A target detection method based on fusion of 4D millimeter wave radar and laser radar

Abstract

The application discloses a target detection method based on fusion of 4D millimeter wave radar and laser radar, and the method generates ROI when 4D millimeter wave radar point clouds are clustered in three-dimensional space based on the characteristics of the 4D millimeter wave radar, and only laser radars in the ROI are sent into a neural network for target detection; when the 4D millimeter wave point clouds are clustered, a density clustering based on Mahalanobis distance considering reflection intensity and speed information is proposed; compared with the method of sending all laser radar point clouds into the neural network, the technical scheme of the application greatly reduces the calculation amount, improves the speed of target detection, and has a low probability of misrecognition; when laser radars fail in scenes such as rain, snow, fog and dust, the position, size and reflection intensity of obstacles in three-dimensional space can still be perceived by the 4D millimeter wave radar, and simple classification can be performed, so that the anti-interference ability and environmental adaptability of the perception system are greatly increased.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic driving perception, and particularly relates to a target detection method based on fusion of 4D millimeter wave radar and laser radar. BACKGROUND

[0002] The external environment faced by an automatic driving car is complex and changeable, and how to accurately perceive the position, size and category of an obstacle is a big problem faced by the automatic driving car. Since target detection of a sensor is subject to defects of the sensor itself, the detection capability is restricted, and it is difficult to overcome from the algorithm level. Therefore, target detection based on sensor fusion is valued and rapidly developed.

[0003] Among them, target detection based on fusion of millimeter wave radar and laser radar is a research direction with great potential. Laser radar can output a large number of 3D point clouds, which can better represent the geometric information of the target; millimeter wave radar has good penetration and can still output target two-dimensional point clouds well under rain, snow, fog and dust conditions. Fusion of the two can improve the accuracy and robustness of target detection.

[0004] Most of the current target detection algorithms based on fusion of millimeter wave radar and laser radar encode the millimeter wave radar point cloud and the laser radar point cloud respectively, extract features through a neural network, and regress and classify the detection frame through a neural network. For example, the method proposed by Li Chao et al. in the paper "Millimeter-wave-laser radar fusion target detection based on attention" can improve the accuracy of target recognition, but has the following shortcomings:

[0005] 1) The above method will cause huge calculation consumption and slow processing speed, which is not conducive to the safety of automatic driving, because the number of laser radar point clouds is very large. 2) When the targets are very close, the detection method is prone to misidentifying two targets as one target. 3) In rain, snow, fog, dust and other scenes, due to the poor penetration of laser radar, the above method and other target detection methods based on fusion of millimeter wave radar and laser radar are difficult to obtain accurate laser radar point clouds, and can only obtain sparse two-dimensional millimeter wave radar point clouds, making it difficult to determine the position, size and category of the target in three-dimensional space.

[0006] 4D millimeter wave radar is a new sensor that has emerged in recent years. Compared with traditional millimeter wave radar, it adds a height dimension to the two-dimensional point cloud, and the number of point clouds is greatly increased, so that it can truly measure the position and size of the obstacle in three-dimensional space. SUMMARY

[0007] In order to solve the above-mentioned prior art, the present application provides a target detection method based on fusion of 4D millimeter wave radar and laser radar, generates ROI (region of interest) through 4D millimeter wave radar, extracts laser radar point cloud in ROI, and encodes and uses deep learning method for target detection. The specific technical scheme of the present application is as follows: a target detection method based on fusion of 4D millimeter wave radar and laser radar, comprising the following steps:

[0008] S1: receiving 4D millimeter wave radar point cloud and laser radar point cloud synchronized in time and space, and performing high-dimensional density clustering on 4D millimeter wave radar point cloud based on Mahalanobis distance;

[0009] S2: fitting a directed bounding box for each cluster according to the clustering result of step S1;

[0010] S3: classifying each cluster according to prior information, average reflection intensity of point cloud in the bounding box, and geometric features of the bounding box, and filtering out ground point cloud clusters and their bounding boxes;

[0011] S4: extracting laser radar point cloud in the remaining bounding box, inputting the point cloud into a neural network in turn according to the cluster, and performing target detection;

[0012] S5: judging whether a target is detected in step S4, if yes, covering the classification result and the bounding box result in step S3, if no, keeping the results in steps S2 and S3, and adding a label Radar Result.

[0013] Further, in step S1, the received 4D millimeter wave radar point cloud is traversed, the covariance matrix ∑ of the multidimensional random variable P(x, y, z, v x , v y , I) is calculated, and high-dimensional density clustering based on Mahalanobis distance is performed, wherein,

[0014] Input: 4D millimeter wave radar six-dimensional point cloud set D=(p1, p2,..., P m ), neighborhood parameter (μ, MinPts), covariance matrix ∑, sample distance measurement method, i.e. Mahalanobis distance

[0015] Output: cluster division C;

[0016] Wherein, x, y, z are the coordinates of the point cloud in three-dimensional space, v x and v y are the motion velocities of the point cloud along the X axis and along the Y axis, I is the reflection intensity of the point cloud, p1, p2,..., p m is the 4D millimeter wave radar point cloud, μ is the neighborhood distance threshold, MinPts is the minimum number of samples contained in the neighborhood of the core object during clustering, p i , pj For any two points in the point cloud set D;

[0017] The specific clustering process is:

[0018] S1-1: initialize the core object set Cluster number k = 0, unvisited sample set Γ = D, cluster partition

[0019] S1-2: traverse the point cloud set D to find all core objects;

[0020] S1-3: if the core object set The clustering ends, otherwise go to step S1-4;

[0021] S1-4: in the core object set Ω, randomly select a core object o, initialize the current cluster core object queue Ωcur = {o}, initialize the class serial number k = k + 1, initialize the current cluster sample set C k = {o}, update the unvisited sample set Γ = Γ-{o};

[0022] S1-5: if the current cluster core object queue The current clustering cluster C k is generated, the cluster partition C = {C1, C2,..., C k} is updated, the core object set Ω = Ω-C k is updated, go to step S1-3, otherwise update the core object set Ω = Ω-C k ;

[0023] S1-6: take out a core object o' from the current cluster core object queue Ωcur, find all μ-neighborhood sample sets Nμ(o') through the neighborhood distance threshold μ, let Δ = Nμ(o')∩Γ, update the current cluster sample set C k = C k ∪Δ, update the unvisited sample set Γ = Γ-Δ, update Ωcur = Ωcur∪(Δ∩Ω)-o', go to step S1-5;

[0024] Wherein, the definition of μ-neighborhood is: for x j ∈D, its μ-neighborhood contains the sample set D whose distance from x j is not greater than μ, that is, Nμ(x j ) = {x i ∈D | distance(x i , x j ) ≤ μ}, the number of the sample set is denoted as |Nμ(x j )|;

[0025] The core object is defined as follows: for any sample xj∈D, if its μ neighborhood corresponds to Nμ(xj) j It contains at least MinPts samples, that is, if |Nμ(x)| < 10 ... j If x ≥ MinPts, then x j It is the core object.

[0026] Furthermore, in steps S1-2, the method for identifying all core objects is as follows:

[0027] First, find sample x using Mahalanobis distance. j The μ-neighborhood subset Nμ(x) j );

[0028] Then, if the number of samples in the subset satisfies |N∈(x j If |≥MinPts, then the sample x j Add to the core object sample set: Ω=Ω∪{x j}

[0029] Furthermore, in step S2, the input is: 4D millimeter-wave radar point cloud cluster C. i =(p i1 p i2 , ..., p in ), iteration step size δ, p i1 p i2 , ..., p in C i 4D millimeter-wave radar point cloud; Output: bounding box bottom rectangular boundary {a j x′+b j y′=c j |j=1,2,3,4}, bounding box height h, a j b j c j The parameters are for the linear expression; the specific process of step S2 is as follows:

[0030] S2-1: Project the 3D point cloud cluster onto the xy plane to obtain N = ((x1′, y1′), (x2′, y2′)..., (x m ′,y m ′)), (x1′, y1′), (x2′, y2′)..., (x m ′,y m ′) are the coordinates of the point cloud projected onto the xy plane, respectively;

[0031] S2-2: Initialize the angle θ = 0 of any side of the bounding box bottom rectangle relative to the x-axis, and empty dictionary Q;

[0032] S2-3: Let e1=(cos 0, sin 0), e2=(-sin 0, cos 0), calculate

[0033] S2-4: Let Calculate Q <- (0: a);

[0034] S2-5: Determine whether 0 < n / 2 is true, if true, execute 0 = 0 + d, jump to step S2-3; if not, execute step S2-6;

[0035] S2-6: Select the key 0 in Q corresponding to the maximum a * , calculate

[0036] S2-7: Calculate a1=cos 0 * , b1=sin 0 * , a2=-sin 0, b2=cos 0, a3=cos 0 * , b3=sin 0 * , a4=-sin 0, b4=cos 0,

[0037] S2-8: Traverse the point cloud cluster C i , calculate h=p.z max -p.z min , p.z max is the maximum z value of all point clouds in C i , p.z min is the minimum z value of all point clouds in C i ;

[0038] S2-9: The algorithm ends.

[0039] Further, the prior information in the step S3 is: by collecting 4D millimeter wave radar point cloud data, executing the step S1 and the step S2 on each frame of point cloud data, and counting the average reflection intensity of the point cloud cluster and the height information of the point cloud cluster bounding box of the ground, pedestrians, small vehicles and large vehicles.

[0040] Further, the point cloud target detection network in the step S4 is PointPillars.

[0041] The beneficial effects of the present application are:

[0042] 1.The application proposes a target detection method based on fusion of 4D millimeter wave radar and laser radar, based on the characteristics of 4D millimeter wave radar, the ROI is generated when the 4D millimeter wave radar point cloud is clustered in three-dimensional space, only the laser radar in the ROI is sent into the neural network for target detection, compared with the method of sending all laser radar point clouds into the neural network, the calculation amount is greatly reduced, and the speed of target detection is improved.

[0043] 2.When the 4D millimeter wave point cloud is clustered, the application proposes a density clustering based on Mahalanobis distance considering reflection intensity and velocity information, compared with the clustering method based on Euclidean distance, it can distinguish targets with different properties and different velocities that are very close, and reduce the probability of misidentification.

[0044] 3.In rain, snow, fog, dust and other scenes where laser radar fails, compared with the traditional millimeter wave radar and laser radar fusion perception method, the application can still perceive the position, size, reflection intensity of obstacles in three-dimensional space through 4D millimeter wave radar and perform simple classification, greatly increasing the anti-interference ability and environmental adaptability of the perception system. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed in the embodiments, and the features and advantages of the application will be more clearly understood by referring to the drawings, which are schematic and should not be understood as any limitation to the application. Those skilled in the art can obtain other drawings without creative labor. Among them:

[0046] Figure 1 is the target detection flowchart of the application based on fusion of 4D millimeter wave radar and laser radar;

[0047] Figure 2 is the clustering result classification and ground filtering schematic diagram of the application. DETAILED DESCRIPTION

[0048] In order to more clearly illustrate the above-mentioned purposes, features and advantages of the application, the following will further describe the application in combination with the drawings and specific embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other without conflict.

[0049] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, however, the application can also be implemented in other ways different from those described herein, therefore, the protection scope of the application is not limited by the specific embodiments disclosed below.

[0050] As Figure 1As shown, a target detection method based on fusion of 4D millimeter wave radar and laser radar includes the following steps:

[0051] S1: receiving 4D millimeter wave radar point cloud and laser radar point cloud after time and space synchronization, and performing high-dimensional density clustering on 4D millimeter wave radar point cloud based on Mahalanobis distance;

[0052] Traverse the received 4D millimeter wave radar point cloud, calculate the covariance matrix ∑ of the multi-dimensional random variable P(x, y, z, v x , v y , I), and perform high-dimensional density clustering based on Mahalanobis distance, wherein

[0053] Input: 4D millimeter wave radar six-dimensional point cloud set D=(p1, p2,..., p m ), neighborhood parameter (μ, MinPts), covariance matrix ∑, sample distance measurement method, i.e. Mahalanobis distance

[0054] Output: cluster division C;

[0055] Wherein, x, y, z are the coordinates of the point cloud in three-dimensional space, v x and v y are the motion velocities of the point cloud along the X axis and along the Y axis, I is the reflection intensity of the point cloud, p1, p2,..., p m is the 4D millimeter wave radar point cloud, μ is the neighborhood distance threshold, MinPts is the minimum number of samples contained in the neighborhood of the core object during clustering, p i , p j are any two points in the point cloud set D;

[0056] The specific clustering process is as follows:

[0057] S1-1: initialize the core object set Cluster number k=0, unvisited sample set Γ=D, cluster division

[0058] S1-2: traverse the point cloud set D to find all core objects;

[0059] S1-3: if the core object set is empty, the clustering ends, otherwise go to step S1-4;

[0060] S1-4: randomly select a core object o from the core object set Ω, initialize the current cluster core object queue Ωcur={o}, initialize the class serial number k=k+1, initialize the current cluster sample set C k ={o}, and update the unvisited sample set Γ=Γ-{o};

[0061] S1-5: If the current cluster core object queue Then the current cluster C k Once generated, update the cluster partition C = {C1, C2, ..., C}. k}, Update the core object set Ω = Ω - C k Proceed to steps S1-3; otherwise, update the core object set Ω = Ω - C. k ;

[0062] S1-6: Take a core object o′ from the current cluster core object queue Ωcur, find all μ-neighborhood subsets Nμ(o′) using the neighborhood distance threshold μ, let Δ=Nμ(o′)∩Γ, and update the current cluster sample set C. k =C k ∪Δ, update the unvisited sample set Γ=Γ-Δ, update Ωcur=Ωcur∪(Δ∩Ω)-o′, and proceed to step S1-5;

[0063] The μ neighborhood is defined as: for x j ∈D, its μ neighborhood contains samples in the sample set D that are the same as x. j The subset of samples whose distance is no greater than μ, i.e., Nμ(x) j )={x i ∈D|distance(x i x j The number of samples in this subset is denoted as |Nμ(x) ≤ μ}. j )|;

[0064] The core object is defined as: for any sample x j ∈D, if its μ neighborhood corresponds to Nμ(x) j It contains at least MinPts samples, that is, if |Nμ(x)| < 10 ... j If x ≥ MinPts, then x j It is the core object.

[0065] In step S1, the 4D millimeter-wave radar point cloud has its motion velocity information and reflection intensity information. Therefore, during clustering, not only distance is considered, but also velocity and reflection intensity. Since the point cloud coordinates (x, y, z) are related to its reflection intensity I and velocity (v... x v y Because the units of measurement are different and the numerical differences are large, the density clustering algorithm based on Mahalanobis distance is adopted, p1(x1, y1, z1, v) x1 v y1 ,I1),p2(x2,y2,z2,v x2 v y2 The Mahalanobis distance between two points (I2) can be expressed as: where ∑ is the covariance matrix. By the measurement of Mahalanobis distance, the problem of different dimensions and different scales of each dimension data can be eliminated. Density clustering algorithm generally assumes that the class can be determined by the tightness of the sample distribution. The samples of the same class are closely connected, that is, there are samples of the same class in the vicinity of any sample in the class. By classifying the closely connected samples into a class, a clustering class is obtained. By classifying all groups of closely connected samples into different classes, the final clustering class result is obtained.

[0066] S2: fitting a directional bounding box for each cluster according to the clustering result of step S1;

[0067] Input: 4D millimeter wave radar point cloud cluster C i = (p i1 , p i2 ,..., p in ), iteration step δ, p i1 , p i2 ,..., p in 4D millimeter wave radar point cloud in C i ; Output: bounding box bottom rectangular boundary {a j x' + b j y' = c j |j = 1, 2, 3, 4}, bounding box height h, a j , b j , c j are parameters of the straight line expression; the specific process of step S2 is as follows:

[0068] S2-1: project the three-dimensional point cloud cluster to the xy plane to obtain N = ((x1', y1'), (x2', y2')..., (x m ', y m ')}, (x1', y1'), (x2', y2')..., (x m ', y m ') are the coordinates of the point cloud projected onto the xy plane;

[0069] S2-2: initialize the angle θ = 0 of any side of the bounding box bottom rectangle relative to the x axis, and the empty dictionary Q;

[0070] S2-3: let e1 = (cos θ, sin θ), e2 = (-sin θ, cos θ), and calculate

[0071] S2-4: let calculate Q <- (θ: α);

[0072] S2-5: Determine whether θ < π / 2 is true. If it is true, execute θ = θ + δ and jump to step S2-3; if it is not true, execute step S2-6.

[0073] S2-6: Select the bond θ in Q that corresponds to the maximum α. * ,calculate

[0074] S2-7: Calculate a1 = cosθ * b1 = sinθ * , a² = -sinθ, b² = cosθ a3=cosθ * b3 = sinθ * , a4 = -sinθ, b4 = cosθ

[0075] S2-8: Traversing point cloud cluster C i Calculate h = pz max -pz min pz max C i The maximum z-value of all point clouds, pz min C i The minimum z-value of all point clouds;

[0076] S2-9: Algorithm ends.

[0077] In step S2, a directed bounding box is fitted to each clustered point cloud cluster. The basic idea is to traverse all possible directions of the rectangle. In each iteration, a rectangle pointing in that direction and containing all scan points can be easily found. The direction, length, and width of the rectangle are determined according to the principle of minimizing the area of ​​the rectangle. The height of the bounding box is determined by the difference between the highest and lowest points of the point cloud in the point cloud cluster.

[0078] S3: Based on prior information and the average reflection intensity of the point cloud within the bounding box and the geometric features of the bounding box, classify each cluster and filter out ground point cloud clusters and their bounding boxes.

[0079] The prior information is as follows: by collecting 4D millimeter-wave radar point cloud data, steps S1 and S2 are performed on each frame of point cloud data to statistically analyze the average reflection intensity of point cloud clusters of ground, pedestrians, small vehicles, and large vehicles, as well as the height information of the bounding box of the point cloud clusters.

[0080] For example, such as Figure 2 As shown, 1000 frames from a self-made open-pit mine point cloud dataset were selected for statistical analysis, and the following results were obtained:

[0081] The statistical analysis of the average reflection intensity of the ground point cloud cluster, the pedestrian point cloud cluster, the car and other small vehicle point cloud cluster, and the truck and other large vehicle point cloud cluster is performed, and it is obtained that the average reflection intensity of the ground point cloud cluster is between 2 and 6, the average reflection intensity of the pedestrian point cloud cluster is between 4 and 8, and the reflection intensity of the car and large vehicle is between 15 and 25.

[0082] The statistical analysis of the geometric features of the bounding box of the ground point cloud cluster, the bounding box of the pedestrian point cloud cluster, the bounding box of the car and other small vehicle point cloud cluster, and the bounding box of the truck and other large vehicle point cloud cluster is performed, and it is obtained that the height of the bounding box of the ground point cloud cluster is between 0 m and 0.3 m, the height of the bounding box of the pedestrian point cloud cluster is between 1.2 m and 1.9 m, the height of the bounding box of the car and other small vehicle point cloud cluster is between 1.2 m and 1.9 m, and the height of the bounding box of the truck and other large vehicle point cloud cluster is above 2.3 m.

[0083] According to the above prior information, the ground point cloud cluster and its bounding box that meet the reflection intensity condition and the geometric condition are filtered out, the remaining bounding boxes are classified according to their geometric features and the average reflection intensity of the point cloud in them, and the point cloud clusters that meet the pedestrian, small vehicle, and large vehicle are assigned a class label.

[0084] S4: Extract the laser radar point cloud in the remaining bounding box, input the neural network in turn according to the cluster, and perform target detection;

[0085] S5: Determine whether a target is detected in step S4, if yes, cover the classification result and the bounding box result in step S3, if no, keep the results in steps S2 and S3, and add a label Radar Result.

[0086] In some embodiments, in step S1-2, the method for finding all core objects is:

[0087] First, find the μ neighborhood subsample set Nμ(x j ) of the sample x j by the Mahalanobis distance measurement method.

[0088] Then, if the number of subsample set samples satisfies |N∈(x j )|≥MinPts, add the sample x j to the core object sample set: Ω=Ω∪{x j}.

[0089] Preferably, the point cloud target detection network in step S4 includes but is not limited to PointPillars.

[0090] In the present application, unless specifically defined otherwise and limited, the terms "mount", "connect", "connection", "fixed", and the like, should be understood broadly, for example, can be fixed connection, can also be detachable connection, or integral; can be mechanical connection, can also be electrical connection; can be direct connection, can also be indirect connection through an intermediate medium, can be internal communication of two elements or interaction relationship of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0091] In the present application, unless specifically defined otherwise and limited, the first feature "on" or "under" the second feature can include that the first and second features are in direct contact, or that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, the first feature "on", "above" and "over" the second feature includes that the first feature is directly above and obliquely above the second feature, or only indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature "under", "below" and "under" the second feature includes that the first feature is directly below and obliquely below the second feature, or only indicates that the horizontal height of the first feature is less than that of the second feature.

[0092] In the present application, the terms "first", "second", "third", "fourth" are only for descriptive purpose, and should not be understood as indicating or implying relative importance. The term "a plurality of" means two or more, unless otherwise specifically limited.

[0093] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A target detection method based on fusion of 4D millimeter wave radar and lidar, characterized in that, The method comprises the following steps: S1: receiving a 4D millimeter wave radar point cloud and a laser radar point cloud which are spatio-temporally synchronized, and performing high-dimensional density clustering on the 4D millimeter wave radar point cloud based on Mahalanobis distance; S2: fitting a directed bounding box for each cluster according to the clustering result of step S1; S3: classifying each cluster according to prior information and average reflection intensity of point cloud in the bounding box, geometric features of the bounding box, and filtering out ground point cloud clusters and their bounding boxes; S4: extracting laser radar point clouds in the remaining bounding boxes, inputting the point clouds into a neural network in turn according to the clusters, and performing target detection; S5: judging whether a target is detected in step S4, if yes, covering the classification result and the bounding box result in step S3, if no, retaining the results in steps S2 and S3, and adding a label Radar Result.

2. The object detection method of claim 1, wherein, In the step S1, the received 4D millimeter wave radar point cloud is traversed, the covariance matrix Σ of the multi-dimensional random variables P(x, y, z, v x , v y , I) is calculated, and high-dimensional density clustering based on Mahalanobis distance is performed, Wherein, Input: 4D mmWave radar six-dimensional point cloud set D = (p1, p2,..., pN) m neighborhood parameters (μ, MinPts), covariance matrix Σ, sample distance metric, i.e., Mahalanobis distance Output: cluster division C; where x, y, z are the coordinates of the point cloud in three-dimensional space, v x and v y are the motion velocities of the point cloud along the X-axis and along the Y-axis, I is the reflection intensity of the point cloud, p1, p2,..., p m is a 4D millimeter wave radar point cloud, μ is a neighborhood distance threshold value, MinPts is the minimum number of samples contained in the neighborhood of a core object when clustering, p i , p j are any two points in the point cloud set D; The specific clustering process is: S1-1: Initialize the core object set Cluster number of clusters k = 0, unvisited sample set Γ = D, cluster partition S1-2: traversing the point cloud set D to find all core objects; S1-3: If the core object set then the clustering ends, otherwise go to step S1-4; S1-4: In the core object set Ω, randomly select a core object o, initialize the current cluster core object queue Ωcur = {o}, initialize the category index k = k + 1, and initialize the current cluster sample set C. k ={o}, update the unvisited sample set Γ = Γ - {o}; S1-5: If the current cluster core object queue then the current cluster C k is generated, the cluster division C = {C1, C2,..., C k} is updated, the core object set Ω = Ω - C k is updated, and the process returns to step S1-3, otherwise the core object set Ω = Ω - C k is updated. S1-6: Take out a core object o' from the current cluster core object queue Ωcur, find all μ-neighborhood sub-sample sets Nμ(o') through the neighborhood distance threshold μ, let Δ = Nμ(o')∩Γ, update the current cluster sample set C = C k k ∪Δ, update the unvisited sample set Γ = Γ-Δ, update Ωcur = Ωcur∪(Δ∩Ω)-o', and go to step S1-5.​ wherein the definition of the μ-neighborhood is: for x j ∈D, its μ-neighborhood contains the sub-sample set in the sample set D with the distance not greater than μ from x j , i.e. Nμ(x j ) = {x i ∈D | distance(x i , x j ) ≤ μ}, and the number of the sub-sample set is denoted as |Nμ(x j )|; A core object is defined as: for any sample x j ∈ D, if its μ-neighborhood corresponds to Nμ(x j ) containing at least MinPts samples, i.e., if |Nμ(x j )| ≥ MinPts, then x j is a core object.

3. The object detection method of claim 2, wherein, In step S1-2, the method for finding all core objects is: First, find the μ-neighborhood sub-sample set Nμ(x j ) of x j by Mahalanobis distance metric Then, if the number of samples in the sub-sample set satisfies |N∈(x j )|≥MinPts, the sample x j is added to the core object sample set: Ω=Ω∪{x j}.

4. The object detection method according to claim 1 or 2, characterized in that, In the step S2, input: 4D mmWave radar point cloud cluster C i = (p i1 ,p i2 ,...,p in ), iteration step size δ, p i1 ,p i2 ,...,p in 4D mmWave radar point cloud in C i ; Output: bounding box bottom rectangle boundary {a j x' + b j y' = c j | j = 1, 2, 3, 4}, bounding box height h, a j , b j , c j are parameters of the straight line expression; the specific process of step S2 is: S2-1: Project the three-dimensional point cloud cluster to the xy-plane to obtain N = ((x1', y1'), (x2', y2')..., (xN', yN')). m m ′,y​ (x1',y1'), (x2',y2')..., (x m ′,y m ′) are the coordinates of the point cloud projected onto the xy plane, respectively; S2-2: initializing the angle of any side of the bottom rectangle of the bounding box relative to the x-axis θ=0, and an empty dictionary Q; S2-3: Let e1 = (cos 0, sin 0), e2 = (-sin 0, cos 0), compute S2-4: Let Compute Q←(θ:α); S2-5: judging whether θ<π / 2 is true, if true, executing θ=θ+δ and jumping to step S2-3; if not true, executing step S2-6; S2-6: Select the bond θ in Q corresponding to the maximum α * , calculate S2-7: Compute a1 = cos θ * , b1 = sin θ * , a2 = -sin θ, b2 = cos θ, a3 = cos θ * b3 = sin θ * , a4 = -sin θ, b4 = cos θ, S2-8: Traverse the point cloud cluster C i , compute h = p.z max -p.z min , p.z max is the maximum z value of all point clouds in C i , p.z min is C i minimum z value for all point clouds in the middle S2-9: the algorithm ends.

5. The object detection method according to claim 1 or 2, characterized in that, The prior information in step S3 is: by collecting 4D millimeter wave radar point cloud data, performing steps S1 and S2 on each frame of point cloud data, and counting the average reflection intensity of point cloud clusters and the height information of point cloud cluster bounding boxes of ground, pedestrians, small vehicles or large vehicles.

6. The object detection method of claim 5, wherein, The point cloud target detection network in step S4 is PointPillars.

Images