Lidar point cloud data processing method based on convolutional neural network and denoising regularization method, storage medium and equipment

Abstract

This invention relates to the field of lidar point cloud data processing, specifically to a method, storage medium, and device for processing lidar point cloud data, based on convolutional neural networks and denoising regularization. The purpose of this invention is to address the problems of low denoising accuracy and low processing efficiency in existing lidar point cloud data denoising methods, leading to large errors between the denoised image and the real image. The process is as follows: obtaining preprocessed lidar point cloud data; obtaining a denoiser; constructing a regularization term; constructing an objective function; solving the objective function to obtain denoised lidar point cloud data; using the preprocessed lidar point cloud data as input and the denoised lidar point cloud data as output, performing self-supervised training based on the penalty function in the objective function to obtain a trained network; and processing the lidar point cloud data to be tested using the trained network.

Description

Technical Field

[0001] This invention relates to the field of lidar point cloud data processing, and to lidar point cloud data processing methods, storage media, and devices. Background Technology

[0002] LiDAR (Light Detection and Ranging) technology, as a core tool in modern surveying and remote sensing, constructs point cloud data composed of countless three-dimensional coordinate points by emitting laser beams and accurately measuring their reflection time differences. These datasets not only record the spatial location (X, Y, Z) of target objects in detail, but may also contain additional information such as intensity and color, providing a valuable data foundation for fields such as autonomous driving, robot navigation, terrain mapping, and urban 3D modeling.

[0003] However, lidar point cloud data is inevitably subject to various noise interferences during the acquisition process, such as equipment errors, environmental factors, and changes in the reflectivity of the target object. This noise not only reduces data accuracy but can also severely impact subsequent data processing, feature extraction, and decision analysis. Therefore, effectively removing noise from point cloud data while maintaining its fine structure and key features has become a critical problem that urgently needs to be solved in the application of lidar technology.

[0004] In recent years, the method of using denoisers as regularization terms—namely, regularization by denoising (RED)—has achieved significant results in the field of image processing. Traditional image denoising methods are mostly based on statistical models or filtering algorithms, such as mean filtering, median filtering, and Wiener filtering. These methods can remove noise to a certain extent, but they often come at the cost of blurring image details or losing edge information. The core idea of ​​the method of using denoisers as regularization terms is to embed the denoising process into the optimization framework of image processing as part of the regularization term. In optimization problems, regularization terms are usually used to constrain the solution space to prevent overfitting or the generation of unstable solutions. In the context of denoising, regularization terms are used to balance the relationship between denoising effect and image fidelity, that is, to remove noise while preserving as much of the original information and structure of the image as possible.

[0005] With the rise of deep learning technology, deep learning-based denoisers have demonstrated powerful capabilities in image processing. Deep learning models, especially convolutional neural networks (CNNs), can automatically learn feature representations of images and utilize these features for denoising. Incorporating such deep learning denoisers as part of a regularization term can further improve image denoising performance and adapt to different types of noise and image content. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of low denoising accuracy and low processing efficiency of existing LiDAR point cloud data denoising methods, which result in large errors between the denoised images and the real images. Therefore, this invention proposes a LiDAR point cloud data processing method, storage medium, and device based on convolutional neural networks and denoising regularization methods.

[0007] The specific process of the lidar point cloud data processing method based on convolutional neural networks and denoising regularization is as follows:

[0008] Step 1: Preprocess the raw lidar point cloud data to obtain preprocessed lidar point cloud data;

[0009] Step 2: Obtain the noise denoiser;

[0010] Step 3: Based on the denoiser obtained in Step 2, construct the regularization term;

[0011] Step 4: Combine the data fidelity term and the regularization term constructed in Step 3 to construct the objective function;

[0012] Step 5: Solve the objective function constructed in Step 4 using an optimization algorithm to obtain the denoised lidar point cloud data;

[0013] Step 6: Use the preprocessed LiDAR point cloud data from Step 1 as the input to the deep learning network, and the denoised LiDAR point cloud data as the output of the deep learning network. Perform self-supervised training based on the penalty function in the objective function constructed in Step 4 until convergence, and obtain the trained deep learning network.

[0014] Step 7: Input the point cloud data of the LiDAR to be tested into the trained deep learning network, and the trained deep learning network outputs the denoised point cloud data of the LiDAR to be tested.

[0015] A computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods.

[0016] A lidar point cloud data processing device based on convolutional neural networks and denoising regularization methods is disclosed. The device includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods. The beneficial effects of this invention are:

[0017] This invention proposes a LiDAR point cloud data processing method based on convolutional neural networks and denoising regularization methods. While denoising, it more completely preserves the original features and structural information of the point cloud data, improves data processing efficiency, denoising accuracy, and data quality.

[0018] Compared with the prior art, the present invention has obvious beneficial effects. Through the above technical solution, the method provided by the present invention can achieve considerable technical progress and practicality, and has broad industrial application value.

[0019] This invention utilizes a denoising regularization method to constrain the complexity of the model by introducing a regularization term, thereby achieving effective noise suppression. Compared to traditional denoising methods, this method can better preserve the original features and structural information of point cloud data while removing noise.

[0020] LiDAR point cloud data is typically characterized by its large scale and high dimensionality, making traditional methods inefficient in processing it. This invention utilizes convolutional neural networks, whose unique convolution and pooling operations effectively extract key features from point clouds, thereby significantly improving the speed and efficiency of data processing.

[0021] This invention sets the denoiser as a regularization term, paired with a convolutional neural network denoiser. Unlike existing methods that use only prior knowledge as a penalty function for regularization, this invention uses denoising to constrain the solution space of the final data. This is similar to how regularization uses prior knowledge to constrain the solution space. Therefore, the denoiser is embedded into the optimization framework, and the regularization term is defined using a denoising algorithm. This achieves regularization while simultaneously denoising, resulting in higher-quality data, enhanced model generalization ability, improved denoising accuracy, and reduced error between the denoised image and the real image. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0023] Specific Implementation Method 1: The specific process of the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods in this implementation method is as follows:

[0024] This invention introduces the concept of denoising regularization into the processing of lidar point cloud data. The core of this method lies in treating the denoiser as a regularization term in an optimization problem, which, together with the data fidelity term (i.e., the difference between the original point cloud data and the observed point cloud data), constructs a comprehensive objective function.

[0025] The advantages of using a denoiser as a regularization term in processing LiDAR point cloud data lie in its flexibility and scalability. On one hand, the denoiser can be specifically designed or selected based on the characteristics of the point cloud data and the type of noise, such as denoisers based on statistical models, filtering algorithms, or deep learning. These denoisers can effectively suppress different types of noise, improving the denoising effect. On the other hand, the introduction of a regularization term provides additional constraints to the denoising process, helping to find the optimal balance between noise removal and data structure preservation. Furthermore, incorporating a CNN denoiser as part of the regularization term can further improve the denoising effect and adapt to different types of LiDAR point cloud data. Finally, a LiDAR point cloud data processing method and system based on convolutional neural networks and denoising regularization methods are proposed.

[0026] Step 1: Preprocess the raw lidar point cloud data to obtain preprocessed lidar point cloud data;

[0027] Step 2: Obtain the noise denoiser;

[0028] Based on the characteristics and noise type of the lidar point cloud data obtained in step 1, design or select a suitable denoiser.

[0029] Step 3: Based on the denoiser obtained in Step 2, construct the regularization term;

[0030] Incorporating the denoiser's output or some metric based on its performance as part of the regularization term can be achieved by defining a penalty function that is related to the denoiser's output.

[0031] Step 4: Combine the data fidelity term (i.e., the difference between the original point cloud data and the observed point cloud data) and the regularization term constructed in Step 3 to construct the objective function;

[0032] Step 5: Solve the objective function constructed in Step 4 using an optimization algorithm to obtain the denoised lidar point cloud data;

[0033] The estimated values ​​of the point cloud data are updated iteratively, and a denoiser is used to process them in each iteration;

[0034] Step 6: Use the preprocessed LiDAR point cloud data from Step 1 as the input to the deep learning network, and the denoised LiDAR point cloud data as the output of the deep learning network. Perform self-supervised training based on the penalty function in the objective function constructed in Step 4 until convergence, and obtain the trained deep learning network.

[0035] Step 7: Input the point cloud data of the LiDAR to be tested into the trained deep learning network, and the trained deep learning network outputs the denoised point cloud data of the LiDAR to be tested.

[0036] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step 1, the original lidar point cloud data is preprocessed to obtain preprocessed lidar point cloud data; the specific process is as follows:

[0037] If the point cloud data is too dense, methods such as voxel grid downsampling can be used to reduce the number of points and improve the efficiency of subsequent processing. The point cloud space is divided into multiple voxel grids. Within each grid, the centroid of all points or a random point is taken as a representative point.

[0038]

[0039] Among them, Voxel k It is the number of points in the k-th voxel grid;

[0040] Assume the original LiDAR point cloud dataset is P = {p1, p2, p3, ..., p...} N};

[0041] Where p i It is a lidar point cloud in three-dimensional space, i = 1, 2, ..., N;

[0042] For each lidar point cloud p i Calculate p i Compared with all other point clouds p in the original LiDAR point cloud dataset j distance d ij , and find p i The average distance of the k nearest neighbors

[0043] Set a threshold T;

[0044] If point cloud p i The average distance of the k nearest neighbors If the value is greater than the threshold T, then the point cloud p is considered to be... i These are outliers or anomalies; the point cloud p... i Remove from the original LiDAR point cloud dataset;

[0045] The condition 1≤k≤200;

[0046] Represented as

[0047]

[0048]

[0049] Among them, NN k (i) represents the point cloud p i The set of k nearest neighbors, where Remove means remove;

[0050] For efficiency reasons, typically only the distances to a fixed number (e.g., k) of the nearest neighbors of each point are calculated, and decisions are made based on these distances.

[0051] The other steps and parameters are the same as in Specific Implementation Method 1.

[0052] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step 2, the noise denoiser is obtained;

[0053] Based on the characteristics and noise type of the lidar point cloud data obtained in step 1, design or select a suitable denoiser.

[0054] The specific process is as follows:

[0055] Noise is unwanted interference or distortion caused by various factors during image acquisition, transmission, storage, or processing. Common types of noise include Poisson noise, Gaussian noise, and salt-and-pepper noise.

[0056] The goal of image denoising is to recover a clean image x from a noisy observed image y. Its model is: y = x + n, where n is typically additive white Gaussian noise or Poisson noise. This model shows that image denoising is an inverse problem.

[0057] Design or select a noise denoiser:

[0058] The denoiser can be a statistical model-based denoiser, a deep learning model, or a denoising algorithm.

[0059] The denoiser based on the statistical model is a bilateral filter or a Gaussian filter, etc. These filters can smooth noise while preserving edge information as much as possible.

[0060] The deep learning model is PointNet or PointConv, etc. PointNet, PointConv and other convolutional neural networks are designed specifically for point cloud data. They can automatically learn the feature representation of point cloud data and use it for noise reduction.

[0061] The denoising algorithm is the Non-Local Means algorithm, which uses non-local similarity in point clouds to remove noise.

[0062] Other steps and parameters are the same as in specific implementation method one or two.

[0063] Specific Implementation Method Four: This implementation method differs from one of the specific implementation methods one to three in that: in step 3, a regularization term is constructed based on the denoiser obtained in step 2;

[0064] Incorporating the denoiser's output or some metric based on its performance as part of the regularization term can be achieved by defining a penalty function that is related to the denoiser's output.

[0065] The regularization expression is:

[0066]

[0067] Where R(y) is the regularization term; y is the noise denoiser; y is the preprocessed lidar point cloud data; the superscript T indicates transpose.

[0068] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0069] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: in step 4, the objective function is constructed based on the data fidelity term (i.e., the difference between the original point cloud data and the observed point cloud data) and the regularization term constructed in step 3; the specific process is as follows:

[0070] Represented as:

[0071] The objective function is expressed as:

[0072]

[0073] Where x represents the target image (unknown) and y represents the preprocessed lidar point cloud data;

[0074] l(y,x) represents the data fidelity term, which guarantees that y≈x;

[0075] λ represents the degree of penalty for R(y). The weights between the data fidelity term and the regularization term are adjusted according to actual needs to balance the relationship between the two.

[0076] λR(y) represents the penalty function.

[0077] The other steps and parameters are the same as those in specific implementation methods one through four.

[0078] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that: in step 5, an optimization algorithm is used to solve the objective function constructed in step 4 to obtain the denoised lidar point cloud data;

[0079] The estimated values ​​of the point cloud data are updated iteratively, and a denoiser is used to process them in each iteration;

[0080] The specific process is as follows:

[0081] Step 51: Construct the Lagrange augmented function:

[0082]

[0083] in, f(x) represents an intermediate variable. The expression is defined as follows: p(y|x) represents the prior probability (unknown quantity) of finding x given y;

[0084] g(x) represents an intermediate variable, and p(x) represents the probability of x.

[0085] f(x) + λg(v) is the original optimization objective;

[0086] v represents the target image;

[0087] Under the condition that x = v, (xv) is an equality constraint, and u is a Lagrange multiplier associated with the constraint (xv) = 0;

[0088] ρ is the penalty coefficient;

[0089] Step 52: Set the maximum number of iterations K;

[0090] Step 53, iteration number k = 1;

[0091] Step 54: To solve the minimum problem, the Lagrange augmented function is transformed into solving the following subproblems:

[0092]

[0093] u (k+1) =u (k) +(x (k+1) -v (k+1) )

[0094] ρ k+1 =γ k ρ k

[0095] Where, x (k+1) Let x represent the target image in the (k+1)th iteration. (k) This represents the target image in the k-th iteration;

[0096] ρ k ρ represents the penalty coefficient for the k-th iteration. k+1 This represents the penalty coefficient for the (k+1)th iteration;

[0097] v (k) Let v represent the target image in the k-th iteration. (k+1) This represents the target image in the (k+1)th iteration;

[0098] u (k)Let u be the Lagrange multiplier associated with the constraint (xv) = 0 in the k-th iteration. (k+1) Let represent the Lagrange multiplier associated with the constraint (xv) = 0 in the (k+1)th iteration;

[0099] This represents the denoiser for the k-th iteration;

[0100] γ k Represents ρ k Adjustment parameters;

[0101] It is a denoising algorithm (or "denoiser" for short); It is a parameter that controls the strength of the noise denoiser; ρ is improved to ρ k+1 =γ k ρ k γ k ≥1 is the adjustment parameter for ρ.

[0102] Step 55: Let the iteration number k = k + 1, and repeat step 54 until the maximum iteration number K is reached to obtain x. K That is, the target image x in the objective function (satisfying...) The corresponding x).

[0103] Derivation of the principle:

[0104] The Plug-and-Play ADMM algorithm performs well in terms of convergence and final data recovery. Traditional ADMM algorithms:

[0105] The data recovery problem can be viewed as a maximum a posteriori probability estimation problem.

[0106]

[0107] Where x is the target image and y is the image output by the sensor. It is an estimate of x given y.

[0108] Transform the unconstrained problem into a constrained problem:

[0109]

[0110] Obey x = v

[0111] Consider the Lagrange augmentation:

[0112]

[0113] Where f(x)+λg(v) is the original optimization objective, under the condition that x=v, (xv) is the equality constraint, and u is the Lagrange multiplier associated with the constraint (xv)=0.

[0114] The regularization term constructed in step 3 Combined,

[0115]

[0116] To solve the minimum value problem, we can transform it into solving the following subproblems:

[0117]

[0118] u (k+1) =u (k) +(x (k+1) -v (k+1) )

[0119] Plug-and-PlayADMM algorithm:

[0120]

[0121] in, It is an existing denoising algorithm.

[0122] Improved Plug-and-PlayADMM algorithm:

[0123]

[0124] u (k+1) =u (k) +(x (k+1) -v (k+1) )

[0125] ρ k+1 =γ k ρ k

[0126] in, It is a denoising algorithm (or simply "denoiser"). It is a parameter that controls the strength of the noise denoiser; ρ is improved to ρ k+1 =γ k ρ k γ k ≥1 is the adjustment parameter for ρ.

[0127] The other steps and parameters are the same as those in specific implementation methods one through five.

[0128] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One to Six in that: in step 6, the preprocessed lidar point cloud data from step 1 is used as the input of the deep learning network, and the denoised lidar point cloud data is used as the output of the deep learning network. Self-supervised training is performed based on the penalty function in the objective function constructed in step 4 until convergence, and a trained deep learning network is obtained.

[0129] The specific process is as follows:

[0130] The preprocessed LiDAR point cloud data from step 1 is used as the input to the deep learning network, and the target image x obtained in step 5 is used as the output of the deep learning network. Self-supervised training is performed based on the penalty function in the objective function constructed in step 4 until convergence, thus obtaining the trained deep learning network.

[0131] The other steps and parameters are the same as those in specific implementation methods one through six.

[0132] Specific Implementation Method Eight: This implementation method differs from one of the specific implementation methods one to seven in that the deep learning network is the PointNet deep learning network.

[0133] PointNet was the first deep learning network to directly process point cloud data. It treats point cloud data as an unordered set of points and uses symmetric functions (such as max pooling) to ensure the network's permutation invariance to the point cloud data. The core idea of ​​PointNet is to use a multilayer perceptron (MLP) to extract features from each point, then use a max pooling layer to aggregate the features of the points into global features, and finally use these features for tasks such as classification or regression.

[0134] Direct point cloud processing: PointNet can directly process raw point cloud data without converting it to other formats (such as voxel meshes or images).

[0135] Permutation invariance: Through symmetric functions such as max pooling layers, PointNet is invariant to the order of points in the point cloud.

[0136] PointConv is a deep convolutional network for point cloud data. It implements convolution operations on point cloud data by treating the convolution kernel as a non-linear function of the local coordinates of 3D points, consisting of weights and a density function. PointConv can handle non-uniformly sampled point cloud data and learns effective convolution kernels for feature extraction.

[0137] Density-reweighted convolution: PointConv compensates for the effects of non-uniform sampling through a density reweighting mechanism, enabling the network to process point cloud data more accurately.

[0138] Translation invariance and permutation invariance: PointConv can maintain invariance to the translation and permutation of points in the point cloud, thereby improving the robustness of the network.

[0139] For the neural network used to process point cloud data, a self-supervised training method is adopted. The network parameters are fine-tuned using training samples to obtain the final point cloud data denoising system for point cloud data processing tasks.

[0140] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0141] Specific Implementation Method Nine: This implementation method is a computer storage medium that stores at least one instruction. The at least one instruction is loaded and executed by a processor to implement the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods as described in any one of 1 to 8.

[0142] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions.

[0143] Specific Implementation Method 10: This implementation method is a lidar point cloud data processing device based on convolutional neural networks and denoising regularization methods. The device includes a processor and a memory. The memory stores at least one instruction. The at least one instruction is loaded and executed by the processor to implement the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods as described in any one of claims 1 to 8.

[0144] The device includes a processor and a memory. It should be understood that the device includes any device including a processor and a memory as described in this invention. The device may also include other units or modules that perform display, interaction, processing, control and other functions via signals or instructions.

[0145] The memory stores at least one instruction, which is loaded and executed by the processor to implement the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods.

[0146] Those skilled in the art will understand that at least one stored instruction constitutes a computer program product corresponding to a method or system. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0147] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0148] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0149] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0150] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0151] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0152] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for processing lidar point cloud data based on convolutional neural networks and denoising regularization methods, characterized in that: The specific process of the method is as follows: Step 1: Preprocess the raw lidar point cloud data to obtain preprocessed lidar point cloud data; Step 2: Obtain the noise denoiser; Step 3: Based on the denoiser obtained in Step 2, construct the regularization term; Step 4: Combine the data fidelity term and the regularization term constructed in Step 3 to construct the objective function; Step 5: Solve the objective function constructed in Step 4 using an optimization algorithm to obtain the denoised lidar point cloud data; Step 6: Use the preprocessed LiDAR point cloud data from Step 1 as the input to the deep learning network, and the denoised LiDAR point cloud data as the output of the deep learning network. Perform self-supervised training based on the penalty function in the objective function constructed in Step 4 until convergence, and obtain the trained deep learning network. Step 7: Input the point cloud data of the LiDAR to be tested into the trained deep learning network, and the trained deep learning network outputs the denoised point cloud data of the LiDAR to be tested.

2. The lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods according to claim 1, characterized in that: In step 1, the original lidar point cloud data is preprocessed to obtain preprocessed lidar point cloud data; the specific process is as follows: Assume the original LiDAR point cloud dataset is P = {p1, p2, p3, ..., p...} N }; Where p i It is a lidar point cloud in three-dimensional space, i = 1, 2, ..., N; For each lidar point cloud p i Calculate p i Compared with all other point clouds p in the original LiDAR point cloud dataset j distance d ij And find p i The average distance of the k nearest neighbors i≠j; Set a threshold T; If point cloud p i The average distance of the k nearest neighbors If the value is greater than the threshold T, then the point cloud p is considered to be... i These are outliers or anomalies; the point cloud p... i Removed from the original LiDAR point cloud dataset.

3. The method of claim 2, wherein: In step 2, the noise denoiser is obtained; The specific process is as follows: The denoiser can be a statistical model-based denoiser, a deep learning model, or a denoising algorithm. The denoiser based on the statistical model is either a bilateral filter or a Gaussian filter. The deep learning model is either PointNet or PointConv; The denoising algorithm is a nonlocal mean algorithm.

4. The lidar point cloud data processing method based on convolutional neural networks and denoising regularization method according to claim 3, characterized in that: In step 3, a regularization term is constructed based on the denoiser obtained in step 2; The regularization expression is: Where R(y) is the regularization term; y is the noise denoiser; y is the preprocessed lidar point cloud data; the superscript T indicates transpose.

5. The lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods according to claim 4, characterized in that: In step 4, the objective function is constructed based on the data fidelity term and the regularization term built in step 3; the specific process is as follows: Represented as: The objective function is expressed as: Where x represents the target image and y represents the preprocessed lidar point cloud data; l(y,x) represents the data fidelity term, which guarantees that y≈x; λ represents the degree of penalty imposed on R(y); λR(y) represents the penalty function.

6. The method of claim 5, wherein: In step 5, an optimization algorithm is used to solve the objective function constructed in step 4 to obtain the denoised lidar point cloud data. The specific process is as follows: Step 51: Construct the Lagrange augmented function: in, f(x) represents an intermediate variable. The expression is defined as follows: p(y|x) represents the prior probability of finding x given y. g(x) represents an intermediate variable, and p(x) represents the probability of x. v represents the target image; Under the condition that x = v, (xv) is an equality constraint, and u is a Lagrange multiplier associated with the constraint (xv) = 0; ρ is the penalty coefficient; Step 52: Set the maximum number of iterations K; Step 53, iteration number k = 1; Step 54: Transform the Lagrange augmented function into solving the following subproblems: u (k+1) = u (k) + (x (k+1) - v (k+1) ) p k+1 = γ k p k Where, x (k+1) Let x represent the target image in the (k+1)th iteration. (k) This represents the target image in the k-th iteration; ρ k ρ represents the penalty coefficient for the k-th iteration. k+1 This represents the penalty coefficient for the (k+1)th iteration; v (k) target image at the kth iteration, v (k+1) target image at the k+1th iteration; u (k) Let u be the Lagrange multiplier associated with the constraint (xv) = 0 in the k-th iteration. (k+1) Let represent the Lagrange multiplier associated with the constraint (xv) = 0 in the (k+1)th iteration; denoiser representing the kth iteration; gamma k represents a tuning parameter; and k represents a tuning parameter; and Step 55: Let the iteration number k = k + 1, and repeat step 54 until the maximum iteration number K is reached to obtain x. K That is, the target image x in the objective function.

7. The method of claim 6, wherein the method is characterized by: In step 6, the preprocessed lidar point cloud data from step 1 is used as the input to the deep learning network, and the denoised lidar point cloud data is used as the output of the deep learning network. Self-supervised training is performed based on the penalty function in the objective function constructed in step 4 until convergence, and a trained deep learning network is obtained. The specific process is as follows: The preprocessed LiDAR point cloud data from step 1 is used as the input to the deep learning network, and the target image x obtained in step 5 is used as the output of the deep learning network. Self-supervised training is performed based on the penalty function in the objective function constructed in step 4 until convergence, thus obtaining the trained deep learning network.

8. The lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods according to claim 7, characterized in that: The deep learning network mentioned is the PointNet deep learning network.

9. A computer storage medium, characterized in that: The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the lidar point cloud data processing method based on convolutional neural networks and denoising regularization methods as described in any one of claims 1 to 8.

10. A lidar point cloud data processing device based on convolutional neural networks and denoising regularization methods, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the lidar point cloud data processing method based on a convolutional neural network and a denoising regularization method as described in any one of claims 1 to 8.

Images