A method, device and medium for detecting aerial targets based on laser radar

Abstract

The present invention discloses a method, device and medium for detecting aerial targets based on laser radar, the method comprising: using laser radar to continuously scan aerial scenes to obtain continuous frame three-dimensional point cloud data; using statistical filters to perform data denoising on each frame of three-dimensional point cloud data; designing and using a minimum spanning tree cutting algorithm to cluster point clouds, dividing each frame of point clouds into different clustered point cloud clusters; performing feature extraction and feature analysis on the segmented clustered point cloud clusters, respectively, obtaining clustered point cloud clusters of the target of interest, and calculating the size of the target; using a principal component analysis algorithm on the clustered point cloud clusters of the target of interest, the center position coordinates and attitude parameters of the target of interest are calculated. Based on the minimum spanning tree cutting algorithm and the principal component analysis algorithm, the present invention can effectively extract the point cloud of the aerial target of interest, and calculate the center position coordinates and attitude parameters of the aerial target.

Description

Technical Field

[0001] The present invention belongs to the field of laser radar data processing, and in particular relates to a laser radar-based aerial target detection method, device and medium. Background Art

[0002] Using LiDAR for data collection and processing to detect aerial targets, such as drones, birds, balloons, and other objects, can meet the needs of some industries. For example, in the security field, potential aerial threat targets can be detected and identified to issue warnings; in the military field, aerial targets can be detected to track and locate aerial targets.

[0003] The three-dimensional point cloud data obtained by LiDAR scanning the aerial scene usually has a strong distinction between the background and the foreground. This is because most of the laser rays will be projected into the clear area without the target object, and only the objects reflected by the target will be collected by the LiDAR sensor to form three-dimensional point cloud data. LiDAR has an increasingly wide application in the field of vehicle-mounted assisted driving, but there are relatively few studies on aerial scene detection, and the reference materials are relatively limited. In the field of autonomous driving, researchers use target detection methods based on Bird's eye view (BEV) and Deep Neural Network (DNN) models and other technical routes to perform scene analysis, detect and measure targets of interest.

[0004] The BEV-based method is to project a frame of three-dimensional point cloud scanned by the lidar onto the XY coordinate plane with the ground as the reference plane to obtain a two-dimensional projection image. This method requires manual setting of the resolution of the bird's-eye view during discretization, and for aerial scenes, there is a lack of ideal reference planes for projection.

[0005] The target detection method based on the DNN model uses a trained deep neural network model to detect and identify targets on a 3D point cloud. This method requires a large amount of training data to achieve good recognition results. Moreover, since the deep neural network model usually has many parameters, there are problems such as low computational efficiency and slow running speed.

[0006] There are some special data quality issues in the three-dimensional point cloud data of laser radar for aerial scene scanning, including: the three-dimensional point cloud of the scanned target surface is relatively sparse due to the wide range of aerial scenes; data noise points are generated due to the interference of the atmospheric environment and background light, etc. To meet the needs of aerial target detection, it is of great significance to design a set of laser radar data processing methods, devices and media with high computational efficiency and stable detection effect to solve the needs of aerial target detection. Summary of the invention

[0007] Purpose of the invention: The present invention provides a method, device and medium for aerial target detection based on laser radar, which can effectively extract the point cloud cluster of the target of interest from the three-dimensional point cloud scanned by the laser radar, and calculate the center position coordinates and posture parameters of the target of interest.

[0008] Technical solution: The laser radar-based aerial target detection method of the present invention comprises the following steps:

[0009] (1) Obtain three-dimensional point cloud data frame by frame from continuous scanning by the LiDAR;

[0010] (2) De-noising each frame of 3D point cloud data using a statistical filter based on a distance discrimination condition to obtain a frame of denoised 3D point cloud data;

[0011] (3) For the denoised 3D point cloud data, the minimum spanning tree cutting algorithm is used to cluster the point cloud and divide each frame of point cloud into different clustered point cloud clusters;

[0012] (4) Perform feature extraction and feature analysis on the segmented clustered point cloud clusters to obtain the clustered point cloud clusters of the target of interest and calculate the size of the target of interest;

[0013] (5) For the clustered point cloud clusters of the target of interest, the principal component analysis algorithm is used to calculate the center position coordinates and attitude parameters of the target of interest to achieve the detection of aerial targets.

[0014] Furthermore, the three-dimensional point cloud data in step (1) are acquired frame by frame by continuously scanning the aerial scene using a rotating mirror or MEMS galvanometer laser radar.

[0015] Furthermore, the implementation process of step (2) is as follows:

[0016] For each frame of 3D point cloud data, a KD tree data structure is established; according to the KD tree data structure, each point is used as a search point to find the closest points;

[0017] Calculate the nearest The distance between each point and the search point is calculated, and the distance values ​​are sorted from large to small, and the largest distance value is used as the discrete measurement value of the current search point; each point in a frame of 3D point cloud data is used as a search point, and a corresponding discrete measurement value is calculated for each point. If the discrete measurement value of a point is larger, it is more likely to be an outlier point, and the probability that the point is a noise point is greater;

[0018] Sort all points in a frame of 3D point cloud data from small to large according to their discrete measurement values, and select a preset ratio The points ranked at the back are regarded as noise points, which are deleted; the remaining points constitute a frame of 3D point cloud after noise removal.

[0019] Furthermore, the implementation process of step (3) is as follows:

[0020] Calculate the distance from each point in the 3D point cloud to The Euclidean distance of each field point is calculated; the point closest to the laser radar is selected as the source point, the shortest path from the source point to other points is calculated, and the minimum spanning tree MST is constructed using the Dijkstra algorithm; threshold judgment is performed along the edge of the MST, and when the distance value of an edge is greater than the preset distance segmentation threshold, this edge is cut to obtain the cut MST subgraph; for the cut MST subgraph, continue to judge along its edge, and when the distance value of an edge is greater than the preset distance segmentation threshold, this edge is cut until there is no MST subgraph that can be further segmented; all points in each segmented MST subgraph constitute an independent clustering point cloud cluster.

[0021] Furthermore, the implementation process of step (4) is as follows:

[0022] Calculate the minimum rectangular bounding box of each segmented point cloud cluster, and extract the length L, width W, and height H of the bounding box as the size of the target in the point cloud cluster; count the number of points num in each point cloud cluster;

[0023] According to the pre-set prior size of the object of interest, a cluster of the object of interest is extracted from the clustered point cloud clusters.

[0024] Furthermore, the implementation process of step (5) is as follows:

[0025] For the clustered point cloud cluster of the target of interest, calculate all the points in it The geometric center coordinates of , as the center position coordinates of the object of interest:

[0026] ;

[0027] Using the principal component analysis algorithm, calculate the cofactor matrix , the cofactor matrix Computing eigenvalues ​​via singular value decomposition and the eigenvector , :

[0028] ;

[0029] Cofactor Matrix The three eigenvectors of , and The three axes that are orthogonal in space are formed. The attitude parameters of the three orthogonal axes of the target of interest in the laser radar coordinate system are calculated by the following formula, that is, the three attitude angles :

[0030] .

[0031] A device according to the present invention includes a memory and a processor, wherein:

[0032] A memory for storing computer programs that can be run on the processor;

[0033] A processor is used to execute the steps of the above-mentioned method for detecting aerial targets based on laser radar when running the computer program.

[0034] The present invention provides a storage medium having a computer program stored thereon, and when the computer program is executed by at least one processor, the steps of the above-mentioned method for detecting aerial targets based on laser radar are implemented.

[0035] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention first designs a minimum spanning tree cutting algorithm to perform point cloud clustering, the calculation process is simple, and the point clouds belonging to different targets can be effectively segmented into different point cloud clusters; the principal component analysis algorithm proposed in the present invention can calculate the center position coordinates and posture parameters of the aerial target from the point cloud cluster of the aerial target of interest, thereby meeting the detection and positioning requirements for the aerial target. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the aerial target detection method based on laser radar;

[0037] Figure 2 Schematic diagram of scanning aerial targets for LiDAR detection;

[0038] Figure 3 Schematic diagram of the point cloud clustering process based on the minimum spanning tree cutting algorithm, where (a) is a frame of point cloud after denoising; (b) is the constructed minimum spanning tree; (c) is the three clustered point cloud clusters after cutting based on the minimum spanning tree. DETAILED DESCRIPTION

[0039] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0040] like Figure 1 As shown, the present invention proposes an aerial target detection method based on laser radar, which specifically includes the following steps:

[0041] Step 1: Use a rotating mirror or MEMS galvanometer laser radar to continuously scan the aerial scene and obtain three-dimensional point cloud data frame by frame from the laser radar continuous scanning.

[0042] A rotating mirror laser radar is fixedly installed on a communication tower at a certain height from the ground. The drone target is taken as the target of interest, and the laser radar is used to detect the drone target in the air. The schematic diagram of laser radar detecting and scanning the aerial target is shown in the figure below. Figure 2 As shown. Point the laser radar toward the sky, power it on and connect it to the processor via Ethernet. Continuously scan the aerial scene to achieve the effect of continuous observation and obtain three-dimensional point cloud data scanned by the laser radar frame by frame.

[0043] Step 2: Each time a new frame of 3D point cloud data is obtained, denoising is performed using a statistical filter based on a distance discrimination condition to obtain a frame of denoised 3D point cloud data.

[0044] For each frame of 3D point cloud data, a KD tree data structure is established. According to the KD tree data structure, each point is used as a search point to find the closest points, The value range is between 3 and 10.

[0045] Calculate the nearest The distance between each point and the search point is calculated, and the distance values ​​are sorted from large to small, and the largest distance value is used as the discrete measurement value of the current search point. Each point in a frame of 3D point cloud data is used as a search point, and a corresponding discrete measurement value is calculated for each point. The larger the discrete measurement value of a point, the more likely it is an outlier point, and the greater the probability that the point is a noise point.

[0046] All points in a frame of 3D point cloud data are sorted from small to large according to their discrete metric values, and 10% of the points at the bottom of the list are selected as noise points, which are deleted. The remaining points constitute a frame of 3D point cloud after noise removal.

[0047] Step 3: If Figure 3 As shown in , for the denoised three-dimensional point cloud data, Figure 3 As shown in (a), the minimum spanning tree cutting algorithm is used to cluster the point cloud, and each frame of point cloud is divided into different clustered point cloud clusters.

[0048] Calculate the distance from each point in the 3D point cloud to The Euclidean distance of the domain points, The value range of is between 6 and 11. The point closest to the laser radar is selected as the source point. Then, the shortest path from the source point to other points is calculated, and the minimum spanning tree (MST) is constructed using the Dijkstra algorithm. The minimum spanning tree is an undirected graph model, such as Figure 3 As shown in (b), threshold judgment is performed along the edge of MST. When the distance value of an edge is greater than the preset distance segmentation threshold, the edge is cut to obtain the MST subgraph after cutting.

[0049] For the cut MST subgraph, continue to judge along its edge. When the distance value of an edge is greater than the preset distance segmentation threshold, continue to cut on this edge until there is no MST subgraph that can be further segmented. At this time, all the points in each segmented MST subgraph constitute an independent cluster point cloud cluster, such as Figure 3 As shown in (c).

[0050] Step 4: Perform feature extraction and feature analysis on the segmented clustered point cloud clusters to obtain the clustered point cloud clusters of the UAV target of interest and calculate the size of the UAV target of interest.

[0051] Calculate the minimum rectangular bounding box of each segmented point cloud cluster, and extract the length L, width W, and height H of the bounding box as the size of the target in the point cloud cluster. Count the number of points num in each point cloud cluster. Extract a cluster of targets of interest from the point cloud cluster according to the pre-set prior size of the drone target of interest.

[0052] Step 5: For the clustered point cloud clusters of the target of interest, use the principal component analysis algorithm to calculate the center position coordinates and posture parameters of the target of interest.

[0053] For the clustered point cloud clusters of the drone target of interest, calculate all the points in it The geometric center coordinates of , as the center position coordinates of the UAV target:

[0054]

[0055] Using the principal component analysis algorithm, calculate the cofactor matrix , the cofactor matrix Computing eigenvalues ​​via singular value decomposition and the eigenvector ( ):

[0056]

[0057] Cofactor Matrix The three eigenvectors of , and The three axes that are orthogonal in space are formed. The following formula is used to calculate the attitude parameters of the three orthogonal axes of the UAV target of interest in the laser radar coordinate system, that is, the three attitude angles :

[0058] .

[0059] The present invention also provides an apparatus, comprising a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the above-mentioned method for detecting aerial targets based on laser radar when running the computer program.

[0060] The present invention also provides a storage medium having a computer program stored thereon, and when the computer program is executed by at least one processor, the steps of the above-mentioned method for detecting aerial targets based on laser radar are implemented.

[0061] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting aerial targets based on laser radar, characterized in that: The following steps are involved: (1) Obtain three-dimensional point cloud data from continuous scanning by laser radar frame by frame; (2) De-noising each frame of three-dimensional point cloud data using a statistical filter based on a distance discrimination condition to obtain a frame of denoised three-dimensional point cloud data; (3) For the denoised 3D point cloud data, the minimum spanning tree cutting algorithm is used to cluster the point cloud, and each frame of point cloud is divided into different clustered point cloud clusters; (4) Perform feature extraction and feature analysis on the segmented clustered point cloud clusters to obtain the clustered point cloud clusters of the target of interest and calculate the size of the target of interest; (5) Using the principal component analysis algorithm, the center position coordinates and attitude parameters of the target of interest are calculated for the clustered point cloud clusters to achieve aerial target detection; The implementation process of step (3) is as follows: Calculate the Euclidean distance from each point in the three-dimensional point cloud to k domain points; select the point closest to the lidar as the source point, calculate the shortest path from the source point to other points, and use the Dijkstra algorithm to build the minimum spanning tree MST; perform threshold judgment along the edge of the MST, and when the distance value of an edge is greater than the preset distance segmentation threshold, cut on this edge to obtain the cut MST subgraph; for the cut MST subgraph, continue to judge along its edge, and when the distance value of an edge is greater than the preset distance segmentation threshold, cut on this edge until there is no MST subgraph that can be further segmented; all points in each segmented MST subgraph constitute an independent clustering point cloud cluster.

2. The method for detecting aerial targets based on laser radar according to claim 1, characterized in that: The three-dimensional point cloud data in step (1) is obtained by continuously scanning the aerial scene using a rotating mirror or MEMS galvanometer laser radar.

3. The method for detecting aerial targets based on laser radar according to claim 1, characterized in that: The implementation process of step (2) is as follows: For each frame of 3D point cloud data, a KD tree data structure is established; based on the KD tree data structure, each point is used as a search point to find the n points closest to it; Calculate the distance between the nearest n points and the search point, and sort them from large to small according to the distance value, and use the largest distance value as the discrete measurement value of the current search point; take each point in a frame of 3D point cloud data as the search point, and calculate a corresponding discrete measurement value for each point. The larger the discrete measurement value of a point, the more likely it is an outlier point, and the greater the probability that the point is a noise point; All points in a frame of 3D point cloud data are sorted from small to large according to their discrete measurement values, and a preset proportion p% of the points in the last position are selected as noise points, and these noise points are deleted; The retained points constitute a frame of 3D point cloud after noise removal.

4. The method for detecting aerial targets based on laser radar according to claim 1, characterized in that: The implementation process of step (4) is as follows: The minimum rectangular bounding box of the segmented clustered point cloud clusters is calculated respectively, and the length L, width W and height H of the bounding box are extracted as the size of the target in the clustered point cloud cluster; the number of points num in each clustered point cloud cluster is counted respectively; according to the pre-set prior size of the target of interest, a cluster of targets of interest is extracted from the clustered point cloud cluster.

5. The method for detecting aerial targets based on laser radar according to claim 1, characterized in that: The implementation process of step (5) is as follows: For the clustered point cloud cluster of the target of interest, calculate all the points {p1,p2,...,p i ,...,p num }'s geometric center coordinates As the center position coordinates of the object of interest: Using the principal component analysis algorithm, calculate the cofactor matrix Calculate the eigenvalue λ of the cofactor matrix C by singular value decomposition j and the eigenvector v j , j=1,2,3: C·v j =λ j ·v j ,0<λ1<λ2<λ3 The three eigenvectors v1 of the cofactor matrix C = (v 11 ,v 12 ,v 13 ) T ,v2=(v 21 ,v 22 ,v 23 ) T and v3=(v 31 ,v 32 ,v 33 ) T The three axes that are orthogonal in space are formed. The attitude parameters of the three orthogonal axes of the target of interest in the laser radar coordinate system are calculated by the following formula, that is, the three attitude angles 6. The method for detecting aerial targets based on laser radar according to claim 3, characterized in that: The value range of n is between 3 and 10.

7. The method for detecting aerial targets based on laser radar according to claim 1, characterized in that: The value range of k is between 6 and 11.

8. A device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of a laser radar-based aerial target detection method as described in any one of claims 1 to 7 when running the computer program.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of a laser radar-based aerial target detection method as described in any one of claims 1 to 7.

Images