step S52, the point cloud clusters are associated from the first frame to the Uth frame, and a total of E preliminary motion trajectories Trail(1)~Trail(E), e=1, 2...E; step S53, according to the length l corresponding to Trail(e) ur , width w ur And the number of radar target points Nur', calculate the mean length of the point cloud cluster under the preliminary motion trajectory one by one Width Mean And the mean number of radar target points Step S54: calculate the length l of all point cloud clusters under the e-th preliminary motion trajectory one by one pr , width w pr and the number of radar target points Npr' and the mean length of the e-th preliminary motion trajectory Width Mean And the mean number of radar target points Compare and verify; optimize the frames where the point cloud clusters with large differences are located, and output the motion trajectory Trail outcome (1)~Trail outcome (E).
[0010] Preferably, the optimization process in step S54 includes deleting frames where point cloud clusters that are too large are located.
[0011] Preferably, the optimization processing in step S54 includes, for the frame where the point cloud cluster with excessively large differences is located, performing Kalman filter function calculation on the point cloud data of the target in the current frame and the motion trajectory of the target in the previous frame to obtain the motion trajectory of the target in the corresponding frame.
[0012] Preferably, in step S21, the algorithm operation includes: executing distance dimension Fourier transform, velocity dimension Fourier transform, constant false alarm detection, Doppler phase compensation, and DOA angle estimation.
[0013] Preferably, in step S53, the mean value calculation includes calculating the corresponding length l ur , width w ur The number of radar target points Nur' is calculated by adding and averaging.
[0014] Preferably, the step S43 includes: t Greater than 16 meters, width w t If the length is greater than 3 meters and the number of radar target points Nt' is greater than 15, it is judged as a large vehicle type; t Greater than 16 meters, width w tThose with a distance greater than 3 meters and a number of radar target points Nt' greater than 8 and less than 15 are judged as medium-sized vehicles; the rest are small vehicles.
[0015] In summary, compared with the prior art, the method for processing millimeter wave radar point cloud data and classifying targets provided by the present invention has the following beneficial effects:
[0016] First, the present invention is intelligent, comprehensive, accurate and real-time in application and can output the target trajectory with high precision.
[0017] Second, the present invention uses reliable, real-time and accurate detection data and speed to make target judgments with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 Schematic diagram of the use of millimeter-wave radar in road traffic.
[0019] Figure 2 It is the overall flow chart of the present invention.
[0020] Figure 3 This is a marking flow chart in step S3 of the present invention.
[0021] Figure 4 This is a diagram of a traffic arrival scene under real conditions in an embodiment of the present invention.
[0022] Figure 5 This is a radar point cloud data diagram according to an embodiment of the present invention.
[0023] Figure 6 This is a comparison diagram of the clustering results of the present invention and traditional DBSCAN clustering.
[0024] Figure 7 Schematic diagram of constructing virtual measurement point cloud data. DETAILED DESCRIPTION
[0025] The following will be combined with the attached embodiment of the present invention Figure 1 ~Attached Figure 7 , the technical solutions, structural features, objectives achieved and effects in the embodiments of the present invention are described in detail.
[0026] It should be noted that the drawings are in a very simplified form and use non-precise proportions. They are only used to conveniently and clearly assist in explaining the embodiments of the present invention, and are not used to limit the conditions for the implementation of the present invention. Therefore, they have no substantive technical significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical content disclosed by the present invention.
[0027] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only the elements explicitly listed, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0028] like Figure 1 and Figure 2 As shown, the present invention discloses a method for processing millimeter wave radar point cloud data and classifying targets. The method obtains target movement data of a total of M targets in a certain frame within the capture range of the millimeter wave radar through a millimeter wave radar, t=1, 2...M, and labels them one by one; the targets here include but are not limited to large vehicles, small vehicles and other vehicles belonging to the field of road traffic.
[0029] The target must have a relatively large volume, so the target movement data of a single target may be composed of one or more radar target points; therefore, there are N radar target points in the target movement data generated by M targets, i = 1, 2...N, and they are numbered one by one, and N ≥ M, and they are associated with the aforementioned M, so that the data can be directly matched to the corresponding target later.
[0030] The radar target point is measured, and the relevant parameters of the measured radar target point are counted as point cloud data, wherein the point cloud data includes relevant parameters such as moving speed and target moving direction; and then the point cloud data set P is obtained as follows: i |i=1,2…,N};P i Represents the point cloud data of the i-th radar target point; analyzes the point cloud data to cluster and classify the target and conducts linkage analysis with other frames to obtain the target trajectory.
[0031] like Figure 2 As shown, the method for processing and classifying targets based on millimeter wave radar point cloud data includes: step S1, obtaining target movement data of a total of M targets in one frame through a millimeter wave radar, t=1, 2...M, and labeling them one by one; the target movement data of the M targets exist in N radar target points, i=1, 2..., N, and labeling them one by one, N≥M; obtaining M target movement data corresponding to the N radar target points;
[0032] As mentioned above, the target movement data obtained by the millimeter-wave radar has a certain volume, so one target may correspond to multiple radar target points. In addition, the purpose of the labeling is that the subsequent calculations are all centered around the radar target points, so that the calculation results in the subsequent calculations can be linked to the corresponding target through the radar target point.
[0033] Step S2, obtaining a point cloud data set P = {P1...P i |i=1,2…,N}; and preprocess the point cloud dataset P;
[0034] Specifically, the target movement data of the M targets corresponding to the N existing radar target points are measured one by one to obtain the point cloud data of a single radar target point, and the point cloud data set P is sorted out, and then the point cloud data set P is preprocessed to obtain the data required for subsequent calculations.
[0035] Step S3, introduce the Mahalanobis distance and use the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm to perform point cloud clustering processing on the point cloud data in the preprocessed point cloud data set P one by one to obtain the point cloud cluster S of each target t ;
[0036] Because of the numbered corresponding processing mentioned above, N point cloud data can be directly processed here and then linked to the corresponding target.
[0037] For example, the point cloud cluster S1 of the first target is obtained as follows:
[0038] The target movement data of the first target is obtained. At this time, M=1. Assuming that there are 3 corresponding radar target points in the target movement data of the first target, point cloud data P1, P2, and P3 are obtained at this time. The point cloud data belong to the target movement data of the first target. Then, point cloud clustering processing is performed on the point cloud data P1, P2, and P3. Because the numbers correspond, the aforementioned point cloud data can be counted as the point cloud cluster S1 of the first target according to the order.
[0039] Step S4: point cloud clusters S1 to S M Transform to obtain intuitive point cloud clusters S1(direct)~S M (direct), and identify the target type to obtain the target classification result in the current frame;
[0040] Specifically, step S4 is to perform point cloud clusters S1 to S MThe heading angle (target moving direction) is transformed into the length in the same direction in all intuitive point cloud clusters, thereby achieving accurate measurement, as described later.
[0041] Step S5, calculating the target classification results of other consecutive frames with the frame as the center frame, and performing linkage analysis to output the real-time trajectory.
[0042] Specifically, after obtaining the classification result of a single frame, it is necessary to realize the output of the real-time trajectory by linking with other continuous frames; the step S5 is to obtain a continuous target classification result by applying the steps S1 to S4 to other continuous frames with the frame as the center frame, and analyze the corresponding results to obtain the real-time trajectory.
[0043] The specific implementation manner is described in detail below, and the step S2 includes:
[0044] Step S21, sequentially perform distance dimension Fourier transform, velocity dimension Fourier transform, constant false alarm detection, Doppler phase compensation, and DOA (Direction Of Arrival) angle estimation on the target movement data in the current frame; obtain a point cloud data set P = {P1...P i |i=1,2...,N}, where P i represents the point cloud data of the i-th radar target point, and N represents the total number of radar target points in the point cloud dataset P.
[0045] The target movement data includes, but is not limited to, parameters such as the current coordinate position of the target, the moving direction of the target, and the physical size of the target.
[0046] Regarding the distance-dimensional Fourier transform and the velocity-dimensional Fourier transform, both are implemented based on the Fourier transform, and their function is to automatically search for the acquired target movement data, adjust the phase, and facilitate the selection of corresponding parameters in subsequent target operations.
[0047] Regarding constant false alarm detection, this technology uses the statistical characteristics of the signal and the statistical characteristics of the noise to determine whether the target signal exists in the interfered observation. Constant false alarm detection can ensure the accuracy of the subsequent measurement data of the radar target point.
[0048] Regarding Doppler phase compensation, the Doppler effect refers to the change in the frequency of the wave emitted by the object that the observer will feel when the observer and the object are in relative motion. In the fields of radar, sonar, wireless communication, etc., it will cause the frequency shift of the signal, thus affecting the analysis and processing of the signal. Doppler phase compensation is used to process the received target motion data to eliminate the phase shift caused by the Doppler effect, so that the target motion data can be accurately analyzed and processed.
[0049] Regarding DOA angle estimation, the principle is to use radar to estimate the direction of signal transmission from the target.
[0050] Through the above steps, we can obtain the point cloud data set P = {P1...P i |i=1,2…,N}, but the point cloud dataset P needs to be preprocessed before it can be used in subsequent steps.
[0051] Step S22: point cloud data P in point cloud data set P i Carte-Cartesian coordinate system transformation is performed one by one to obtain the point cloud dataset P in the Cartesian coordinate system. Cartesian ={P′1……P′ i |i=1,2…,N}; where P′ i =(x i ,y i ,v i ,σ i ), P′ i Represents the point cloud data of the i-th radar target point in the Cartesian coordinate system, x i is the lateral distance of the ith radar target point in the Cartesian coordinate system, y i is the longitudinal distance of the ith radar target point in the Cartesian coordinate system, v i is the velocity of the ith radar target point in the Cartesian coordinate system, σ i is the motion angle of the i-th radar target point in the Cartesian coordinate system.
[0052] How to get x i With y i , here we propose a method, x i =r i cosσ i ,y i =r i sinσ i , where r i is the distance of the radar target point relative to the radar; about σ i With v i The measurement method can be to lock the first frame first, and then use the direction of the radar target point in the current frame relative to the first frame as the motion angle σ i , the moving distance of the radar target point in the current frame relative to the first frame is divided by the elapsed time to obtain v i , the v obtained here is i is an average speed; of course, the first frame can also be replaced by the previous frame for the above calculation, in which case the data obtained is equivalent to an instantaneous value; or the frame 5 seconds ago can be selected for the above calculation; of course, the movement angle σ can also be determined by the direction of the road i, and through the use of auxiliary measurement tools such as electronic eyes set at the entrance, the speed is matched according to the vehicle model / license plate to obtain v i ; The method here is not limited, the purpose is to obtain relevant parameters.
[0053] Step S23, using the velocity v of the i-th radar target point in the Cartesian coordinate system i To determine the object, classify the point cloud dataset P in the Cartesian coordinate system Cartesian ={P′1……P′ i |i=1,2…,N}, and obtain the approaching radar target point cloud set P containing A radar target points Cartesian (close) = {P′ 1- ...P' k- |k=1,2…,A}; where P′ k- =(x i ,y i ,v i ,σ i |v i <0); and a cloud set P of target points far from the radar including B radar target points Cartesian (far) = {P′ 1+ ...P' q+ |q=1,2...,B}; where P′ q+ =(x i ,y i ,v i ,σ i |v i >0);
[0054] Among them, v i is the velocity of the ith radar target point in the Cartesian coordinate system, and the velocity v of the ith radar target point in the Cartesian coordinate system is determined by i The positive and negative values of the radar target point can be obtained, and then the moving direction of the radar target point relative to the radar can be obtained, thereby obtaining the approaching radar target point cloud set P Cartesian (close) = {P1′ - ...P k ' - |k=1,2...,A} and the point cloud set P of the target far from the radar Cartesian (far) = {P1′ + ...P q ' + |q=1,2…,B}.
[0055] It should be explained here that step S23 is only for the point cloud data set P in the Cartesian coordinate system. Cartesian = {P iClassify them according to {i = 1, 2..., N} without any parameter operations or changes, and A + B = N; thus approaching the radar target point cloud set P Cartesian (close) = {P1′ - ……P k ′ - |k = 1, 2..., A} and the radar target point cloud set P Cartesian (far) = {P′ 1+ ……P′ q+ |q = 1, 2..., B} The content of a single radar target point is the same as that of the corresponding single radar target point in the point cloud data set P Cartesian = {P′1……P′ i |i = 1, 2..., N}.
[0056] As Figure 3 shown, the step S3 includes:
[0057] Step S31, calculate the Mahalanobis distance D Cartesian (close) = {P′ 1- ……P′ k- |k = 1, 2..., A} between different two points, and then obtain the Mahalanobis distance set D(close) = {D kk’ ……D 12 ……D kk’ 丨k = 1, 2..., A, k′ = 1, 2..., A, k ≠ k’};
[0058]
[0059] Among them, ∑ kk’D- represents the covariance between the kth radar target point and the k′th radar target point in the approaching radar target point cloud set; among them, the covariance is introduced to reduce the overall error of two variables and increase the stability of the calculation. In addition, the present invention creatively uses the Mahalanobis distance instead of the Euclidean distance because the greatest advantage of the Mahalanobis distance is that it is not affected by the dimension between different dimensions, thereby realizing the accuracy of distance measurement.
[0060] Step S32, calculate the Mahalanobis distance D Cartesian (far) = {P′ 1+ ……P′ q+ |q = 1, 2..., B} between different two radar target points, and then obtain the Mahalanobis distance set D(far) = {D qq’ ……D 12 ……D qq’|q = 1, 2…, B, q′ = 1, 2…, B, q ≠ q’};
[0061]
[0062] Among them, ∑ kk’D- represents the covariance between the q-th radar target point and the q'-th radar target point in the set of radar target points far from the radar, and its function is as described above.
[0063] It can be seen that the principles of step S31 and step S32 are basically the same, and both need to obtain the Mahalanobis distance between different two points in the set; in addition, since the relevant parameters of the radar target points in the approaching radar target point cloud set P Cartesian (close) and the set of radar target points far from the radar P Cartesian (far) can be found one-to-one through the numbers, the situation of parameter confusion can be avoided.
[0064] At the same time, the order of step S31 and step S32 can be interchanged. The purpose of using separate sequential calculations in this method is to make the calculation clearer and prevent errors.
[0065] Step S33, set the Mahalanobis distance threshold ε m and the threshold MinP for the number of radar points in the spatial range, and use ε m and MinP as the annotation parameters to mark the core points and noise points of the radar target points for each radar target point in the approaching radar target point cloud set P Cartesian (close) and the set of radar target points far from the radar P Cartesian (far);
[0066] Among them, the design principle of the above two parameters is that since the Mahalanobis distance D kk’ between different two radar target points has been obtained in steps S31 and S32, the Mahalanobis distance threshold ε m and the threshold MinP for the number of radar points in the spatial range are manually set, and then by judging the number of the remaining radar target points existing within the range of ε m for any radar target point and comparing it with the threshold MinP for the number of radar points in the spatial range, if it is greater than MinP, the point is marked as a core point; if it is less than MinP, the point is marked as a noise point;
[0067] Regarding how to manually determine the Mahalanobis distance threshold ε m and the threshold MinP for the number of radar points in the spatial range, it is necessary to jointly design according to the nature of the radar actually used, such as signal strength, and the application scenario, such as spatial size, etc., so as to obtain appropriate thresholds.
[0068] A specific determination method is given below. Since the Mahalanobis distance set D(far) of the far-away radar target point cloud and the Mahalanobis distance set D(close) of the close-to-radar target point cloud have been determined above, taking the Mahalanobis distance set D(far) of the far-away radar target point cloud as an example, after determining the Mahalanobis distance threshold ε m After the threshold MinP of the number of radar points in the spatial range is reached, a core point can be determined. At this time, the core point is selected to form a radius of ε m The set domain is then judged one by one for other radar target points in the set domain, and the points that meet the conditions are also marked as the same core points. Then, the newly added core points are judged recursively until all the core points are judged.
[0069] Step S34: classify the marked core points and noise points according to the corresponding targets to form M point cloud clusters S1 to S M ;
[0070] Specifically, this step is essentially to classify the radar target points belonging to the same target. Since there are M targets, a total of M point cloud clusters will be formed, S1~S M At this point, a point cloud cluster includes all point cloud data corresponding to the radar target point, and each point cloud data has been identified as a core point or a noise point through the above steps, forming a preliminary classification result.
[0071] Furthermore, the step S4 comprises:
[0072] Step S41, calculate the target motion heading angle corresponding to each point cloud cluster one by one Among them, x center Represents the lateral distance of the radar target point as the core point in the point cloud cluster in the Cartesian coordinate system, y center is the longitudinal distance of the radar target point as the core point in the point cloud cluster in the Cartesian coordinate system; and the point cloud data of each radar target point in the point cloud cluster is calculated according to the heading angle Rotate clockwise to turn the target's motion direction to the horizontal direction, and obtain the corresponding intuitive point cloud clusters S1 (direct) ~ S M (direct);
[0073] This step calculates the target motion heading angle corresponding to the point cloud cluster. The purpose is to calculate By rotating clockwise, the target's moving direction is turned to the horizontal direction, so that each radar target point can be at the same size in calculating parameters such as target length, avoiding differences in calculation due to different moving directions; Figure 7 As shown, the virtual measurement point cloud data of the point cloud cluster is obtained.
[0074] Step S42, through the intuitive point cloud clusters S1 (direct) ~ S M (direct) Calculate the length l of each point cloud cluster t 、Width w t ;
[0075] where w t =x max -x min , l t =y max -y min ;
[0076] x max Indicates the maximum lateral distance among all radar target points in the point cloud cluster;
[0077] y max Indicates the maximum longitudinal distance among all radar target points in the point cloud cluster;
[0078] x min Indicates the minimum lateral distance among all radar target points in the point cloud cluster;
[0079] y min Indicates the minimum longitudinal distance among all radar target points in the point cloud cluster.
[0080] Step S43: According to the length l of each point cloud cluster t , width w t and the number of radar target points N t ', get the target classification result outcome(t) in the current frame;
[0081] Specifically, the classification results are generally set as large vehicles (buses, trucks and other large motor vehicles), medium vehicles (small trucks, SUVs, vans, etc.), and small vehicles (sedans, motorcycles, etc.);
[0082] According to the length l of each point cloud cluster t , width w t and the number of radar target points Nt' for classification results. In the embodiment of the present invention, the length l t Greater than 16 meters, width w t If the length is greater than 3 meters and the number of radar target points Nt' is greater than 15, it is judged as a large vehicle type; t Greater than 16 meters, width w t Those with a distance greater than 3 meters and a number of radar target points Nt' greater than 8 and less than 15 are judged as medium-sized vehicles; the rest are small vehicles.
[0083] At this point, step S43 completes the analysis of all radar signal points in the current frame, and step S5 further applies this method to the remaining frames, thereby combining and analyzing consecutive frames to obtain real-time trajectories.
[0084] The step S5 comprises:
[0085] Step S51, taking the current frame as the center frame, select the previous and next consecutive frames, a total of U frames, u = 1, 2...U, and obtain all point cloud clusters under the 1st frame to the Uth frame one by one according to steps S1 to S4, and the length l corresponding to each point cloud cluster ur , width w ur and the number of radar target points N ur 'And the corresponding target classification result outcome(ur); r represents the number of the point cloud cluster in different frames;
[0086] Among them, r is only used as a number here. Because new targets appear and old targets disappear, the number of point cloud clusters (targets) in each frame is unknown. For example, the relevant parameters of the first point cloud cluster in the first frame are expressed as follows: The length l corresponding to the first point cloud cluster 11 , width w 11 and the number of radar target points N 11 ' and the corresponding target classification result outcome(11); For the case where ur may contain multiple digits that may cause confusion in recognition, separators such as "," and "-" can be added to ur to prevent misjudgment.
[0087] Regarding the calculation of the center frame, since the center frame has been calculated, only the label needs to be changed. In addition, the meaning of the center frame does not mean that it must be in the exact center of the continuous frames, because the purpose of calculating the frames (current frame) in steps S1 to S4 is to select a frame with better data as the basis for analysis. Therefore, there is no need to require that the current frame must be in the exact center of the continuous frames.
[0088] Step S52, associating the point cloud clusters from the first frame to the Uth frame to obtain a total of E preliminary motion trajectories Trail(1) to Trail(E), where e=1, 2, ..., E;
[0089] Because the movement of the same target is continuous, the point cloud clusters within the threshold range can be attributed to the same target by setting the point cloud cluster difference threshold between consecutive frames. At this time, the preliminary motion trajectory of the same target can be obtained. However, due to the existence of errors, the preliminary motion trajectory needs to be further optimized later. A single preliminary motion trajectory Trail(e) includes the corresponding point cloud cluster and the length l of the point cloud cluster. ur , width w ur and the number of radar target points Nur 'And the corresponding target classification result outcome(ur).
[0090] In the actual calculation process, the target may suddenly disappear or appear, so it is uncertain how many targets there are. Therefore, Trail (1) to Trail (E) here should include all targets that have appeared in consecutive frames; among which Trail (e) represents the initial motion trajectory of the same target.
[0091] Step S53: according to the length l corresponding to Trail(e) ur , width w ur And the number of radar target points Nur', calculate the mean length of the point cloud cluster under the preliminary motion trajectory one by one Width Mean And the mean number of radar target points
[0092] For example, for the first preliminary motion trajectory, according to the average length of the point cloud cluster from the first frame to the Uth frame under the first preliminary motion trajectory, Width Mean And the mean number of radar target points It can be obtained by adding the average.
[0093] Step S54: calculate the length l of all point cloud clusters under the e-th preliminary motion trajectory one by one pr , width w pr and the number of radar target points Npr' and the mean length of the e-th preliminary motion trajectory Width Mean And the mean number of radar target points Compare and verify; optimize the frames where the point cloud clusters with large differences are located, and output the motion trajectory Trail outcome (1)~Trail outcome (E).
[0094] The difference can be determined by setting a threshold. For example, for the first preliminary motion trajectory, the mean length of the point cloud cluster from the first frame to the Uth frame is obtained. Width Mean And the mean number of radar target points In the case of , it is combined with the length l of all point cloud clusters (from the first frame to the U frame, if it disappears in the middle, it reaches the last frame) of the first preliminary motion trajectory pr , width w pr Compare them one by one with the number of radar target points Npr'.
[0095] Generally speaking, without affecting the overall output result, the processing described in step S54 is generally to delete the corresponding frame. In a preferred embodiment, a trajectory prediction method can also be added to step S54.
[0096] The processing of the frames corresponding to the point cloud clusters with excessive differences in step S54 also includes:
[0097] If there is a point cloud cluster with too large a difference, the Kalman filter function is calculated on the point cloud data of the current frame of the target and the motion trajectory of the target in the previous frame to obtain the motion trajectory of the target in the corresponding frame.
[0098] The Kalman filter is an algorithm that uses a linear system state equation to optimally estimate the system state. By inputting the point cloud data of the target in the current frame and the motion trajectory of the target in the previous frame, the Kalman filter is used to optimally estimate the motion trajectory.
[0099] like Figure 4 , Figure 5 In the specific embodiment shown in FIG, the figure is a real traffic scene. According to the distance, the farthest distance is target 1, followed by targets 2 to 7, among which target 3 is a stationary vehicle. If the traditional DBSCAN clustering algorithm is used, such as Figure 6 As shown in (a), since the vehicles are close enough, it is not possible to effectively distinguish between target 1 and target 2, target 4 and target 5, and target 6 and target 7. When using the improved DBSCAN clustering algorithm proposed in the present invention, Figure 6 As shown in (b), since the speed information in the target point cloud data is introduced, and the speed information of each target is different, the point clouds belonging to different target vehicles can be effectively distinguished and the target vehicles can be distinguished.
[0100] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.
Claims
1. A method for processing millimeter wave radar point cloud data and classifying targets, characterized in that: The method comprises: Step S1, obtaining target movement data of a total of M targets in one frame through a millimeter wave radar, t=1, 2...M, and labeling them one by one; the target movement data of the M targets exist in N radar target points, i=1, 2..., N, and labeling them one by one, N≥M; obtaining target movement data corresponding to the radar target points; Step S2, obtaining a point cloud data set P = {P1...P i |i=1,2…,N}; and preprocess the point cloud dataset P; where P i Represents the point cloud data of the i-th radar target point; Step S3, introduce the Mahalanobis distance and use the DBSCAN clustering algorithm to perform point cloud clustering on the point cloud data in the preprocessed point cloud dataset P one by one to obtain the point cloud cluster S of each target t ; Step S4: point cloud clusters S1 to S M Transform to obtain intuitive point cloud clusters S1(direct)~S M (direct), and identify the target type to obtain the target classification result in the current frame; Step S5, calculating the target classification results of other consecutive frames with the frame as the center frame, and performing linkage analysis to output the real-time trajectory.
2. The method for millimeter wave radar point cloud data processing and target classification according to claim 1 is characterized in that: The step S2 comprises: Step S21, perform algorithm calculations on the target movement data in the current frame in turn; obtain a point cloud data set P = {P1...P i |i=1,2…,N}; Step S22: point cloud data P in point cloud data set P i Carte-Cartesian coordinate system transformation is performed one by one to obtain the point cloud dataset P in the Cartesian coordinate system. Cartesian ={P′1……P′ i |i=1,2…,N}; where P′ i =(x i ,y i ,v i ,σ i ), P′ i Represents the point cloud data of the i-th radar target point in the Cartesian coordinate system, x i is the lateral distance of the ith radar target point in the Cartesian coordinate system, y i is the longitudinal distance of the ith radar target point in the Cartesian coordinate system, v i is the velocity of the ith radar target point in the Cartesian coordinate system, σ i is the motion angle of the i-th radar target point in the Cartesian coordinate system; Step S23: The velocity v of the i-th radar target point in the Cartesian coordinate system i To determine the object, classify the point cloud dataset P in the Cartesian coordinate system Cartesian ={P′1……P′ i |i=1,2…,N}, and obtain the approaching radar target point cloud set P containing A radar target points Cartesian (close) = {P′ 1- ...P' k- |k=1,2…,A}; where P′ k- =(x i ,y i ,v i ,σ i |v i <0); and a cloud set P of target points far from the radar including B radar target points Cartesian (far) = {P′ 1+ ...P' q+ |q=1,2…,B}; where P′ q+ =(x i ,y i ,v i ,σ i |v i >0).
3. The method for millimeter wave radar point cloud data processing and target classification according to claim 2 is characterized in that: The step S3 comprises: Step S31, calculate the Mahalanobis distances D between different two points in the approaching radar target point cloud set P Cartesian (close) = {P′ 1- ……P′ k- |k = 1, 2…, A}, and then obtain the approaching radar target point cloud Mahalanobis distance set D(close) = {D kk’ ……D 12 ……D kk’ 丨k = 1, 2…, A, k′ = 1, 2…, A, k ≠ k’}; where Σ kk’D- represents the covariance between the kth radar target point and the k'th radar target point in the approaching radar target point cloud set; Step S32, calculate the Mahalanobis distance D between different pairs of radar target points in the set P(far) = {P′ Cartesian Cartesian ……P′ 1+ 1+ ……P′ q+ q+ |q = 1, 2…, B} one by one, and then obtain the set of Mahalanobis distances of the radar target points far away D(far) = {D qq’ qq’ ……D 12 12 ……D qq’ qq’ 丨q = 1, 2…, B, q′ = 1, 2…, B, q ≠ q’}; Σ kk’D- represents the covariance between the qth radar target point and the q'th radar target point in the far-away radar target point cloud set; Step S33, setting the Mahalanobis distance threshold ε m And the threshold value of the number of radar points in the spatial range MinP, one by one approaching the radar target point cloud set P Cartesian (close) and far away from the radar target point cloud set P Cartesian The radar target points in (far) are marked with core points and noise points of the radar target points; By judging whether any radar target point is within ε m The number of other radar target points in the range is compared with the spatial range radar point number threshold MinP. If it is greater than MinP, the point is marked as a core point; if it is less than MinP, the point is marked as a noise point; Step S34: the marked core points and noise points are classified according to the corresponding targets to form M point cloud clusters S1 to S M .
4. The method for processing millimeter wave radar point cloud data and classifying targets according to claim 3 is characterized in that: The step S4 comprises: Step S41, calculate the target motion heading angle corresponding to each point cloud cluster one by one Among them, x center Represents the lateral distance of the radar target point as the core point in the point cloud cluster in the Cartesian coordinate system, y i is the longitudinal distance of the radar target point as the core point in the point cloud cluster in the Cartesian coordinate system; and the point cloud data of each radar target point in the point cloud cluster is calculated according to the heading angle Rotate clockwise to turn the target's motion direction to the horizontal direction, and obtain the corresponding intuitive point cloud clusters S1 (direct) ~ S M (direct); Step S42, through the intuitive point cloud clusters S1 (direct) ~ S M (direct) Calculate the length l of each point cloud cluster t , width w t ; where w t =x max -x min , l t =y max -y min ;x max Indicates the maximum lateral distance among all radar target points in the point cloud cluster; y max Indicates the maximum longitudinal distance among all radar target points in the point cloud cluster; x min Indicates the minimum lateral distance among all radar target points in the point cloud cluster; y min Indicates the minimum longitudinal distance among all radar target points in the point cloud cluster; Step S43: According to the length l of each point cloud cluster t , width w t and the number of radar target points N t ', get the target classification result outcome(t) in the current frame.
5. The method for millimeter wave radar point cloud data processing and target classification according to claim 4 is characterized in that: The step S5 comprises: Step S51, taking the current frame as the center frame, select a total of U frames of continuous frame data, u = 1, 2...U, and obtain all point cloud clusters under the 1st frame to the Uth frame one by one according to steps S1 to S4, and the length l corresponding to each point cloud cluster ur , width w ur and the number of radar target points N ur 'And the corresponding target classification result outcome(ur); r represents the number of the point cloud cluster in different frames; Step S52, associating the point cloud clusters from the first frame to the Uth frame to obtain a total of E preliminary motion trajectories Trail(1) to Trail(E), where e=1, 2, ..., E; Step S53: according to the length l corresponding to Trail(e) ur , width w ur And the number of radar target points Nur', calculate the mean length of the point cloud cluster under the preliminary motion trajectory one by one Width Mean And the mean number of radar target points Step S54: calculate the lengths l of all point cloud clusters under the e-th preliminary motion trajectory one by one. pr , width w pr and the number of radar target points Npr' and the mean length of the e-th preliminary motion trajectory Width Mean And the mean number of radar target points Compare and verify; optimize the frames where the point cloud clusters with large differences are located, and output the motion trajectory Trail outcome (1)~Trail outcome (E).
6. The method for millimeter wave radar point cloud data processing and target classification according to claim 5 is characterized in that: The optimization process in step S54 includes: Delete the frames where the point cloud clusters that are too large are located.
7. The method for millimeter wave radar point cloud data processing and target classification according to claim 5 is characterized in that: The optimization process in step S54 includes: For the frame where the point cloud cluster with too large difference is located, the Kalman filter function is calculated by the point cloud data of the current frame of the target and the motion trajectory of the target in the previous frame to obtain the motion trajectory of the target in the corresponding frame.
8. The method for millimeter wave radar point cloud data processing and target classification according to claim 5 is characterized in that: In the step S21, the algorithm operation includes: executing distance dimension Fourier transform, velocity dimension Fourier transform, constant false alarm detection, Doppler phase compensation, and DOA angle estimation.
9. The method for millimeter wave radar point cloud data processing and target classification according to claim 5, characterized in that: In step S53, the mean value calculation includes calculating the corresponding length l ur , width w ur The number of radar target points Nur' is calculated by adding and averaging.
10. The method for millimeter wave radar point cloud data processing and target classification according to claim 5, characterized in that: The step S43 includes: t Greater than 16 meters, width w t If the length is greater than 3 meters and the number of radar target points Nt' is greater than 15, it is judged as a large vehicle type; t Greater than 16 meters, width w t If the distance is greater than 3 meters and the number of radar target points Nt' is greater than 8 and less than 15, it is judged as a medium vehicle type; The rest are small car types.
