Laser radar anti-interference system and method based on graph convolutional neural network
By processing lidar signals using graph convolutional neural networks, the problem of anti-interference in complex environments is solved, achieving efficient denoising and feature extraction, improving the signal-to-noise ratio and meeting real-time requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-07
- Publication Date
- 2026-04-07
AI Technical Summary
Existing lidar signal processing methods struggle to adapt to dynamic interference in complex environments, cannot effectively distinguish between real targets and noise, and lack an understanding of the global context of the signal, resulting in insufficient anti-interference performance.
A graph convolutional neural network is used to process the one-dimensional echo signal of lidar. Through pulsed laser polarization Monte Carlo simulation, full waveform decomposition and starfish optimization algorithm, graph structure construction and GCN-GAT stacking module, the signal is denoised and feature extracted, while preserving the spatiotemporal correlation and global features of the signal.
It improves the anti-interference capability of LiDAR in complex environments, maintains a signal-to-noise ratio improvement of more than 15dB, and controls the latency to within 10ms, making it suitable for applications with high real-time requirements such as autonomous driving.
Smart Images
Figure CN121805985A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lidar signal processing technology, specifically to a lidar anti-interference system and method based on graph convolutional neural networks. Background Technology
[0002] One-dimensional echo signals from lidar face severe challenges in practical applications due to complex environmental interference. This interference mainly originates from natural environmental factors (such as the scattering effect of precipitation particles from rain, snow, fog, and haze on the laser) and dynamic obstacles (such as multipath reflections caused by moving vehicles and pedestrians). Traditional signal processing methods, such as wavelet transform and Kalman filtering, perform reasonably well in stationary noise environments, but their limitations become increasingly apparent when dealing with non-stationary noise overlapping with the target signal spectrum.
[0003] The existing technologies have the following main problems: on the one hand, filtering strategies based on fixed thresholds or preset frequency bands are difficult to adapt to dynamically changing interference patterns; on the other hand, these methods have limited ability to extract local features of signals, especially when weak target signals (such as echoes from small objects at a distance) coexist with strong noise (such as reflections from dense rain and snow), which often leads to the target signal being mistakenly filtered out or too much noise remaining.
[0004] While emerging deep learning models (such as Convolutional Neural Networks) have improved feature representation capabilities through end-to-end learning, their inherent limitations are equally prominent. The local receptive field design of standard CNNs essentially processes Euclidean space data, while the temporal correlations (such as motion correlations between consecutive frames of a target) and non-uniform sampling characteristics (such as non-uniform angular sampling caused by radar rotation scanning) inherent in one-dimensional echo signals are difficult to model effectively. This leads to the network potentially destroying crucial spatiotemporal structural information of the signal during noise suppression.
[0005] Furthermore, existing methods generally lack an understanding of the global context of the signal. For example, they cannot effectively distinguish between the continuous trajectory of a real target and the transient peaks of random noise. This deficiency is particularly evident in dynamic occlusion scenarios. These problems collectively limit the noise suppression performance of existing technologies in complex scenarios, necessitating a novel processing method that can simultaneously model the spatiotemporal characteristics of the signal, adaptively distinguish noise from the target, and maintain physical interpretability. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for anti-interference of lidar based on graph convolutional neural networks. By remodeling traditional one-dimensional time-domain signals into graph structure data, the invention fully utilizes the unique advantages of graph convolutional neural networks in processing non-Euclidean spatial data, effectively improving the anti-interference capability of lidar in complex environments.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] On one hand, the present invention provides a lidar anti-interference system based on graph convolutional neural networks, comprising:
[0009] The pulsed laser polarization Monte Carlo simulation module is used to study the transmission characteristics of pulsed lasers in smoke environments, including the composition, scale distribution, volume optical properties, and geometric influence of smoke particles on laser transmission.
[0010] The echo signal denoising module uses full waveform decomposition and starfish optimization algorithm to denoise the original echo signal and achieves accurate fitting of Gaussian sub-waveforms to extract the characteristic echoes of smoke echoes and targets.
[0011] The graph construction module, based on the denoised echo signal, uses three polarized echoes (0°, 45°, 135°) as nodes to construct a graph structure;
[0012] The graph convolutional neural network module is used to process the node graph after the graph is constructed. Based on the attention mechanism of Transformer and the GCN-GAT stacking module, it realizes the extraction of node features and the differential aggregation of information.
[0013] The anti-interference output module reconstructs the signal from the result processed by the graph convolutional neural network and outputs the anti-interference signal and noise distribution.
[0014] Preferably, the above-mentioned pulsed laser polarization Monte Carlo simulation module uses Stokes vectors and Mueller matrices to establish a polarization Monte Carlo simulation model, and identifies the differences in echo characteristics under different polarization states through optical path analysis and correlation coefficient verification.
[0015] Preferably, the full waveform decomposition in the above-mentioned echo signal denoising module uses variational mode decomposition algorithm to denoise the original echo, and constructs a dual threshold criterion through curvature extremum detection and gradient field analysis to accurately identify the peak position and neighborhood inflection point distribution characteristics of the Gaussian component.
[0016] Preferably, the starfish optimization algorithm in the above echo signal denoising module mimics the starfish's behavior of turning over rocks to find prey. The starfish are divided into explorers, developers, and collaborators according to their roles, and the explorer position update formula is:
[0017] ;
[0018] Where, in the formula, , These represent the current optimal individual position and the random individual position, respectively; rand is a random number uniformly distributed in [0,1]. , To control the weighting coefficients between global and random searches.
[0019] Preferably, the formula for constructing the graph structure using the above-mentioned graph construction module is:
[0020] ;
[0021] Where σ() is the activation function, Represents the node features of the Lth layer. To add self-connected adjacency matrices; for The degree matrix; Let be the learnable weight matrix for layer (l).
[0022] Preferably, the GCN-GAT stacked structure in the above graph convolutional neural network module first extracts local features of nodes through a graph convolutional network, and then realizes differentiated aggregation of information between nodes through a graph attention network.
[0023] On the other hand, the present invention provides a lidar anti-interference method based on graph convolutional neural networks, comprising the following steps:
[0024] (1) Obtain the original one-dimensional echo signal of the lidar;
[0025] (2) The transmission characteristics of laser in a smoke environment were studied by pulsed laser polarization Monte Carlo simulation;
[0026] (3) The echo signal is denoised using the full waveform decomposition and starfish optimization algorithm;
[0027] (4) Based on the denoised echo signal, a graph structure is constructed with three polarized echoes (0°, 45°, 135°) as nodes;
[0028] (5) Input the constructed graph structure into a graph convolutional neural network based on the Transformer attention mechanism and GCN-GAT stack;
[0029] (6) Reconstruct the signal from the network output and output the anti-interference signal and noise distribution.
[0030] Preferably, the above-mentioned pulsed laser polarization Monte Carlo simulation includes:
[0031] The photon polarization state was characterized using the Stokes parameter method;
[0032] The Mueller matrix is used to describe the changes in polarized light during the scattering process;
[0033] Simulation analysis identifies that the 90° polarization state has fewer object echo features in the signal due to the depolarization effect of multiple scattering, thus eliminating the 90° polarization state in subsequent processing.
[0034] Preferably, the denoising process combining the above-mentioned full waveform decomposition with the starfish optimization algorithm includes:
[0035] The original echo was denoised using the variational mode decomposition algorithm to obtain denoised echo data;
[0036] A dual threshold criterion is constructed by combining curvature extremum detection and gradient field analysis to identify the peak position of Gaussian components;
[0037] The variational mode decomposition parameters are adaptively optimized using the starfish optimization algorithm.
[0038] Gaussian fitting is performed on the denoised signal to extract the characteristic parameters of the smoke echo and the target echo.
[0039] Preferably, the above graph convolutional neural network processing includes:
[0040] After removing the 90° polarization state from the smoke and target polarization echo obtained by full waveform decomposition, the three polarization state waveforms under the same label are used as nodes to construct a subgraph.
[0041] First, the input signal is subjected to global feature extraction through the Transformer attention mechanism to capture long-range dependencies;
[0042] Then, the GCN-GAT stacking module is used to extract local features of the nodes in the graph, and the mutual influence between different polarization states is fused through the graph attention mechanism.
[0043] The features are mapped to the output vector through a fully connected layer, and the Softmax function is used for normalization to obtain the classification result.
[0044] The beneficial effects of this invention are as follows:
[0045] 1. Innovatively applying graph convolutional networks to one-dimensional signal processing of lidar effectively solves the problem of anti-interference in complex environments using traditional methods;
[0046] 2. Graph structure modeling fully preserves the spatiotemporal correlation characteristics of the signal, and multi-scale feature learning enables comprehensive perception from local details to global context;
[0047] 3. The adaptive filtering mechanism ensures stable performance under different interference scenarios, and can still maintain a signal-to-noise ratio improvement of more than 15dB in strong noise environment;
[0048] 4. The processing latency is controlled within 10ms, providing a reliable technical guarantee for applications with high real-time requirements such as autonomous driving;
[0049] 5. The system adopts a modular design with reasonable allocation of computing resources. The GPU is used only for network inference, while the remaining processing is completed by the CPU, which significantly reduces hardware costs. Attached Figure Description
[0050] Figure 1 This is a basic flowchart of a LiDAR anti-interference method based on graph convolutional neural networks;
[0051] Figure 2 This is a schematic diagram of the full waveform decomposition;
[0052] Figure 3 It is a graph convolutional neural network architecture;
[0053] Figure 4 This is a schematic diagram of the anti-interference algorithm. Detailed Implementation
[0054] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0055] Example 1:
[0056] This embodiment provides a lidar anti-interference system based on graph convolutional neural networks. The system is mainly divided into five modules: pulsed laser polarization Monte Carlo simulation module, echo signal denoising module, mapping module, graph convolutional neural network module, and anti-interference output module.
[0057] The pulsed laser polarization Monte Carlo simulation module is used to study the transmission characteristics of pulsed lasers in smoke environments, including the composition, scale distribution, volume optical properties, and geometric influence of smoke particles on laser transmission.
[0058] The echo signal denoising module uses full waveform decomposition and starfish optimization algorithm to denoise the original echo signal and achieves accurate fitting of Gaussian sub-waveforms to extract the characteristic echoes of smoke echoes and targets.
[0059] The graph construction module, based on the denoised echo signal, uses three polarized echoes (0°, 45°, 135°) as nodes to construct a graph structure;
[0060] The graph convolutional neural network module is used to process the node graph after the graph is constructed. Based on the attention mechanism of Transformer and the GCN-GAT stacking module, it realizes the extraction of node features and the differential aggregation of information.
[0061] The anti-interference output module reconstructs the signal from the result processed by the graph convolutional neural network and outputs the anti-interference signal and noise distribution.
[0062] Example 2:
[0063] This embodiment provides a LiDAR anti-interference method based on graph convolutional neural networks, such as... Figure 1 As shown, it includes the following steps:
[0064] (1) Obtain the original one-dimensional echo signal of the lidar;
[0065] (2) The transmission characteristics of laser in a smoke environment were studied by pulsed laser polarization Monte Carlo simulation;
[0066] The Monte Carlo method was used to simulate the transmission characteristics of pulsed lasers in a smoke environment, and the influence of the composition, size distribution, bulk optical properties and geometry of smoke particles on laser transmission was analyzed.
[0067] By simulating and calculating the characteristics of echo signals under different smoke concentrations, particle sizes, and polarization states, prior knowledge is provided for subsequent signal processing.
[0068] The pulsed laser polarization Monte Carlo simulation module uses Stokes vectors and Mueller matrices to establish a polarization Monte Carlo simulation model. Through optical path analysis and correlation coefficient verification, it identifies the differences in echo characteristics under different polarization states.
[0069] (3) The echo signal is denoised using the full waveform decomposition and starfish optimization algorithm;
[0070] After acquiring the raw one-dimensional echo signal from the lidar, noise reduction is performed using a combination of full waveform decomposition and Starfish Optimization (SSA) algorithm.
[0071] Full waveform decomposition: This algorithm uses full waveform decomposition to obtain the best-fit sub-echoes of the smoke and the target, such as... Figure 2 As shown, firstly, the original echo is denoised using a variational mode decomposition algorithm to obtain denoised echo data. Then, a dual threshold criterion is constructed through curvature extremum detection and gradient field analysis to accurately identify the peak position and neighborhood inflection point distribution characteristics of the Gaussian component, thereby achieving quantitative determination of the mixed waveform components.
[0072] Starfish optimization algorithm: Adaptively optimizes VMD parameters to improve decomposition accuracy and suppress noise interference.
[0073] The starfish optimization algorithm mimics the behavior of starfish turning over rocks to find prey and utilizes group cooperation to improve foraging efficiency. Starfish are divided into explorers, developers, and collaborators based on their roles. During the iteration process, the explorer position update formula is:
[0074] ;
[0075] In the formula, , These represent the current optimal individual position and the random individual position, respectively; rand is a random number uniformly distributed in [0,1]. , To control the weighting coefficients between global and random searches.
[0076] Gaussian sub-waveform fitting: Gaussian fitting is performed on the denoised signal to extract the characteristic parameters (such as peak intensity, pulse width, polarization characteristics, etc.) of the smoke echo and the target echo.
[0077] (4) Based on the denoised echo signal, a graph structure is constructed with three polarized echoes (0°, 45°, 135°) as nodes;
[0078] Based on the denoised echo signal, a graph structure is constructed using the three polarized echoes (0°, 45°, 135°) as nodes, where:
[0079] ;
[0080] In the formula, σ() is the activation function. Represents the node features of the Lth layer. To add self-connected adjacency matrices (to avoid ignoring the characteristics of the nodes themselves); for The degree matrix, Let be the learnable weight matrix for layer (l).
[0081] Node: Each node represents an echo signal, and its feature vector includes intensity, degree of polarization, timing information, etc.
[0082] An adjacency matrix is constructed based on the temporal continuity and polarization correlation of the signal. If two nodes are temporally continuous or have similar polarization characteristics, a connection is established.
[0083] (5) Input the constructed graph structure into a graph convolutional neural network based on Transformer attention mechanism and GCN-GAT stacking, such as Figure 3 As shown.
[0084] After removing the 90° polarization state from the smoke and target polarization echo obtained by full waveform decomposition, the three polarization state waveforms under the same label are used as nodes to construct a subgraph.
[0085] First, the input signal is subjected to global feature extraction through the Transformer attention mechanism to capture long-range dependencies;
[0086] Then, the GCN-GAT stacking module is used to extract local features of the nodes in the graph, and the mutual influence between different polarization states is fused through the graph attention mechanism.
[0087] The features are mapped to the output vector through a fully connected layer, and the Softmax function is used for normalization to obtain the classification result.
[0088] (6) Reconstruct the signal from the network output and output the anti-interference signal and noise distribution.
[0089] Optional implementation methods
[0090] During the signal reconstruction stage, wavelet transform or Kalman filtering can be used to further optimize the temporal continuity. If the characteristics of the target echo and the smoke echo overlap significantly, multimodal data (such as infrared or millimeter-wave radar) can be introduced to assist in classification. The system supports online learning and adapts to the dynamic interference environment by updating the network weights in real time.
[0091] Using the above method, this invention can effectively distinguish between smoke interference and real target echoes, improving the anti-interference capability of lidar in complex environments. The deployment process at the hardware and algorithm levels is as follows: Figure 4 As shown in the figure, the experiment shows that the method can still maintain a signal-to-noise ratio improvement of more than 15dB in a strong noise environment, while the processing delay is controlled within 10ms.
[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A lidar anti-interference system based on graph convolutional neural networks, characterized in that, include: The pulsed laser polarization Monte Carlo simulation module is used to study the transmission characteristics of pulsed lasers in smoke environments, including the composition, scale distribution, volume optical properties, and geometric influence of smoke particles on laser transmission. The echo signal denoising module uses full waveform decomposition and starfish optimization algorithm to denoise the original echo signal and achieves accurate fitting of Gaussian sub-waveforms to extract the characteristic echoes of smoke echoes and targets. The graph construction module, based on the denoised echo signal, uses three polarized echoes (0°, 45°, 135°) as nodes to construct a graph structure; The graph convolutional neural network module is used to process the node graph after the graph is constructed. Based on the attention mechanism of Transformer and the GCN-GAT stacking module, it realizes the extraction of node features and the differential aggregation of information. The anti-interference output module reconstructs the signal from the result processed by the graph convolutional neural network and outputs the anti-interference signal and noise distribution.
2. The lidar anti-interference system based on graph convolutional neural networks according to claim 1, characterized in that, The pulsed laser polarization Monte Carlo simulation module uses Stokes vectors and Mueller matrices to establish a polarization Monte Carlo simulation model. Through optical path analysis and correlation coefficient verification, it identifies the differences in echo characteristics under different polarization states.
3. The lidar anti-interference system based on graph convolutional neural networks according to claim 1, characterized in that, The echo signal denoising module uses variational mode decomposition algorithm to denoise the original echo, and constructs a dual threshold criterion through curvature extremum detection and gradient field analysis to accurately identify the peak position and neighborhood inflection point distribution characteristics of Gaussian components.
4. The lidar anti-interference system based on graph convolutional neural networks according to claim 1, characterized in that, The starfish optimization algorithm in the echo signal denoising module mimics the starfish's behavior of turning over rocks to find prey. Starfish are divided into explorers, developers, and collaborators based on their roles. The explorer position update formula is: ； Where, in the formula, , These represent the current optimal individual position and the random individual position, respectively; rand is a random number uniformly distributed in [0,1]. , To control the weighting coefficients between global and random searches.
5. A lidar anti-interference system based on graph convolutional neural networks according to claim 1, characterized in that, The formula for constructing the graph structure by the graph composition module is: ； Where σ() is the activation function, Represents the node features of the Lth layer. To add self-connected adjacency matrices; for The degree matrix; Let be the learnable weight matrix for layer (l).
6. The lidar anti-interference system based on graph convolutional neural networks according to claim 1, characterized in that, The GCN-GAT stacked structure in the graph convolutional neural network module first extracts local features of nodes through a graph convolutional network, and then realizes differentiated aggregation of information between nodes through a graph attention network.
7. A method for anti-interference of lidar based on graph convolutional neural networks, characterized in that... The system according to any one of claims 1-6 includes the following steps: (1) Obtain the original one-dimensional echo signal of the lidar; (2) The transmission characteristics of laser in a smoke environment were studied by pulsed laser polarization Monte Carlo simulation; (3) The echo signal is denoised using the full waveform decomposition and starfish optimization algorithm; (4) Based on the denoised echo signal, a graph structure is constructed with three polarized echoes (0°, 45°, 135°) as nodes; (5) Input the constructed graph structure into a graph convolutional neural network based on the Transformer attention mechanism and GCN-GAT stack; (6) Reconstruct the signal from the network output and output the anti-interference signal and noise distribution.
8. The anti-interference method for lidar based on graph convolutional neural networks according to claim 7, characterized in that, The pulsed laser polarization Monte Carlo simulation includes: The photon polarization state was characterized using the Stokes parameter method; The Mueller matrix is used to describe the changes in polarized light during the scattering process; Simulation analysis identifies that the 90° polarization state has fewer object echo features in the signal due to the depolarization effect of multiple scattering, thus eliminating the 90° polarization state in subsequent processing.
9. A laser radar anti-interference method based on graph convolutional neural networks according to claim 7, characterized in that, The denoising process combining full waveform decomposition and starfish optimization algorithm includes: The original echo was denoised using the variational mode decomposition algorithm to obtain denoised echo data; A dual threshold criterion is constructed by combining curvature extremum detection and gradient field analysis to identify the peak position of Gaussian components; The variational mode decomposition parameters are adaptively optimized using the starfish optimization algorithm. Gaussian fitting is performed on the denoised signal to extract the characteristic parameters of the smoke echo and the target echo.
10. A method for anti-interference of lidar based on graph convolutional neural networks according to claim 7, characterized in that, The graph convolutional neural network processing includes: After removing the 90° polarization state from the smoke and target polarization echo obtained by full waveform decomposition, the three polarization state waveforms under the same label are used as nodes to construct a subgraph. First, the input signal is subjected to global feature extraction through the Transformer attention mechanism to capture long-range dependencies; Then, the GCN-GAT stacking module is used to extract local features of the nodes in the graph, and the mutual influence between different polarization states is fused through the graph attention mechanism. The features are mapped to the output vector through a fully connected layer, and the Softmax function is used for normalization to obtain the classification result.