CN-121978653-A - Multi-scale feature phase-preserving laser radar signal denoising model and ranging method

CN121978653ACN 121978653 ACN121978653 ACN 121978653ACN-121978653-A

Abstract

The invention relates to a multi-scale feature phase-preserving laser radar signal denoising model and a ranging method, wherein the denoising model comprises an input layer, a multi-scale feature encoder, a phase maintaining layer, a decoder and an output layer, wherein the input layer carries out convolution and activation function processing on a noise-containing laser radar echo signal, the multi-scale feature encoder carries out multi-scale feature extraction on local details and global periodicity of the signal through convolution layers with different expansion factors and different kernel sizes, then carries out fusion output on a multi-scale feature matrix, the phase maintaining layer carries out multi-layer recursion on the fused multi-scale feature matrix by adopting a gate control mechanism of GRU, restores local phase deviation, optimizes the consistency of the global period of the signal through dynamic adjustment of the dependency relationship between historical phase information and the current input phase, finally maintains the continuity of the phase, and the decoder gradually restores the local details and the global periodicity of the signal through up-sampling operation and reconstructs and outputs a noise-reducing signal with the same dimension as the input signal. And finally, performing phase difference ranging on the denoised echo signals and the transmitting signals.

Inventors

ZHANG ZHI
ZHAO HAO
LIN XULING
JIANG CHENG
XING KUN

Assignees

北京空间机电研究所

Dates

Publication Date: 20260505
Application Date: 20251215

Claims (10)

1. The laser radar signal denoising model integrating the multi-scale features and protecting the phase is characterized by comprising an input layer, a multi-scale feature encoder, a phase maintaining layer, a decoder and an output layer; the input layer is used for carrying out convolution and activation function processing on the echo signals of the noisy laser radar; The multi-scale feature encoder realizes multi-scale feature extraction of local details and global periodicity of signals through convolution layers with different expansion factors and different kernel sizes, and then performs fusion output on a multi-scale feature matrix; The phase maintaining layer adopts a gate control mechanism of GRU to carry out multi-layer recursion processing on the fused multi-scale feature matrix, and repairs local phase deviation and optimizes the consistency of the global period of the signal by dynamically adjusting the dependency relationship between the historical phase information and the current input phase, and finally maintains the continuity of the phase; A decoder for gradually recovering local details and global periodicity of the signal through an up-sampling operation; And the output layer is used for reconstructing and outputting a noise reduction signal with the same dimension as the input signal.
2. The laser radar signal denoising model of claim 1, wherein the processing of the input layer is as follows: Wherein, the Sampling data for the echo signals; Representing the convolution kernel of the input layer, For the input layer convolution operation, In order to activate the processing of the function, Is the output of the input layer.
3. The laser radar signal denoising model for fusing multi-scale feature phase preservation according to claim 1, wherein the multi-scale feature encoder comprises a multi-scale feature extraction layer and a feature fusion layer, the multi-scale feature extraction layer processes the output of an input layer by adopting convolution layers with different expansion factors and different kernel sizes to obtain features with different scales, and the feature fusion layer fuses and outputs a multi-scale feature matrix.
4. A multi-scale feature phase preserving fused lidar signal denoising model as claimed in claim 3, wherein the multi-scale feature extraction layer comprises three branches: The first branch is a big receptive field branch, and the processing process is as follows: Wherein, the For the output of the input layer, The output of the branch of the large receptive field; For the convolution kernel of the large receptive field branch, selecting a convolution kernel with an expansion factor of 31 multiplied by 1, wherein the expansion factor is 4; convolution operation for large receptive field branches; Processing for activating the function; The second branch is a middle receptive field branch, and the treatment process is as follows: Wherein, the Output for middle receptive field branches; For the convolution kernel of the middle receptive branch, selecting a 15 multiplied by 1 convolution kernel with an expansion factor, and selecting an expansion factor of 2; convolution operation for large receptive field branches; The third branch is a small receptive field branch, and the treatment process is as follows: Wherein, the For the output of the small receptive field branch, Selecting a convolution kernel with an expansion factor of 5 multiplied by 1 for the convolution kernel of the middle receptive branch, wherein the expansion factor is 1; Convolution operation for large receptive field branches.
5. The laser radar signal denoising model of claim 4, wherein the feature fusion layer is processed as follows: Wherein, the Representing branches to large receptive fields Output of middle receptive field branches Output of small receptive field branches Splicing is carried out in a mode of splicing according to rows, For the convolution sum of the feature fusion layer, a 5 x 1 convolution kernel is selected, For the convolution operation of the feature fusion layer, Processing for activating a function.
6. The multi-scale feature phase preserving fused lidar signal denoising model of claim 4, wherein the phase preserving layer comprises M hidden units, M being equal to or greater than 128, each hidden unit comprising an update layer, a reset gate; The process of updating the door is as follows: The process of resetting the gate is as follows: Wherein, the Wherein, the Is in a final hidden layer state, Is in a hidden state at the previous moment, Is candidate hidden layer state, 、、 Is a bias vector, 、、 Is a weight matrix, 。
7. The laser radar signal denoising model of claim 4, wherein the decoder comprises a signal restoration layer and a detail restoration layer, and the signal restoration layer is processed as follows: The signal reduction layer is processed as follows: the detail recovery layer is processed as follows: Wherein, the For the convolution kernel of the signal reduction layer, a convolution kernel of 5 x 1 is selected, For the convolution operation of the signal reduction layer, For the output of the signal reduction layer, For the convolution kernel of the middle signal reduction layer, a convolution kernel of 5 x 1 is selected, For the convolution operation of the signal reduction layer, Is the output of the signal reduction layer.
8. The laser radar signal denoising model of claim 4, wherein the output layer is processed as follows: Wherein, the To denoise the output signal For the convolution kernel of the output layer, a3 x1 convolution kernel is selected, Is a convolution operation of the output layer.
9. The phase-preserving lidar signal denoising model of claim 4, wherein the lidar signal denoising model is trained by using a sine wave as an ideal sample and adding a mixed noise of noise gaussian noise and poisson noise as a degenerate sample, and the model optimizes model parameters by minimizing a mean square error MSE between an input signal and an output signal during training.
10. A ranging method using the model of claim 1, characterized by comprising the steps of: S1, acquiring a laser radar emission signal S; S2, acquiring a laser radar echo signal S'; S3, sending the laser radar echo signal S ' into the laser radar signal denoising model fused with the multi-scale characteristic phase-preserving phase according to any one of claims 1-8 to obtain a denoised laser radar signal S ' '; S4, after carrying out Fourier transformation on the transmitting signal S and the denoised laser radar signal S '', extracting respective phase information to obtain a difference absolute value, thereby obtaining a phase difference between the transmitting signal and the echo signal ; S5, constructing a basic ranging formula: Will be Expressed as an integer number 2 And less than an integer number 2 A kind of electronic device And converting the basic ranging formula into a fuzzy ranging formula: ; s6, obtaining the result Take the minimum value Number and phase difference of (2) Substituting, calculating the distance to be measured by a fuzzy ranging formula , wherein, For the wavelength of the lidar signal, Is a preset test distance.

Description

Multi-scale feature phase-preserving laser radar signal denoising model and ranging method Technical Field The invention relates to a laser radar signal denoising model and a ranging method integrating multi-scale features and protecting phases, which aim to furthest reserve phase features, improve the precision of a phase difference method and further improve the ranging precision while denoising a laser radar echo signal through a deep learning technology. The invention has wide application in the fields of space target monitoring, environment detection, automatic driving, topographic mapping and the like. Background In 2017, xu Fan et al propose a method for denoising a laser radar echo signal based on synchronous compression transformation (CN 201710961744.3), which mainly considers that the signal denoising is realized by respectively performing continuous wavelet transformation on a transmitted laser pulse signal as a reference signal to obtain a wavelet time spectrum. In 2018, chang Jianhua et al proposed a method for denoising lidar echo signals based on variation modal decomposition (cn201810732305. X), which adopts a method for denoising lidar echo signals by variation modal decomposition. In 2024, feng Shuai et al, issued in "improved CEEMDAN combined with novel wavelet transform" and "Jidadenoise algorithm", published in "System engineering and electronic technology" and used a method of combining classical modal decomposition with wavelet transform image to denoise. The three methods are all denoising methods by adopting an empirical model modeling method, and the noise suppression is largely determined by the matching degree of signals and the model and the selection range of a threshold value. 2021, Minghuan Hu et al, issued "A Novel Lidar Signal Denoising Method Based on Convolutional Autoencoding Deep Learning Neural Network" to atmosphere, discloses denoising by deep learning from a coded neural network. However, the function of the model in the literature is only denoising without considering the change of denoising on the phase of the echo signal, and when the laser radar echo signal is denoised, the phase information in the signal is lost, and phase distortion is seriously possibly caused, which can reduce the distance accuracy calculated by adopting a phase difference method. Disclosure of Invention The technical problem solved by the invention is to overcome the defects of the prior art, provide the laser radar signal denoising model which integrates the multi-scale characteristics and protects the phase, solve the problem of laser radar echo denoising, and simultaneously maintain the phase information to the greatest extent, thereby improving the ranging accuracy of a phase difference method. The technical scheme of the invention is that the laser radar signal denoising model integrating the multi-scale characteristic phase-preserving comprises an input layer, a multi-scale characteristic encoder, a phase-preserving layer, a decoder and an output layer; the input layer is used for carrying out convolution and activation function processing on the echo signals of the noisy laser radar; The multi-scale feature encoder realizes multi-scale feature extraction of local details and global periodicity of signals through convolution layers with different expansion factors and different kernel sizes, and then performs fusion output on a multi-scale feature matrix; The phase maintaining layer adopts a gate control mechanism of GRU to carry out multi-layer recursion processing on the fused multi-scale feature matrix, and repairs local phase deviation and optimizes the consistency of the global period of the signal by dynamically adjusting the dependency relationship between the historical phase information and the current input phase, and finally maintains the continuity of the phase; A decoder for gradually recovering local details and global periodicity of the signal through an up-sampling operation; And the output layer is used for reconstructing and outputting a noise reduction signal with the same dimension as the input signal. Preferably, the processing procedure of the input layer is expressed as follows: Wherein, the Sampling data for the echo signals; Representing the convolution kernel of the input layer, For the input layer convolution operation,In order to activate the processing of the function,Is the output of the input layer. Preferably, the multi-scale feature encoder comprises a multi-scale feature extraction layer and a feature fusion layer, wherein the multi-scale feature extraction layer processes the output of an input layer by adopting convolution layers with different expansion factors and different kernel sizes to obtain features with different scales, and the feature fusion layer fuses and outputs the multi-scale feature matrix. Preferably, the multi-scale feature extraction layer comprises three branches: The first branch is a big receptive field branch, and the proce