Laser radar signal de-noising model fusing multi-scale feature to keep phase and ranging method

Abstract

The invention relates to a multi-scale feature-fused phase-preserving laser radar signal denoising model and a ranging method. The denoising model comprises an input layer, a multi-scale feature 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 noise-containing laser radar echo signal; the multi-scale feature encoder is used for realizing multi-scale feature extraction of local details and global periodicity of the signal through convolution layers with different expansion factors and different kernel sizes, and fusing and outputting a multi-scale feature matrix; the phase maintaining layer is used for performing multi-layer recursion processing on the fused multi-scale characteristic matrix by adopting a GRU gating mechanism, and finally maintaining the continuity of the phase by dynamically adjusting the dependency relationship between historical phase information and the current input phase, repairing local phase deviation and optimizing the consistency of the global period of the signal; the decoder is used for gradually recovering local details and global periodicity of the signal through up-sampling operation; and the output layer reconstructs and outputs a noise reduction signal with the same dimension as the input signal. And finally, carrying out phase difference method distance measurement on the de-noised echo signal and the emission signal.

Description

Technical Field

[0001] This invention relates to a lidar signal denoising model and ranging method that integrates multi-scale features while preserving phase. The aim is to use deep learning technology to denoise lidar echo signals while preserving phase features to the greatest extent possible, thereby improving the accuracy of the phase difference method and ultimately enhancing ranging accuracy. This invention has wide applications in fields such as space target monitoring, environmental detection, autonomous driving, and terrain mapping. Background Technology

[0002] In 2017, Xu Fan et al. proposed "A Denoising Method for LiDAR Echo Signals Based on Synchronous Compression Transform" (CN201710961744.3), which mainly considered using the emitted laser pulse signal as a reference signal to perform continuous wavelet transforms to obtain the wavelet time spectrum for signal denoising. In 2018, Chang Jianhua et al. proposed "A Denoising Method for LiDAR Echo Signals Based on Variational Mode Decomposition" (CN201810732305.X), which uses variational mode decomposition for LiDAR echo signal denoising. In 2024, Feng Shuai et al. published "An Improved LiDAR Denoising Algorithm Combining CEEMDAN with Novel Wavelet Transform" in *Systems Engineering and Electronics Technology*, which uses a combination of classical mode decomposition and wavelet transform for denoising.

[0003] All three methods described above employ an empirical model for noise reduction. The suppression of noise largely depends on the degree of matching between the signal and the model, as well as the range of threshold selection.

[0004] In 2021, Minghuan Hu et al. published "A Novel Lidar Signal Denoising Method Based on Convolutional Autoencoding Deep Learning Neural Network" in *Atmosphere*, which uses a deep learning autoencoding neural network for denoising. However, the model in the paper only denoises without considering the change in the phase of the echo signal caused by denoising. While denoising the lidar echo signal, it loses the phase information in the signal, which may cause phase distortion in severe cases. This will reduce the accuracy of the distance calculated using the phase difference method. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose a LiDAR signal denoising model that integrates multi-scale features to preserve phase, thereby solving the problem of LiDAR echo denoising and preserving phase information to the greatest extent, thus improving the ranging accuracy of the phase difference method.

[0006] The solution of the present invention is: a LiDAR signal denoising model that integrates multi-scale features and preserves phase, the denoising model including an input layer, a multi-scale feature encoder, a phase-preserving layer, a decoder, and an output layer; The input layer processes the noisy lidar echo signal through convolution and activation functions. The multi-scale feature encoder extracts multi-scale features of local details and global periodicity of a signal through convolutional layers with different dilation factors and kernel sizes, and then fuses and outputs the multi-scale feature matrices. The phase-preserving layer employs the gating mechanism of GRU to perform multi-level recursive processing on the fused multi-scale feature matrix. By dynamically adjusting the dependency between historical phase information and the current input phase, it repairs local phase deviations, optimizes the consistency of the global period of the signal, and ultimately maintains the continuity of the phase. The decoder gradually recovers the local details and global periodicity of the signal through upsampling operations; The output layer reconstructs the noise-reduced signal to the same dimension as the input signal.

[0007] Preferably, the processing procedure of the input layer is as follows:

[0008] in, This is the sampled data for the echo signal; Indicates the input layer convolution kernel. For input layer convolution operations, For activation function processing, This is the output of the input layer.

[0009] Preferably, the multi-scale feature encoder includes a multi-scale feature extraction layer and a feature fusion layer. The multi-scale feature extraction layer processes the output of the input layer using convolutional layers with different dilation factors and different kernel sizes to obtain features at different scales. The feature fusion layer fuses the multi-scale feature matrices and outputs the result.

[0010] Preferably, the multi-scale feature extraction layer comprises three branches: The first branch is the large receptive field branch, and the processing procedure is as follows:

[0011] in, The output of the input layer, The output of the large receptive field branch; For the convolution kernel of the large receptive field branch, a 31×1 convolution kernel with an inflation factor of 4 is selected. Convolution operation for the large receptive field branch; For activation function processing; The second branch is the mid-receptive field branch, and the processing procedure is as follows:

[0012] in, The output of the middle receptive field branch; For the convolution kernel of the mid-receptive field branch, a 15×1 convolution kernel with an inflation factor of 2 is selected. Convolution operation for the large receptive field branch; The third branch is the small receptive field branch, and the processing procedure is as follows:

[0013] in, For the output of the small receptive field branch, For the convolution kernel of the middle receptive field branch, a 5×1 convolution kernel with an inflation factor of 1 is selected. This is the convolution operation for the large receptive field branch.

[0014] Preferably, the feature fusion layer is processed as follows:

[0015]

[0016] in, Indicates the large receptive field branch Output of the middle receptive field branch Output of small receptive field branches Perform the splicing, using the row-by-row splicing method. For the convolution sum of the feature fusion layer, a 5×1 convolution kernel is selected. For the convolution operation of the feature fusion layer, This is for activation function processing.

[0017] Preferably, the phase-preserving layer includes M hidden units, where M is greater than or equal to 128, and each hidden unit includes an update layer and a reset gate; The door update process is as follows:

[0018] The procedure for resetting the door is as follows:

[0019] in,

[0020]

[0021] in, For the final hidden state, The hidden state at the previous moment, For candidate hidden state, , , For bias vector, , , For the weight matrix, .

[0022] Preferably, the decoder includes a signal restoration layer and a detail recovery layer, and the signal restoration layer processes data as follows: The signal restoration layer processes the signal as follows:

[0023] The detailed restoration layer is handled as follows:

[0024] in, For the convolution kernel of the signal reconstruction layer, a 5×1 convolution kernel is selected. For the convolution operation of the signal reconstruction layer, The output of the signal restoration layer, For the convolution kernel of the mid-signal reconstruction layer, a 5×1 convolution kernel is selected. For the convolution operation of the signal reconstruction layer, This is the output of the signal restoration layer.

[0025] Preferably, the output layer is processed as follows:

[0026] in, For the noise-reduced output signal For the output layer's convolution kernel, a 3×1 convolution kernel is selected. This is the convolution operation for the output layer.

[0027] Preferably, the lidar signal denoising model uses a sine wave as the ideal sample, and adds a mixture of Gaussian noise and Poisson noise as the degraded sample for training. During the training process, the model optimizes the model parameters by minimizing the mean square error (MSE) between the input signal and the output signal.

[0028] Another technical solution of the present invention is: a distance measurement method using the above-mentioned model, which includes the following steps: S1. Acquire the lidar transmission signal S; S2. Obtain the lidar echo signal S'; S3. Input the lidar echo signal S' into the lidar signal denoising model that fuses multi-scale features and preserves phase as described in any one of claims 1 to 8 to obtain the denoised lidar signal S''; S4. After performing Fourier transform on the transmitted signal S and the denoised lidar signal S'', extract their respective phase information, calculate the absolute value of the difference, and obtain the phase difference between the transmitted signal and the echo signal. ; S5. Construct the basic distance measurement formula: ,Will Represented as an integer number of 2s and fewer than an integer number of 2 of The summation transforms the basic distance measurement formula into a fuzzy distance measurement formula: ; S6. Find the condition that makes... Take the minimum value Number and phase difference Substitute the values ​​into the fuzzy ranging formula to calculate the distance to be measured. ,in, The wavelength of the lidar signal. The pre-set test distance.

[0029] The advantages of this invention compared to the prior art are: (1) The phase-preserving convolutional autoencoder neural network that integrates multi-scale features can more effectively preserve the local features of the signal and avoid phase distortion caused during the denoising process.

[0030] (2) The present invention can improve the signal-to-noise ratio of the lidar echo signal while preserving phase information to the greatest extent and improving ranging accuracy.

[0031] (3) This invention can be applied to fields such as space target monitoring, autonomous driving, and terrain mapping.

[0032] (4) The model of this invention adopts a multi-scale feature fusion method to expand the receptive field, which can enhance the ability to extract local features in the echo signal. At the same time, the phase preservation layer is used to preserve the phase features to the greatest extent, which can better improve the ranging accuracy calculated by the phase difference method. Attached Figure Description

[0033] Figure 1 This invention provides a phase-preserving lidar signal denoising model that integrates multi-scale features in an embodiment of the present invention.

[0034] Figure 2 This invention provides a comparison of the post-denoising performance indicators of various denoising methods in its embodiments. Figure 3(a) shows the echo signal waveform of the noisy lidar in an embodiment of the present invention; Figure 3(b) shows the waveform of the denoised lidar echo signal obtained by the Buttw method in an embodiment of the present invention; Figure 3(c) shows the waveform of the denoised lidar echo signal obtained by the wavlt method in an embodiment of the present invention; Figure 3(d) shows the waveform of the denoised lidar echo signal obtained by the tvemd method in an embodiment of the present invention; Figure 3(e) shows the waveform of the denoised lidar echo signal obtained by the VOAMD method in an embodiment of the present invention; Figure 3(f) shows the waveform of the denoised lidar echo signal obtained by the CEEMD-SF method in an embodiment of the present invention; Figure 3(g) shows the waveform of the denoised lidar echo signal obtained by using the model of the present invention in an embodiment of the present invention; Figure 4(a) shows the SNR of the model in the embodiment of the present invention after 1000 tests; Figure 4(b) shows the RMSE of an embodiment of the present invention; Figure 4(c) shows the ranging accuracy of the embodiment of the present invention (echo signal SNR: 5dB). Detailed Implementation

[0035] The present invention will be further described below with reference to the embodiments.

[0036] 1. Model Building like Figure 1 As shown, the present invention provides a lidar signal denoising model that integrates multi-scale features and preserves phase. The denoising model includes an input layer, a multi-scale feature encoder, a phase-preserving layer, a decoder, and an output layer. The input layer processes the noisy lidar echo signal through convolution and activation functions. The multi-scale feature encoder extracts multi-scale features of local details and global periodicity of a signal through convolutional layers with different dilation factors and kernel sizes, and then fuses and outputs the multi-scale feature matrices. The phase-preserving layer employs the gating mechanism of GRU to perform multi-level recursive processing on the fused multi-scale feature matrix. By dynamically adjusting the dependency between historical phase information and the current input phase, it repairs local phase deviations, optimizes the consistency of the global period of the signal, and ultimately maintains the continuity of the phase. The decoder gradually recovers the local details and global periodicity of the signal through upsampling operations; The output layer reconstructs the noise-reduced signal to the same dimension as the input signal.

[0037] (1) Input layer The processing procedure of the input layer is as follows:

[0038] in, This is the sampled data for the echo signal; Indicates the input layer convolution kernel. For input layer convolution operations, For activation function processing, This is the output of the input layer.

[0039] In a specific embodiment of the present invention, 7×1 convolution kernel , For convolution kernel weights, Features of the current time and its six previous time points, .

[0040] The ReLU activation function is used after each convolutional layer to enhance the model's non-linear expressiveness and avoid the gradient vanishing problem.

[0041] In this embodiment, the input dimension of the input layer is [1×800] (the signal has 800 sampling points).

[0042] (2) Multiscale encoder with expanded receptive field The multi-scale encoder, which expands the receptive field, uses convolutional layers with different dilation factors and kernel sizes to extract local details and global periodicity from periodic signals at multiple scales. By combining dilated convolutions with multi-scale feature extraction and feature fusion layers, it extracts deep features of the signal layer by layer, gradually compresses the signal dimension, and finally generates a low-dimensional feature vector.

[0043] The specific design is as follows: The multi-scale feature encoder includes a multi-scale feature extraction layer and a feature fusion layer. The multi-scale feature extraction layer processes the output of the input layer using convolutional layers with different dilation factors and different kernel sizes to obtain features at different scales. The feature fusion layer fuses the multi-scale feature matrices and outputs the result.

[0044] (2.1) Multi-scale feature extraction layer The multi-scale feature extraction layer contains three branches: The first branch is the large receptive field branch, and the processing procedure is as follows:

[0045] in, The output of the input layer, The output of the large receptive field branch; For the convolution kernel of the large receptive field branch, a 31×1 convolution kernel with an inflation factor of 4 is selected. Convolution operation for the large receptive field branch; For activation function processing; in, , For convolution kernel, The expansion factor for the receptive field is set to 4 here. The second branch is the mid-receptive field branch, and the processing procedure is as follows:

[0046] in, The output of the middle receptive field branch; For the convolution kernel of the mid-receptive field branch, a 15×1 convolution kernel with an inflation factor of 2 is selected. Convolution operation for the large receptive field branch; in, , For convolution kernel, The expansion factor of the receptive field is set to 2 here.

[0047] The third branch is the small receptive field branch, and the processing procedure is as follows:

[0048] in, For the output of the small receptive field branch, For the convolution kernel of the middle receptive field branch, a 5×1 convolution kernel with an inflation factor of 1 is selected. Convolution operation for the large receptive field branch: in, , For convolution kernel, The expansion factor of the receptive field is set to 1 here.

[0049] Therefore, the structure of the multi-scale feature extraction layer is as follows: kernel size: 31 (dilation factor: 4), 15 (dilation factor: 2), 5; The function is as follows: Through multi-layer convolutional operations, the encoder progressively extracts local features (such as peak and trough shapes) and global periodic features (such as phase shifts) of the signal. Multi-scale feature extraction of the signal is achieved within the encoder. The encoder uses convolutional layers with different dilation factors (31×1, 15×1, 5×1) to cover local details and global periodic features. Feature concatenation (concat1) fuses the three multi-scale features (low, medium, and high-frequency signal features) to form a rich feature representation. A dilated convolutional design that expands the receptive field is adopted to avoid signal truncation caused by traditional pooling operations.

[0050] (2.2) Feature Fusion Layer The feature fusion layer is processed as follows:

[0051]

[0052] in, Indicates the large receptive field branch Output of the middle receptive field branch Output of small receptive field branches Perform the splicing, using the row-by-row splicing method. For the convolution sum of the feature fusion layer, a 5×1 convolution kernel is selected. For the convolution operation of the feature fusion layer, This is for activation function processing. This is the result of multi-scale feature fusion.

[0053] Therefore, the structure of the feature fusion layer is as follows: ; Its function is to splice multi-scale features and normalize them.

[0054] (3) Phase preservation layer The periodicity of an ideal input sinusoidal signal depends on the continuity of the time series.

[0055] The phase preservation layer includes M hidden units, where M is greater than or equal to 128, and each hidden unit includes an update layer and a reset gate. The door update process is as follows:

[0056] The procedure for resetting the door is as follows:

[0057] The final state is:

[0058] Candidate state:

[0059] in, For the final hidden state, The hidden state at the previous moment, For candidate hidden state, , , For bias vector, , , For the weight matrix, .

[0060] This layer combines the gating mechanism of GRU (update gate + reset gate) to memorize the phase information of the previous cycle, thereby avoiding phase breakage caused by noise. By dynamically adjusting the dependency between historical phase information and the current input phase, the continuity of phase is ultimately maintained.

[0061] The characteristics of the phase-preserving layer are as follows: (1) Dynamic weight adjustment: The GRU layer gradually corrects noise interference through multi-layer recursive processing. The two-layer convolution design can better preserve the phase information in the echo signal. The first GRU layer repairs local phase deviations, and the second GRU layer optimizes global periodic consistency. (2) Handling long dependencies: GRU's gating mechanism can capture dependencies spanning multiple cycles (such as phase shifts across multiple complete cycles), which is particularly important for continuous ranging signals of lidar.

[0062] (4) Decoder The decoder consists of two deconvolutional layers (transposed convolutional layers), including a signal restoration layer and a detail recovery layer. The signal restoration layer processes data as follows: The signal restoration layer processes the signal as follows:

[0063] The detailed restoration layer is handled as follows:

[0064] in, For the convolution kernel of the signal reconstruction layer, a 5×1 convolution kernel is selected. For the convolution operation of the signal reconstruction layer, The output of the signal restoration layer, For the convolution kernel of the mid-signal reconstruction layer, a 5×1 convolution kernel is selected. For the convolution operation of the signal reconstruction layer, This is the output of the signal restoration layer.

[0065] The function of a decoder is to gradually recover the local details (such as the steepness of the waveform) and global periodicity (such as phase continuity) of a signal through deconvolution operations.

[0066] (5) Output layer The output layer is processed as follows:

[0067] in, For the noise-reduced output signal For the output layer's convolution kernel, a 3×1 convolution kernel is selected. This is the convolution operation for the output layer.

[0068] In this embodiment, , These are the kernel weights; Features of the current time and the two time points preceding it.

[0069] Output layer structure: final_conv2: 3×1 convolution; output: Regression output layer, using the mean squared error (MSE) loss function.

[0070] Its function is to reconstruct the noise-reduced signal with the same dimension as the input.

[0071] (6) Design of other parameters of the model Batch size: 32; Initial learning rate: 0.0005; Epochs: 30; Adopt an early shutdown mechanism; The optimization algorithm uses the Adam optimization method.

[0072] 2. Model Training The training set sample composition and parameter design are as follows: The LiDAR signal denoising model uses a sine wave as the ideal sample and adds a mixture of Gaussian noise and Poisson noise as the degraded sample for training. During the training process, the model optimizes the model parameters by minimizing the mean square error (MSE) between the input and output signals.

[0073] To improve the robustness of the model, the sample set uses a mixture of Gaussian noise (mainly -5~30dB) and Poisson noise (gain 50~2000). The input to the model in this paper is a noisy lidar echo signal, and the output is the denoised signal. During training, the model optimizes the network parameters by minimizing the mean square error (MSE) between the input and output signals.

[0074] In a specific embodiment of the present invention, since the transmitted signal is, .in, f is the amplitude of the signal modulation, and f is the modulation frequency. The initial phase of the signal is defined as . Therefore, the signals in the sample library are mainly sine waves, representing ideal samples, with a mixture of Gaussian and Poisson noise added as degenerate samples. To improve the robustness of the model, the sample set uses a mixture of Gaussian noise (mainly -5~30dB) and Poisson noise (gain 50~2000), as detailed below: Gaussian noise simulation: 80% of the samples are concentrated in the medium noise intensity of 3~5dB (simulating the actual ranging scenario), and 20% of the samples cover the extreme case of -10dB~30dB, which improves the generalization ability of the model.

[0075] Poisson noise simulation: The Poisson gain range of 50~2000 covers different photon counting scenarios (such as low gain in low light environment and high gain in strong light environment), enhancing the model's adaptability to photon noise.

[0076] Regularization and gradient clipping: L2 regularization prevents the model from overfitting noisy features, with the regularization parameter set to 0.0002; gradient clipping avoids gradient explosion caused by noise, with the threshold set to 1 to stabilize the training process.

[0077] 3. Phase difference method for distance measurement (1) Phase difference calculation: S1. Obtain the lidar transmitted signal S as follows:

[0078] S2. Obtain the noisy lidar echo signal S' as follows:

[0079] Among them, the transmitted signal S and the noisy lidar echo signal S', f is the amplitude of the signal modulation, and f is the modulation frequency. This represents the initial phase of the signal.

[0080] S3. Input the lidar echo signal S' into the lidar signal denoising model that fuses multi-scale features and preserves phase, and obtain the denoised lidar signal S''. S4. After performing Fourier transform on the transmitted signal S and the denoised lidar signal S'', extract their respective phase information, calculate the absolute value of the difference, and obtain the phase difference between the transmitted signal and the echo signal. ; This represents the phase difference between the transmitted signal and the echo signal.

[0081]

[0082] in, and The phase information is extracted after performing Fourier transforms on the transmitted signal S and the echo signal S', respectively.

[0083] (2) Distance measurement using the phase difference method Due to the distance to be measured .because , ,therefore t is the time it takes for the signal to travel from the emission point to the target and back to the lidar detector, and c is the speed of light in air.

[0084] Constructing the basic distance measurement formula: ,Will Represented as an integer number of 2s and fewer than an integer number of 2 of The summation transforms the basic distance measurement formula into a fuzzy distance measurement formula: ; S6. Find the condition that makes... Take the minimum value Number and phase difference Substitute the values ​​into the fuzzy ranging formula to calculate the distance to be measured. ,in, The wavelength of the lidar signal.

[0085] Figure 2 This paper compares the denoising performance of various denoising methods according to embodiments of the present invention. The results of denoising by various methods under typical operating conditions (SNR = 5 dB) are shown in the table below: Table 1. Noise reduction metrics after various denoising methods (echo signal SNR = 5dB)

[0086] Figures 3(a) to 3(f) show a comparison of the waveforms after denoising using various denoising methods of the present invention.

[0087] The denoising results of this model applied to 1000 test signals are shown in Figures 4(a) to 4(c). According to the statistical results, under the conditions of a typical SNR of 5dB, a repetition frequency of 100MHz, a set measurement distance of 1000m, and a background illuminance of 1353 W / m2, after denoising using the model proposed in this invention, the SNR can be improved from 5dB to 21.8dB~22.5dB, the RMSE can be reduced to 0.053~0.057, and the ranging accuracy can be achieved to 11.8cm~15.6cm.

[0088] The method of this invention integrates a phase-preserving convolutional autoencoder neural network with multi-scale features, which can more effectively preserve the local features of the signal and avoid phase distortion caused during denoising. This invention can improve the signal-to-noise ratio of lidar echo signals while preserving phase information to the greatest extent, thereby enhancing ranging accuracy. It can be widely applied in fields such as space target monitoring, environmental detection, autonomous driving, and terrain mapping.

[0089] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A phase-preserving lidar signal denoising model that integrates multi-scale features, characterized in that... It includes an input layer, a multi-scale feature encoder, a phase-preserving layer, a decoder, and an output layer; The input layer processes the noisy lidar echo signal through convolution and activation functions. The multi-scale feature encoder extracts multi-scale features of local details and global periodicity of a signal through convolutional layers with different dilation factors and kernel sizes, and then fuses and outputs the multi-scale feature matrices. The phase-preserving layer employs the gating mechanism of GRU to perform multi-level recursive processing on the fused multi-scale feature matrix. By dynamically adjusting the dependency between historical phase information and the current input phase, it repairs local phase deviations, optimizes the consistency of the global period of the signal, and ultimately maintains the continuity of the phase. The decoder gradually recovers the local details and global periodicity of the signal through upsampling operations; The output layer reconstructs the noise-reduced signal to the same dimension as the input signal.

2. The lidar signal denoising model based on phase preservation of multi-scale features according to claim 1, characterized in that, The processing procedure of the input layer is as follows: in, This is the sampled data for the echo signal; Indicates the input layer convolution kernel. For input layer convolution operations, For activation function processing, This is the output of the input layer.

3. The lidar signal denoising model based on phase preservation of multi-scale features according to claim 1, characterized in that, The multi-scale feature encoder includes a multi-scale feature extraction layer and a feature fusion layer. The multi-scale feature extraction layer uses convolutional layers with different dilation factors and different kernel sizes to process the output of the input layer to obtain features at different scales. The feature fusion layer fuses and outputs multi-scale feature matrices.

4. The lidar signal denoising model that integrates multi-scale features while preserving phase, as described in claim 3, is characterized in that... The multi-scale feature extraction layer contains three branches: The first branch is the large receptive field branch, and the processing procedure is as follows: in, The output of the input layer, The output of the large receptive field branch; For the convolution kernel of the large receptive field branch, a 31×1 convolution kernel with an inflation factor of 4 is selected. Convolution operation for the large receptive field branch; For activation function processing; The second branch is the mid-receptive field branch, and the processing procedure is as follows: in, The output of the middle receptive field branch; For the convolution kernel of the mid-receptive field branch, a 15×1 convolution kernel with an inflation factor of 2 is selected. Convolution operation for the large receptive field branch; The third branch is the small receptive field branch, and the processing procedure is as follows: in, For the output of the small receptive field branch, For the convolution kernel of the middle receptive field branch, a 5×1 convolution kernel with an inflation factor of 1 is selected. This is the convolution operation for the large receptive field branch.

5. The lidar signal denoising model based on phase preservation of multi-scale features according to claim 4, characterized in that, The feature fusion layer is processed as follows: in, Indicates the large receptive field branch Output of the middle receptive field branch Output of small receptive field branches Perform the splicing, using the row-by-row splicing method. For the convolution sum of the feature fusion layer, a 5×1 convolution kernel is selected. For the convolution operation of the feature fusion layer, This is for activation function processing.

6. The lidar signal denoising model based on phase preservation of multi-scale features according to claim 4, characterized in that, The phase preservation layer includes M hidden units, where M is greater than or equal to 128, and each hidden unit includes an update layer and a reset gate. The door update process is as follows: The procedure for resetting the door is as follows: in, in, For the final hidden state, The hidden state at the previous moment, For candidate hidden state, , , For bias vector, , , For the weight matrix, .

7. A phase-preserving lidar signal denoising model fusing multi-scale features according to claim 4, characterized in that, The decoder includes a signal restoration layer and a detail recovery layer. The signal restoration layer processes data as follows: The signal restoration layer processes the signal as follows: The detailed restoration layer is handled as follows: in, For the convolution kernel of the signal reconstruction layer, a 5×1 convolution kernel is selected. For the convolution operation of the signal reconstruction layer, The output of the signal restoration layer, For the convolution kernel of the mid-signal reconstruction layer, a 5×1 convolution kernel is selected. For the convolution operation of the signal reconstruction layer, This is the output of the signal restoration layer.

8. A phase-preserving lidar signal denoising model fusing multi-scale features according to claim 4, characterized in that, The output layer is processed as follows: in, For the noise-reduced output signal For the output layer's convolution kernel, a 3×1 convolution kernel is selected. This is the convolution operation for the output layer.

9. A phase-preserving lidar signal denoising model fusing multi-scale features according to claim 4, characterized in that, The LiDAR signal denoising model uses a sine wave as the ideal sample and adds a mixture of Gaussian noise and Poisson noise as the degraded sample for training. During the training process, the model optimizes the model parameters by minimizing the mean square error (MSE) between the input and output signals.

10. The distance measurement method using the model described in claim 1, characterized in that... Includes the following steps: S1. Acquire the lidar transmission signal S; S2. Obtain the lidar echo signal S'; S3. Input the lidar echo signal S' into the lidar signal denoising model that fuses multi-scale features and preserves phase as described in any one of claims 1 to 8 to obtain the denoised lidar signal S''; S4. After performing Fourier transform on the transmitted signal S and the denoised lidar signal S'', extract their respective phase information, calculate the absolute value of the difference, and obtain the phase difference between the transmitted signal and the echo signal. ; S5. Construct the basic distance measurement formula: ,Will Represented as an integer number of 2s and fewer than an integer number of 2 of The summation transforms the basic distance measurement formula into a fuzzy distance measurement formula: ; S6. Find the condition that makes... Take the minimum value Number and phase difference Substitute the values ​​into the fuzzy ranging formula to calculate the distance to be measured. ,in, The wavelength of the lidar signal. The pre-set test distance.

Images