CN-122023157-A - Satellite-borne single-photon laser radar point cloud self-supervision filtering method and system

CN122023157ACN 122023157 ACN122023157 ACN 122023157ACN-122023157-A

Abstract

The invention provides a satellite-borne single-photon laser radar point cloud self-supervision filtering method and a system, and relates to the technical field of laser radar data processing, wherein the filtering method comprises the following steps: by means of the core assumption that source point clouds and target point clouds which are split randomly in the same rough filtering point cloud slice are similar in distribution, a fully self-supervised deep learning filtering framework is constructed, and efficient denoising of the satellite-borne single-photon point clouds can be achieved without manual labeling or external priori knowledge. The method has the advantages that the statistical consistency of the point cloud is used as a supervision signal, the defect that structural noise with a specific mode cannot be effectively filtered by a traditional threshold method is overcome, the distribution characteristics of the point cloud are automatically learned through a deep learning model, the coordinate correction is carried out, the full-automatic processing is realized, the processing efficiency of mass satellite-borne point cloud data is improved, the finely filtered point cloud after the coordinate correction is directly output, and the topography detail can be better kept.

Inventors

YUAN MAN
Cao Bincai
YANG XIUCE
GONG HUI
HU YAN
LU XUELIANG
WANG SAI
ZHANG WEI
HOU HUITAI

Assignees

中国人民解放军61540部队

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (10)

1. The satellite-borne single-photon laser radar point cloud self-supervision filtering method is characterized by comprising the following steps of: Acquiring an original single photon point cloud to be filtered, and performing coarse filtering treatment on the original single photon point cloud to obtain a coarse filtering point cloud; Carrying out space slicing on the coarse filtering point clouds along the flight orbit direction of the laser radar to obtain a plurality of slice point clouds, and randomly splitting each slice point cloud into a source point cloud subset and a target point Yun Ziji; constructing a deep learning model, and training the deep learning model based on the source point cloud subset and the target point Yun Ziji in the same slice point cloud to minimize the distribution difference between the source point cloud subset and the target point cloud subset after model correction, so as to obtain a trained deep learning model; And carrying out coordinate correction on the coarse filtering point cloud based on the trained deep learning model to obtain a coordinate correction corresponding to each photon point in the coarse filtering point cloud, and determining a final fine filtering point cloud according to the coordinate correction.
2. The method for performing self-supervision filtering on the satellite-borne single-photon laser radar point cloud according to claim 1, wherein the performing coarse filtering on the original single-photon point cloud to obtain a coarse filtered point cloud comprises: preprocessing the original single photon point cloud to obtain a first correction point cloud; and performing coarse filtering processing on the first correction point cloud to obtain the coarse filtering point cloud.
3. The method for self-monitoring filtering of the point cloud of the on-board single-photon lidar according to claim 2, wherein the preprocessing the original single-photon point cloud to obtain a first correction point cloud comprises: Acquiring sea surface reference data of a corresponding region; removing photon points with heights positioned on the sea surface and above the sea surface from the original single photon point cloud based on the sea surface reference data to obtain a point cloud under the water surface; And carrying out physical refraction correction on the point cloud under the water surface to obtain the first correction point cloud.
4. The method for self-monitoring filtering of satellite-borne single-photon lidar point clouds according to claim 3, wherein the performing physical refraction correction on the subsurface point cloud to obtain the first corrected point cloud comprises: Acquiring the seawater refractive index of the corresponding region; And carrying out refractive error correction on the three-dimensional coordinates of the subsurface point cloud according to the sea water refractive index, the sea surface reference data and the geometric relationship between the satellite transmitting position and the photon receiving position based on a physical refractive index formula so as to generate the first correction point cloud.
5. The method for self-supervision filtering of the satellite-borne single-photon laser radar point cloud according to claim 1, wherein the deep learning model comprises an encoder and a decoder, the encoder is used for carrying out hash feature coding on three-dimensional coordinates of each photon point in the input source point cloud subset and outputting a high-dimensional feature vector, and the decoder is used for decoding the high-dimensional feature vector and outputting coordinate correction values of each photon point in the source point cloud subset.
6. The method for self-supervision filtering of the point cloud of the satellite-borne single-photon laser radar according to claim 5, wherein the encoder is a multi-stage hash encoder, each stage of hash encoder is provided with an independent hash table with a learnable parameter, and the decoder is a multi-layer perceptron model.
7. The method for self-supervision filtering of a point cloud of a satellite-borne single-photon lidar according to claim 1, wherein the training process of the deep learning model comprises: Sampling from the source point cloud subset and the target point cloud subset corresponding to the same slice point cloud respectively to obtain a source point coordinate and a target point coordinate for training; Inputting the training source point coordinates into the deep learning model to obtain predicted coordinate correction values; Correcting the source point coordinates for training according to the predicted coordinate correction value to obtain corrected source point coordinates; Calculating the earth movement distance between the corrected source point coordinates and the target point coordinates for training as a model training loss function; And optimizing model parameters based on the loss function to obtain the trained deep learning model.
8. The method of claim 7, wherein the loss function further comprises a regularization term, the regularization term being determined based on a gradient direction of the model output.
9. The method for self-monitoring filtering of a point cloud of a satellite-borne single-photon lidar according to claim 1, wherein the determining a final fine-filtered point cloud according to the coordinate correction comprises: adding the original three-dimensional coordinates of each photon point in the coarse filtering point cloud with the corresponding coordinate correction output by the depth learning model to obtain three-dimensional coordinates of each photon point after correction; And forming the final fine filtering point cloud by all the corrected three-dimensional coordinates.
10. The utility model provides a satellite-borne single photon laser radar point cloud self-supervision filter system which characterized in that includes: The device comprises an acquisition unit, a filtering unit and a filtering unit, wherein the acquisition unit is used for acquiring an original single photon point cloud to be filtered, and performing coarse filtering treatment on the original single photon point cloud to obtain a coarse filtering point cloud; the slicing unit is used for spatially slicing the coarse filtering point cloud along the flight orbit direction of the laser radar to obtain a plurality of slicing point clouds, and randomly splitting each slicing point cloud into a source point cloud subset and a target point Yun Ziji; The construction unit is used for constructing a deep learning model, training the deep learning model based on the source point cloud subset and the target point Yun Ziji in the same slice point cloud to minimize the distribution difference between the source point cloud subset and the target point cloud subset after model correction, and obtaining a trained deep learning model; And the processing unit is used for carrying out coordinate correction on the coarse filtering point cloud based on the trained deep learning model to obtain a coordinate correction corresponding to each photon point in the coarse filtering point cloud, and determining a final fine filtering point cloud according to the coordinate correction.

Description

Satellite-borne single-photon laser radar point cloud self-supervision filtering method and system Technical Field The invention relates to the technical field of laser radar data processing, in particular to a satellite-borne single-photon laser radar point cloud self-supervision filtering method and system. Background The single photon laser radar can realize the long-distance and weak reflection detection of the ground object target by the single photon detection sensitivity, and has important application value in the field of satellite-borne earth observation. However, the original single photon point cloud obtained by the method usually contains a large amount of interference signals introduced by atmospheric scattering, solar background noise, system jitter, multiple scattering and the like, and the density of the points is unevenly distributed, and the signal to noise ratio is low, so that the real earth surface signals must be separated through effective filtering treatment, which is a key precondition for generating high-precision standardized topography products. At present, filtering methods for such point clouds are mainly classified into two major categories, namely a threshold-based method and a learning-based method. Threshold-based methods (such as empirical, statistical or adaptive threshold filtering) set thresholds according to photon energy, density or local geometric statistics to reject outliers, simple but dependent on parameter empirical settings, have limited filtering effect on structural noise with specific modes (such as banding noise, clustered noise), and are prone to generating false filtering or residue due to local statistical mutation in regions with severe topography fluctuation. The method based on learning, especially the supervised deep learning method, classifies photons (such as ground points and noise points) through a training model, can learn complex noise modes, but the performance of the method is seriously dependent on a large number of training samples accurately marked, and the cost of manually marking the satellite-borne point cloud for the global scope and various landforms is extremely high, so that the method is not feasible. Disclosure of Invention The present invention solves one or more of the above-mentioned problems of the related art. In order to solve the problems, the invention provides a satellite-borne single-photon laser radar point cloud self-supervision filtering method and system. In a first aspect, the invention provides a satellite-borne single-photon laser radar point cloud self-supervision filtering method, which comprises the following steps: Acquiring an original single photon point cloud to be filtered, and performing coarse filtering treatment on the original single photon point cloud to obtain a coarse filtering point cloud; Carrying out space slicing on the coarse filtering point clouds along the flight orbit direction of the laser radar to obtain a plurality of slice point clouds, and randomly splitting each slice point cloud into a source point cloud subset and a target point Yun Ziji; constructing a deep learning model, and training the deep learning model based on the source point cloud subset and the target point Yun Ziji in the same slice point cloud to minimize the distribution difference between the source point cloud subset and the target point cloud subset after model correction, so as to obtain a trained deep learning model; And carrying out coordinate correction on the coarse filtering point cloud based on the trained deep learning model to obtain a coordinate correction corresponding to each photon point in the coarse filtering point cloud, and determining a final fine filtering point cloud according to the coordinate correction. Optionally, the performing coarse filtering processing on the original single photon point cloud to obtain a coarse filtering point cloud includes: preprocessing the original single photon point cloud to obtain a first correction point cloud; and performing coarse filtering processing on the first correction point cloud to obtain the coarse filtering point cloud. Optionally, the preprocessing the original single photon point cloud to obtain a first correction point cloud includes: Acquiring sea surface reference data of a corresponding region; removing photon points with heights positioned on the sea surface and above the sea surface from the original single photon point cloud based on the sea surface reference data to obtain a point cloud under the water surface; And carrying out physical refraction correction on the point cloud under the water surface to obtain the first correction point cloud. Optionally, the performing physical refraction correction on the subsurface point cloud to obtain the first corrected point cloud includes: Acquiring the seawater refractive index of the corresponding region; And carrying out refractive error correction on the three-dimensional coordinates of the subsu