High-precision positioning method for sparse point cloud of airborne target
By combining neural networks and voting networks with a sparse point cloud replication downsampling strategy, the problem of inaccurate airborne sparse point cloud positioning was solved, achieving high-precision target centroid positioning and meeting real-time positioning requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN FLIGHT SELF CONTROL INST OF AVIC
- Filing Date
- 2025-12-27
- Publication Date
- 2026-04-28
AI Technical Summary
In airborne scenarios, the positioning accuracy of sparse point clouds of distant targets is not high, and existing technologies are difficult to accurately locate them, especially since the uneven distribution of sparse point clouds leads to large positioning errors.
A neural network-based approach is adopted, which performs two centroid localizations through point cloud feature extraction and voting network. Combined with sparse point cloud replication and downsampling strategies, the number of points and feature information are adjusted to improve localization accuracy.
It achieves high-precision positioning of airborne sparse point clouds, meets real-time positioning requirements, and improves the accuracy and efficiency of target centroid positioning.
Smart Images
Figure CN121934044A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target identification and positioning, specifically relating to a high-precision positioning method for sparse point clouds of airborne targets. Background Technology
[0002] To ensure safe flight, aircraft are often equipped with sensors such as lidar to detect surrounding obstacles and other aircraft, thus avoiding flight hazards. Accurately locating obstacles from the point cloud information detected by lidar is a key research focus. When the target is close and the point cloud information is abundant, the average point cloud coordinate method can be used to approximate the target's centroid with high accuracy. However, in airborne scenarios, for flight safety, it is necessary to identify distant targets. Due to limitations in lidar detection mechanisms and laser beam emission mechanisms, the point cloud for distant targets is sparse and unevenly distributed. Using the average point cloud method leads to large positioning errors and unreliable positioning results.
[0003] To address the localization technology of sparse point clouds of detection targets, this solution proposes a high-precision localization method for airborne target sparse point clouds. This method continuously learns the relationship between the distribution characteristics of lidar point clouds and the centroid of the target, enabling rapid localization of the target's position in airborne detection scenarios. Summary of the Invention
[0004] The purpose of this invention is to propose a high-precision positioning method for sparse point clouds of airborne targets. By receiving the sparse point cloud of the target from a distance detected by a lidar, the center position of the target can be quickly and accurately located, thereby improving the accuracy of the measurement.
[0005] A high-precision positioning method for sparse point clouds of airborne targets includes the following steps: S1, Receive the sparse point cloud of the target and complete the preprocessing; S2, build a point cloud feature extraction network to extract feature information of point cloud; S3, build the first point cloud voting network to gather the point cloud locations to the target center area; S4. Based on the point cloud positions obtained in S3, calculate the target centroid position after the first vote. S5. Build a second point cloud voting network to vote on the target centroid position obtained in S4, and calculate the target centroid position after the second vote as the final positioning result.
[0006] Furthermore, S1 receives the sparse point cloud of the target and performs preprocessing. Specifically, assuming that the number of sparse point clouds of the detected distant target is M, and the number of input point clouds of the point cloud feature extraction network is N, the purpose of preprocessing is to adjust the number of sparse point clouds to N.
[0007] Further, when the number of sparse point clouds M of the distant target is detected to be greater than N, the random downsampling method is used to adjust the number of sparse point clouds to N. Let the original sparse point cloud be , and the point cloud after downsampling is . The specific method is as follows: Set the downsampling step: ; The point cloud after downsampling: .
[0008] Further, when the number of sparse point clouds M of the distant target is detected to be less than N, the method of copying the original points is used to adjust the number of sparse point clouds to N. Let the original sparse point cloud be , and the point cloud after copying the original points is . The specific method is as follows: Set the multiple of point cloud copying: c = [N / M], and the copied point cloud Q contains c*M values; When c*M = N, no further processing is performed, and the generated Q is the adjusted point cloud; When c*M < N, calculate the difference ΔN = N - c*M, randomly downsample ΔN point clouds Q' from the original point cloud M, and merge them with the generated Q to obtain the adjusted point cloud.
[0009] Further, in S2, build a point cloud feature extraction network to extract the feature information of the point cloud, including: The point cloud feature extraction network uses the pointnet network to extract the global variables of the point cloud and finally outputs the global features of (N×1024) dimensions; the point cloud feature extraction network also has an intermediate output network that outputs the local features of (N×64) dimensions and the local features of (N×128) dimensions; The feature information extracted from the point cloud includes: obtaining the joint features of (N×1216) dimensions from the point cloud feature extraction network, and splicing the global features of (N×1024) dimensions, the local features of (N×64) dimensions and the local features of (N×128) dimensions to obtain the joint features of (N×1216) dimensions; according to the set feature dimension k, output the point cloud feature extraction result of (N×k) dimensions after dimensionality reduction.
[0010] Further, in S3, aggregate the point cloud positions to the target center, including: According to the point cloud features of (N×k) dimensions, calculate the first position deviation ΔX and the feature deviation ΔF of the original point cloud from the target center; the dimension of the first position deviation ΔX is N×3, and the dimension of the feature deviation is N×k; Calculate the position of the original point cloud after offset: X' = X + ΔX; Calculate the feature of the original point cloud after offset: F' = F + ΔF.
[0011] Furthermore, in S4, based on the point cloud positions obtained in S3, the target centroid position after the first vote is calculated, including: Calculate the target centroid position after the first vote:
[0012] Calculate the target feature results after the first vote: .
[0013] Furthermore, in S5, a second point cloud voting network is built to vote on the target centroid position from S4. Calculate the target centroid position after the first vote The offset ΔX from the second position of the target's centroid; Calculate the target centroid position after the second vote: .
[0014] The advantages and effects of this invention are: (1) To address the problem of inaccurate positioning of airborne sparse point clouds, a high-precision target centroid positioning method based on neural networks is proposed; (2) To address the situation where the sparse point cloud is insufficient to construct neural network input, a strategy of sparse point cloud replication is adopted to supplement the number of point clouds; (3) The high-precision target positioning strategy adopts the neural network method, which is more efficient and can meet the needs of real-time positioning calculation; (4) The high-precision target positioning strategy adopts the two-vote method, which results in higher target centroid positioning accuracy. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a schematic diagram illustrating the principle of a high-precision positioning method for sparse point clouds of airborne targets according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating a high-precision positioning method for sparse point clouds of airborne targets according to an embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] The features of various aspects of the embodiments of the present invention will now be described in detail. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can also be practiced without these specific details. The following description of the embodiments is merely intended to provide a better understanding of the invention by illustrating examples. The invention is not limited to any specific setups and methods provided below, but covers all improvements, substitutions, etc., to product structures and methods without departing from the spirit of the invention. In the various drawings and the following description, well-known structures and techniques are not shown to avoid unnecessarily obscuring the invention.
[0019] It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other, and the various embodiments can be referenced and cited in each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0020] Figure 1 This is a schematic diagram illustrating the principle of a high-precision positioning method for sparse point clouds of airborne targets according to an embodiment of the present invention.
[0021] like Figure 1 As shown, a high-precision positioning method for sparse point clouds of airborne targets includes the following steps: S1, Receive the sparse point cloud of the target and complete the preprocessing; S2, build a point cloud feature extraction network to extract feature information of point cloud; S3, build the first point cloud voting network to gather the point cloud locations to the target center area; S4. Based on the point cloud positions obtained in S3, calculate the target centroid position after the first vote. S5, build a second point cloud voting network to vote on the target center position of S4, and calculate the target centroid position after the second vote.
[0022] The target centroid position obtained from the second vote in S6 and S5 is the final positioning result.
[0023] like Figure 2As shown in the figure, a high-precision positioning method for sparse point clouds of airborne targets includes the following steps: S101, When the program starts, the algorithm receives the sparse point cloud of the target detected by the lidar. S102, Calculate the number of point clouds of the target, denoted as M, and set the number of input point clouds of the neural network to N.
[0024] S103, When the number of target point clouds is greater than the number of input point clouds of the neural network, that is, M≥N, use the method of random downsampling to reduce the number of target point clouds to the number of input point clouds N of the neural network. S104, When the number of target point clouds is less than the number of input point clouds of the neural network, that is, M<N, use the method of copying the original point cloud to expand the number of target point clouds to the number of input point clouds N of the neural network. S105, Update the neural network input data of the target sparse point cloud, and record the point cloud position data at this time as X, and the data format is (N×3). Here, 3 is the dimension, representing the coordinates (x, y, z).
[0025] S106, Build a point cloud feature extraction network and set the dimension k of the extracted point cloud features. S107, Through the feature extraction network in S106, extract the feature information of N points from the input X: (N×3) point cloud, denoted as F, and the output format of this network is (N×k). S108, Build the first point cloud voting network. S109, Through the point cloud voting network built in S108, process the X: (N×3) point cloud position coordinates in S105 and the F: (N×k) point cloud feature information in S107, and calculate the point cloud position X' and point cloud feature F' that gather in the target center area after voting, where the data format of X' is (N×3) and the data format of F' is (N×k). S110, Perform averaging processing on X' and F' obtained in S109, and the method is as follows:
[0026] where x represents the three-dimensional coordinates of the position or the k-dimensional information of the feature. The processed position information is denoted as and the feature information is denoted as ; S111, Build the second voting network. S112, Through the point cloud voting network built in S111, process the point cloud position coordinates and point cloud feature information, and calculate the centroid and the position coordinates of the first voting average again according to the point cloud features The offset ΔX; S113, Calculate the final centroid coordinates X. out :
[0027] S114, outputs the final sparse point cloud localization information X. out .
[0028] To address the problem of inaccurate airborne sparse point cloud positioning, this invention proposes a high-precision target centroid positioning method based on neural networks. For cases where the point cloud is too sparse to construct the neural network input, a sparse point cloud replication strategy is employed to supplement the point cloud quantity. The high-precision target positioning strategy utilizes a neural network approach, which is more efficient and can meet the requirements of real-time positioning calculations. Furthermore, the high-precision target positioning strategy employs a two-stage voting method, resulting in higher target centroid positioning accuracy.
[0029] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.
Claims
1. A high-precision positioning method for sparse point clouds of airborne targets, characterized in that: Includes the following steps: S1, Receive the sparse point cloud of the target and complete the preprocessing; S2, build a point cloud feature extraction network to extract feature information of point cloud; S3, build the first point cloud voting network to gather the point cloud locations to the target center area; S4. Based on the point cloud positions obtained in S3, calculate the target centroid position after the first vote. S5. Build a second point cloud voting network to vote on the target centroid position obtained in S4, and calculate the target centroid position after the second vote as the final positioning result.
2. The method according to claim 1, characterized in that, S1 receives the sparse point cloud of the target and performs preprocessing. Specifically, assuming that the number of sparse point clouds of the detected distant target is M, and the number of input point clouds of the point cloud feature extraction network is N, the purpose of preprocessing is to adjust the number of sparse point clouds to N.
3. The method according to claim 2, characterized in that, When the number of sparse point clouds M of a distant target is detected to be greater than N, a random downsampling method is used to adjust the number of sparse point clouds to N. Let the original sparse point cloud be... The downsampled point cloud is The specific method is as follows: Set the downsampling step size: ; Point cloud after downsampling: .
4. The method according to claim 2, characterized in that, When the number of sparse point clouds M of a distant target is detected to be less than N, the number of sparse point clouds is adjusted to N by copying the original points. Let the original sparse point cloud be... The point cloud after copying the original points is The specific method is as follows: Set the point cloud copy factor: c = [N / M], the copied point cloud Q contains c*M values; When c*M = N, no further processing is needed, and the generated Q is the adjusted point cloud; When c*M < N, calculate the difference ΔN = Nc*M, randomly downsample ΔN point clouds Q' from the original point cloud M, and merge them with the generated Q to obtain the adjusted point cloud.
5. The method according to claim 1, characterized in that, S2, Construct a point cloud feature extraction network to extract feature information from the point cloud, including: The point cloud feature extraction network uses the PointNet network to extract global variables of the point cloud and finally outputs (N×1024)-dimensional global features. The point cloud feature extraction network also has an intermediate output network that outputs (N×64)-dimensional local features and (N×128)-dimensional local features. The extraction of point cloud features includes: obtaining (N×1216) dimensional joint features from the point cloud feature extraction network; concatenating (N×1024) dimensional global features, (N×64) dimensional local features, and (N×128) dimensional local features to obtain (N×1216) dimensional joint features; and outputting the dimensionality-reduced (N×k) dimensional point cloud feature extraction result according to the set feature dimension k.
6. The method according to claim 1, characterized in that, S3, gathers the point cloud locations to the target center, including: Based on the (N×k) dimensional point cloud features, calculate the first position deviation ΔX and the feature deviation ΔF from the original point cloud to the target center; the dimension of the first position deviation ΔX is N×3, and the dimension of the feature deviation is N×k. Calculate the position after offsetting the original point cloud: X' = X + ΔX; Calculate the features after offsetting the original point cloud: F' = F + ΔF.
7. The method according to claim 1, characterized in that, S4, based on the point cloud positions obtained in S3, calculate the target centroid position after the first vote, including: Calculate the target centroid position after the first vote: Calculate the target feature results after the first vote: .
8. The method according to claim 1, characterized in that, S5: Build a second point cloud voting network to vote on the target centroid position from S4. Calculate the target centroid position after the first vote The offset ΔX from the second position of the target's centroid; Calculate the target centroid position after the second vote: .