Point cloud denoising method and device, medium and product

Abstract

The embodiment of the invention provides a point cloud denoising method and device, a medium and a product, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring target point cloud data subjected to abnormal point elimination; according to the target point cloud data, constructing a density-based hierarchical clustering algorithm and a de-noising network of an operation selection strategy, and generating a de-noising model based on the de-noising network; and inputting the to-be-denoised point cloud data into the denoising model, and outputting the denoised point cloud data. According to the scheme, isolated noise possibly misleading path decision is filtered in advance, and interference is cleared for follow-up path selection; a denoising network based on a density hierarchical clustering algorithm and an operation selection strategy is constructed, a point cloud structure is accurately divided, representative elite points are screened in combination with the operation selection strategy, under extreme conditions, the elite points are preferentially used as core extension paths, noise point dominant decision making is avoided, and the accuracy of the system is improved. The problem that in the prior art, a single path is poor in adaptability in an extreme scene is solved, and the denoising accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a point cloud denoising method and device, medium and product. BACKGROUND

[0002] In point cloud data, noise points refer to those points that are incorrect or inaccurate due to various reasons (such as sensor errors, environmental interference, etc.). These noise points generally affect the subsequent use and analysis of data, and removing data noise and improving data accuracy can improve the quality of corona data and subsequent processing effect when processing point cloud data. However, in the process of optimizing and constructing the denoising network by the traditional density clustering algorithm, there is a path decision selection optimization problem in the extreme case, which leads to unsatisfactory denoising effect. SUMMARY

[0003] At least one embodiment of the present application provides a point cloud denoising method, device, medium and product, which is used to solve the problem of low denoising accuracy caused by the path decision selection optimization in the extreme case in the prior art.

[0004] In order to solve the above technical problems, the present application is implemented as follows:

[0005] In a first aspect, the embodiments of the present application provide a point cloud denoising method, comprising:

[0006] Obtaining target point cloud data for abnormal point elimination;

[0007] According to the target point cloud data, a denoising network based on a density-based hierarchical clustering algorithm and a management selection strategy is constructed, and a denoising model is generated based on the denoising network;

[0008] Inputting the point cloud data to be denoised into the denoising model, and outputting the denoised point cloud data.

[0009] Optionally, obtaining the target point cloud data for abnormal point elimination comprises:

[0010] Collecting original point cloud data;

[0011] Adding Gaussian noise and / or random noise to the original point cloud data to generate noise point cloud data; wherein the Gaussian noise is generated by moving the points in the original point cloud according to the Gaussian distribution rule, without adding new data points; the random noise is generated by generating random numbers uniformly distributed in a preset range and superimposing them on the coordinates of the original point cloud, so that random disturbance is generated around each point;

[0012] Based on the local outlier factor algorithm, the noise point cloud data is subjected to abnormal point elimination to obtain the target point cloud data for abnormal point elimination.

[0013] Optionally, based on the local outlier factor algorithm, the noise point cloud data is subjected to outlier elimination to obtain target point cloud data subjected to outlier elimination, comprising:

[0014] Mahalanobis distance between any two data points in the noise point cloud data is calculated; the Mahalanobis distance is calculated through a difference vector of the two data points, a transpose of the difference vector and an inverse matrix of a multi-dimensional random variable covariance matrix of the noise point cloud data;

[0015] A k-distance neighborhood of each data point is determined, and a k-reachable distance of each point in the k-distance neighborhood is solved based on the Mahalanobis distance; the k-reachable distance is a maximum value of the k-distance of a first target point and the Mahalanobis distance between the first target point and a second target point;

[0016] A local reachable density of each data point is calculated according to the k-reachable distance; the local reachable density is a reciprocal of an average value of the k-reachable distances of all points in the k-distance neighborhood of the data point;

[0017] A local outlier factor of each data point is calculated based on the local reachable density;

[0018] n data points with the largest local outlier factors are selected as outlier points to form an outlier point set, and the outlier point set is removed from the noise point cloud data to obtain target point cloud data subjected to outlier elimination.

[0019] Optionally, based on the target point cloud data, a density-based hierarchical clustering algorithm and a denoising network based on a management selection strategy are constructed, and a denoising model is generated based on the denoising network, comprising:

[0020] A density-based hierarchical clustering algorithm is used to cluster the target point cloud data to obtain a clustering result;

[0021] An elite strategy based on a management selection strategy is introduced, and an optimal individual is selected as an elite point from a clustering cluster corresponding to the clustering result to obtain elite point information; the elite point is a center or a most representative data point of each clustering cluster;

[0022] Based on the clustering result and the elite point information, a denoising network is constructed;

[0023] The target point cloud data is divided into a training set and a validation set, the training set is input into the denoising network for training, and the network parameters are adjusted through the validation set to generate a denoising model.

[0024] Optionally, a density-based hierarchical clustering algorithm is used to cluster the target point cloud data to obtain a clustering result, comprising:

[0025] Calculate the mutual reachable distance of each point in the target point cloud data; the mutual reachable distance is the maximum of the core distance of a first position point, the core distance of a second position point of another reference, and the distance between the first position point and the second position point; the core distance is the distance between the first position point and the kth nearest sample point;

[0026] Based on the mutual reachable distance, set the minimum cluster sample point number, and generate a minimum tree with the mutual reachable distance as the weight;

[0027] Sort the weights of the minimum tree in ascending order, and use the density-based hierarchical clustering algorithm to preferentially merge the cluster with the smallest weight, and iteratively perform the merging operation until all clusters are merged into the state with the highest weight, to construct a hierarchical structure of the cluster;

[0028] Set a minimum cluster threshold, traverse the hierarchical structure, and judge at each split; if the number of sample points contained in the new cluster after splitting is less than the minimum cluster threshold, mark the new cluster as an outlier;

[0029] Based on the reciprocal of the mutual reachable distance, calculate the stability of each cluster, and extract the cluster and the corresponding core point whose stability meets the preset condition as the clustering result.

[0030] Optionally, based on the clustering result and the elite point information, a denoising network is constructed, including:

[0031] For the local features of each cluster in the clustering result, the position of the elite point information is combined to configure a corresponding denoising path subnetwork;

[0032] A denoising block composed of a shared subnetwork and the denoising path subnetwork is constructed; the shared subnetwork is used to extract common features of each cluster; the denoising path subnetwork is used to perform differential denoising operations according to the characteristics of different clusters;

[0033] Based on the elite point information, the network feature weight of the denoising network to be constructed is optimized;

[0034] A plurality of denoising blocks are connected in turn according to the clustering hierarchical structure to form a complete denoising network; wherein the output feature of a previous denoising block is used as the input feature of a subsequent denoising block, a displacement vector is obtained through a max-pooling layer and a fully connected layer, and the point cloud coordinates are corrected and output.

[0035] Optionally, the training set of the target point cloud data is input into the denoising network for training, the network parameters are adjusted through the validation set of the target point cloud data, and a denoising model is generated, including:

[0036] The target point cloud data divided training set is input into the denoising network, the denoising network is used to extract and process the noisy point cloud features, and a preliminary denoising result is output.

[0037] During the training process, the target point cloud data divided verification set is input into the network at regular intervals, the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) are calculated, and the evaluation result of the verification set is determined.

[0038] According to the evaluation result of the verification set, the network hyperparameters are adjusted until the denoising effect of the model on the verification set reaches a preset standard, and a denoising model is generated; the network hyperparameters at least include a learning rate, a training batch size, a denoising block number, and a network layer number.

[0039] In a second aspect, an embodiment of the present application provides a point cloud denoising device, comprising:

[0040] An acquisition module is configured to acquire target point cloud data for abnormal point elimination.

[0041] A first processing module is configured to construct a denoising network based on a density-based hierarchical clustering algorithm and a management selection strategy according to the target point cloud data, and generate a denoising model based on the denoising network.

[0042] A second processing module is configured to input denoising point cloud data into the denoising model, and output denoised point cloud data.

[0043] In a third aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the method according to any one of the first aspect are implemented.

[0044] In a fourth aspect, an embodiment of the present application provides a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, the steps of the method according to any one of the first aspect are implemented.

[0045] Compared with the prior art, the point cloud denoising method, device, medium and product provided by the embodiment of the present application can obtain target point cloud data for abnormal point elimination, filter out isolated noise that may mislead path decision in advance, and clear the interference for subsequent path selection; construct a denoising network based on a density-based hierarchical clustering algorithm and a management selection strategy according to the target point cloud data, accurately divide the point cloud structure, select representative elite points in combination with the management selection strategy, and preferentially expand the path with the elite points as the core in extreme cases, avoid the noise points from dominating the decision, and generate a denoising model based on the denoising network; input denoising point cloud data into the denoising model, and output denoised point cloud data, thereby solving the poor adaptability of a single path in an extreme scenario in the prior art, and improving the denoising accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0046] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The detailed description is made with reference to the accompanying drawings.

[0047] Figure 1 A flowchart of a point cloud denoising method provided by an embodiment of the present application;

[0048] Figure 2 An outlier diagram provided by an embodiment of the present application;

[0049] Figure 3 A model training flowchart provided by an embodiment of the present application;

[0050] Figure 4 A flowchart of a point cloud denoising method provided by an embodiment of the present application;

[0051] Figure 5 A structure diagram of a point cloud denoising device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0052] The terms "first", "second", and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second" are usually a class, not limited to the number of objects, for example, the first object can be one or more. In addition, "or" in the present application means at least one of the connected objects. For example, "A or B" covers three scenarios, namely, scenario one: including A and not including B; scenario two: including B and not including A; scenario three: including A and B. The character " / " generally indicates that the objects before and after are in an "or" relationship.

[0053] The term "indicate" in the present application can be a direct indication (or explicit indication) or an indirect indication (or implicit indication). Among them, the direct indication can be understood as that the sender explicitly informs the receiver of specific information, operations to be performed or requested results, etc. in the sent indication; the indirect indication can be understood as that the receiver determines the corresponding information according to the indication sent by the sender, or judges and determines the operation to be performed or the requested result according to the judgment result.

[0054] In order for those skilled in the art to better understand the embodiments of the present application, first make the following description:

[0055] Three-dimensional point cloud data is a set of points defined in three-dimensional space, each point usually contains its coordinates in space (x, y, z) and possibly additional attributes such as color, intensity, etc. These data are widely used in computer vision, virtual reality, geographic information systems (GIS) and three-dimensional modeling, especially in the industrial field, widely used in industrial robots, industrial measurement, etc. At the same time, due to the harsh acquisition environment and insufficient hardware conditions, occlusion between objects, mobile data and other factors, the current acquired point cloud data often has quality damage such as incomplete structure and noise interference, and the above application scenarios often require high-quality point cloud for processing and analysis. Therefore, the denoising technology algorithm for point cloud data is crucial in the field of visual three-dimensional reconstruction.

[0056] The general point cloud denoising method can be roughly divided into two categories: 1) ordered point cloud denoising: point cloud data is organized in space with an ordered structure, that is, ordered point cloud. The commonly used processing methods include: filtering method, global energy method, curve checking method, etc. Gaussian filtering is suitable for normally distributed data, calculating the average distance of each point to its nearest k points, then the distance of all points in the point cloud forms a Gaussian distribution, given the mean and variance, the points outside the specified range can be proposed; Radius filtering draws a circle with a certain point as the center, calculates the number of points falling in the circle, when the number is greater than a given value, then the point is retained, the number is less than the given value, then the point is removed, the remaining points are the densest points. However, in practical applications, most point cloud models are unordered and have no specific rules and distribution.

[0057] In the denoising algorithm for scattered point cloud, there are denoising algorithms based on signal processing, denoising algorithms based on moving least squares, denoising algorithms based on statistics, data-driven denoising algorithms, etc. The point cloud denoising method based on statistical learning is also constantly progressing, for example, the denoising algorithm networks such as PointCleanNet, PointFilter and RePCD-Net, PathNet are proposed, which have good effect in the field of point cloud data noise removal.

[0058] PointCleanNet, PointFilter and RePCD-Net use the same single network to process all noisy point clouds. In complex and noisy point clouds, considering the potential existence of complex noise patterns and various structural elements, it may lead to noise residues in the denoised point cloud in the heavy noise data set, and cause geometric over-smoothing when applied to low-noise point clouds.

[0059] The PathNet denoising network can select a suitable network path for each noise point according to a new PCD paradigm, avoid using a network with a single denoising path to denoise the entire point cloud, and also consider problems such as complex noise patterns. However, PathNet is designed as a structure containing parallel paths in the denoising module, and uses a routing network based on Markov decision process to select paths. The adaptation effect is poor under unknown or significantly changed noise levels, and errors caused by the influence of point cloud density and local features are not considered.

[0060] As described in the background, there is a path decision optimization problem in the prior art point cloud denoising process under extreme conditions, which seriously affects the accuracy of denoising. To solve the above problems, the embodiments of the present application provide a point cloud denoising method, device, medium and product, which can reduce or avoid the occurrence of the above conditions and improve the accuracy of denoising.

[0061] The embodiments of the present application provide a point cloud denoising method, device, medium and product. The method and device are based on the same application concept. Since the principles of the method and device for solving problems are similar, the implementation of the device and the method can be mutually referred to, and the repeated parts will not be described again.

[0062] Please refer to Figure 1 The point cloud denoising method provided by the embodiments of the present application comprises the following steps.

[0063] Step 11, obtaining target point cloud data for outlier removal;

[0064] Step 12, constructing a denoising network based on a density-based hierarchical clustering algorithm and a management selection strategy according to the target point cloud data, and generating a denoising model based on the denoising network;

[0065] Step 13, inputting the point cloud data to be denoised into the denoising model, and outputting the denoised point cloud data.

[0066] In the embodiments of the present application, the step 11 removes extreme abnormal points in the point cloud data through preprocessing, provides high-quality input for the training and inference of the subsequent denoising network, and reduces the interference of noise on the model. The step 12 is used for constructing a denoising network and generating a denoising model. The step 12 constructs a denoising network based on a hierarchical density-based spatial clustering of applications with noise (HDBSCAN) and a non-dominated sorting genetic algorithm (NSGA) according to the target point cloud data. Through the cooperation of density clustering and multi-objective optimization, the problem of insufficient path decision optimization in extreme scenarios in the traditional method is solved, the denoising network suitable for complex point cloud structure is constructed, and the denoising model is generated based on the denoising network.

[0067] Specifically, the target point cloud data is subjected to hierarchical density clustering, the mutual reachable distance between points is calculated to identify high-density core clusters (corresponding to the main structure of the object) and low-density noise regions; the hierarchical structure of the clusters is constructed through hierarchical clustering, and stable clustering clusters (excluding small-scale noise clusters) are compressed and selected to accurately divide the local features of the point cloud, such as sharp edges and smooth surfaces. The NSGA elite strategy is introduced to select elite points from the clustering results. The elite points can be represented as the center or the most representative point of each cluster. In the subsequent network construction, the clustering is preferentially expanded based on the elite points to avoid the interference of noise points on the path decision, especially in extreme scenarios (such as sudden density changes and complex structure regions) to anchor effective features.

[0068] Further, the network is input with a training set of target point cloud data, the model performance is monitored through a validation set, the learning rate, network layer number and other hyperparameters are adjusted, and finally a denoising model with strong generalization ability is generated.

[0069] The step 13 is the application stage of the model, and the trained denoising model is used to realize efficient denoising. Specifically, the point cloud data to be processed containing noise is input into the denoising model, and the model performs adaptive denoising operations on point cloud regions with different local features through the pre-trained dynamic path strategy, such as high-complexity processing for preserving sharp features and efficient simplification processing for smooth regions. The output is the denoised point cloud data, in which the noise is effectively filtered out, while the structural features of the original point cloud, such as object edges and detail textures, are preserved, solving the problem of low denoising accuracy and feature loss in extreme scenarios in the traditional method.

[0070] The present application forms a complete process of preprocessing purification-dynamic network modeling-precise denoising through the cooperation of steps 11 to 13, realizes efficient denoising of complex point cloud data, especially in extreme scenarios, through the deep integration of density clustering and multi-objective optimization.

[0071] Optionally, the step 11 above comprises:

[0072] collecting original point cloud data;

[0073] adding Gaussian noise and / or random noise to the original point cloud data to generate noise point cloud data; wherein the Gaussian noise is generated by moving the points in the original point cloud according to the Gaussian distribution rule to generate no new data points; and the random noise is generated by generating random numbers in a preset range and superimposing them on the coordinates of the original point cloud to generate random disturbances around each point.

[0074] performing outlier rejection on the noise point cloud data based on a local outlier factor algorithm to obtain target point cloud data after outlier rejection.

[0075] In the embodiments of the present application, the purpose of collecting original point cloud data is to obtain basic point cloud data in a real scene as reference materials for subsequent processing. Three-dimensional point clouds of the surrounding environment, such as the surface of industrial parts, outdoor terrain, building outlines, etc., can be collected by visual sensors such as laser radars and depth cameras, containing key information such as the geometric structure and surface details of the object. The original point cloud data is the basis for noise simulation and denoising training, and its quality directly affects the learning effect of the subsequent model. Adding Gaussian noise and / or random noise to the original point cloud data to generate noise point cloud data simulates the noise that may be introduced in the actual scene, such as sensor errors and environmental interference, constructs a training data set close to real applications, and enables the subsequent denoising model to have stronger generalization ability.

[0076] The generation of the two types of noise includes Gaussian noise and random noise. The Gaussian noise moves each point in the original point cloud according to the probability rule of Gaussian distribution (normal distribution) to offset the position (such as random movement along the X, Y, and Z axes), but does not add new data points. For example, in an industrial part point cloud, the Gaussian noise can simulate the random error when the laser radar is measured, so that the offset of the points is concentrated around the mean value, which conforms to the distribution characteristics of the real noise. The random noise generates uniformly distributed random numbers in a preset range and superimposes them on the coordinates of the original point cloud to generate irregular small disturbances around each point. For example, in an outdoor scene point cloud, the random noise can simulate the interference of raindrops and dust on the sensor, causing irregular displacement of the position of the points.

[0077] By artificially adding noise, the diversity of the training data is expanded, ensuring that the model can learn the characteristics of different types and intensities of noise and improving the denoising robustness.

[0078] Based on the local outlier factor algorithm, the noise point cloud data is subjected to outlier point elimination, and extreme outlier points in the noise point cloud are removed, such as isolated points generated by sensor false triggering and outliers caused by sudden interference, so as to avoid interference of these points on subsequent clustering and network training. By calculating the local outlier factor (LOF value) of each point, it is determined whether it is an outlier point. The LOF value is calculated by comparing the density difference between the point and the neighborhood points. If the density of the point is much lower than that of the surrounding points, the LOF value is larger, and the point is determined to be an outlier point. After a batch of outlier points with the largest LOF value are removed, the target point cloud data obtained not only retains the effective structure of the original point cloud, but also removes the extreme noise, thereby providing high-quality input for subsequent construction of the denoising network.

[0079] In an implementation manner, the surrounding environment image can be acquired through the visual sensor, image feature extraction and matching are performed by using a computer vision algorithm, and positioning and navigation point cloud data are generated; the noise point cloud data is divided into a training set and a test set, which are used for subsequent construction and generation of the denoising network.

[0080] Meanwhile, the noise point cloud data can also be obtained by adding noise, such as adding Gaussian noise or random noise. Gaussian noise refers to a type of noise whose probability density function obeys Gaussian distribution. In the generation of the simulation measurement point cloud, the noise is to move the existing points, rather than adding new points. Random noise refers to generating random numbers uniformly distributed in a certain range and adding them to the coordinates of the point cloud, so as to generate random disturbances around each point of the point cloud to evaluate the algorithm data. The point cloud data after adding noise is saved and subjected to data set division.

[0081] The application realizes preprocessing and purification of the point cloud data through the process of real data acquisition, noise simulation enhancement and accurate outlier point elimination, solves the problems of complex noise types and large interference of extreme outlier points in the original data, and provides a guarantee for efficient training and accurate inference of the subsequent denoising model.

[0082] Optionally, based on the local outlier factor algorithm, the noise point cloud data is subjected to outlier point elimination, and target point cloud data subjected to outlier point elimination is obtained, including:

[0083] The Mahalanobis distance between any two data points in the noise point cloud data is calculated; the Mahalanobis distance is calculated by a difference vector of the two data points, a transpose of the difference vector and an inverse matrix of a multi-dimensional random variable covariance matrix of the noise point cloud data;

[0084] The k-distance neighborhood of each data point is determined, and the k-reachable distance of each point in the k-distance neighborhood is solved based on the Mahalanobis distance; the k-reachable distance is the maximum value of the k-distance of a first target point and the Mahalanobis distance between the first target point and a second target point.

[0085] According to the k-reach distance, a local reach density of each data point is calculated; the local reach density is the reciprocal of the average value of k-reach distances of all points in the k-distance neighborhood of the data point;

[0086] Based on the local reach density, a local outlier factor of each data point is calculated;

[0087] n data points with the largest local outlier factors are selected as outlier points to form an outlier point set, and the outlier point set is removed from the noise point cloud data to obtain target point cloud data after outlier removal.

[0088] It should be noted that the local outlier factor algorithm is a density-based local outlier detection algorithm, which is mainly used to identify local outlier points in a data set. The LOF algorithm compares the density of each data point in its neighborhood with the density of surrounding points to determine whether the point is an outlier. Referring to C1 and C2 shown in FIG. 1, where C1 represents a low-density area and C2 represents a high-density area, it can be seen that point p is an obvious isolated point. If a traditional algorithm is used to calculate the distance of p from the C1 and C2 areas, p will be classified into the C2 area. However, the LOF algorithm can more accurately determine that p is an outlier based on the distance of p from the C1 and C2 areas and the distribution of surrounding points, and can detect and remove p, thereby improving the accuracy of data detection. Figure 2

[0089] The embodiments of the present application can realize accurate outlier removal based on the local outlier factor algorithm combined with Mahalanobis distance, and can filter and remove extreme noise points by quantifying the outlier degree of each point in the point cloud data. The purpose of calculating the Mahalanobis distance between any two data points in the noise point cloud data is to accurately measure the real distance between points, and to avoid errors caused by dimensional scale differences or correlations in traditional Euclidean distance.

[0090] Mahalanobis distance is obtained by operating the difference vector of two data points, the transpose of the difference vector, and the inverse matrix of the multi-dimensional covariance matrix of the noise point cloud data. For example, in a three-dimensional point cloud (X, Y, Z coordinates), if the X-axis unit is meters and the Y-axis is centimeters (scale difference), or there is a linear correlation between the X and Y coordinates (such as dense distribution along a certain direction), the Mahalanobis distance can correct these influences through the covariance matrix, so that the distance calculation is more consistent with the real distribution of the data. Mahalanobis distance provides a more reliable distance measurement basis for subsequent outlier detection, and is particularly suitable for point cloud data with complex dimensional characteristics.

[0091] ​In the abnormality detection process, the Mahalanobis distance is used to calculate the distance between two points. The Mahalanobis distance is a method to measure the distance between two points, which can be regarded as a modification of the Euclidean distance. It considers the problem of inconsistent and correlated dimensions, and can eliminate the interference of correlation between variables.

[0092] The Mahalanobis distance of a single data point can be expressed by the formula: Where x is the sample to be tested, that is, the sample data point to be calculated; μ is the mean vector, the mean of the sample population; when calculating, first look at the difference between the sample and the mean, then use the covariance inverse matrix to eliminate the influence of scale and correlation, and finally calculate a standardized distance value.

[0093] The Mahalanobis distance between data points x and y is: Where Σ is the covariance matrix of multi-dimensional random variables. If the covariance matrix is a unit vector, that is, each dimension is independent and identically distributed, the Mahalanobis distance becomes the Euclidean distance.

[0094] In the process of determining the k-distance neighborhood of each data point and solving the k-reachable distance, the k-distance neighborhood represents that for each data point, the k nearest points (based on the Mahalanobis distance calculated above) are selected to form the local neighborhood of the point, such as k = 5, which means that each point refers to the nearest 5 neighbors. The k-reachable distance represents that for any two points (the first target point and the second target point) in the neighborhood, the k-reachable distance of the first target point is taken as the k-distance of the point to the kth nearest neighbor, and the maximum value in the Mahalanobis distance between the two points. For example, if the k-distance of point A is 0.5 meters, and the Mahalanobis distance between points A and B is 0.3 meters, then the k-reachable distance between A and B is 0.5 meters (taking the maximum value); if the Mahalanobis distance between A and B is 0.6 meters, then the k-reachable distance is 0.6 meters. By taking the maximum value, the application avoids misjudgment of local density due to too close distance of individual points, and is more consistent with the actual distribution of local neighborhood density.

[0095] Further, the local reachable density is the inverse of the average value of the k-reachable distances of all points in the k-distance neighborhood of the data point. For example, if the k-reachable distances of the 5 neighborhood points of point P are 0.3, 0.4, 0.5, 0.4, and 0.3, the average value is 0.38, then the local reachable density of point P is 1 / 0.38≈2.63. Wherein, the higher the value, the higher the local area density where the point is located, and the neighborhood points are close and dense; the lower the value, the lower the local density, which may be a noise point. By quantifying the local density of each point, a density benchmark is provided for judging whether it is an abnormal point.

[0096] The local outlier factor is the average value of the ratio of the local reachable density of all points in the k-distance neighborhood of the point to the local reachable density of the point itself. For example, if the local reachable density of point P is 2.63, the local reachable densities of the 5 neighborhood points of P are 3.0, 2.8, 3.2, 2.9, and 3.1, respectively, and the average ratio is (3.0 / 2.63+2.8 / 2.63+...+3.1 / 2.63) / 5≈1.1, then the LOF value of point P is 1.1.

[0097] Judgment logic: if LOF≈1, it indicates that the point has a density similar to that of the neighborhood points, and is a normal point; if LOF>>1 (such as LOF>2), it indicates that the local density of the point is much lower than that of the neighborhood points, and is an outlier (i.e., noise). By comparing the density with that of the neighborhood points, the degree of outlying of each point is accurately quantified.

[0098] In the process of screening and removing outliers and obtaining target point cloud data, the n data points with the largest local outlier factor are selected to form an outlier set. These points are the points with the lowest local density and the most likely extreme noise. Removing the outlier set from the noise point cloud data, the remaining data is the target point cloud data after removing outliers. In this way, the extreme noise that interferes with subsequent processing is completely removed, and the main structure and effective features of the point cloud are retained.

[0099] The above process realizes accurate identification and removal of outliers in noise point cloud data through the layer-by-layer progressive logic of Mahalanobis distance correction, local neighborhood analysis, density comparison quantification, and outlier screening, and provides cleaner and more reliable input data for the training of subsequent denoising networks.

[0100] In a specific implementation, the LOF algorithm calculates the local reachable density of each point, and then calculates the local outlier factor of each point, and selects the n points with the highest outlying degree. First, the k-distance neighborhood of each point is calculated:

[0101] wherein, N k (p) represents the k-distance neighborhood of point p (all points with a distance of ≤k-distance(p) from p). |N k (p) | represents the number of points in the neighborhood (≥k, because there can be multiple points with the same kth nearest neighbor distance). reach_dist k (o, p) represents the reachable distance of point p to o. It can be understood that the smaller the average reachable distance, the larger the LRD (the position of the point is more dense). The larger the average reachable distance, the smaller the LRD (the position of the point is more sparse).

[0102] The k-local outlier factor of each point is calculated:

[0103] where, |N k (p) | denotes the number of points in the neighborhood (≥ k, because there can be multiple points with the kth nearest neighbor distance equal). LRD k (p) denotes the local reachable density of point p. For each neighbor point o of p, the ratio of the density of the neighbor to the density of p is calculated, and then the average is taken; if the result is approximately 1, it means that the density of p is comparable to that of the neighbor, and p is a normal point. If it is >> 1, it means that the density of p is significantly lower than that of the neighbor, and p is in a sparse area and is a possible outlier. If it is < 1, it means that p is more dense than the neighbor, and p can be in the central area of a cluster. For the n largest local outlier factors belonging to the data points (selected), the outlier point set is output for removal: O = {o1, o2, o3, …, on}; the application calculates the outlier value for each point in the point cloud based on the density, local features, etc. of the point, calculates the outlier points and removes them. The removal operation is performed on the points in the set O, and these outlier points are removed from the original point cloud, so as to improve the quality of the point cloud data and the subsequent processing effect.

[0104] Optionally, the step 12 comprises:

[0105] A hierarchical clustering algorithm based on density is used to cluster the target point cloud data, and a clustering result is obtained.

[0106] An elite strategy of business selection strategy is introduced, and an individual with the best performance is selected as an elite point from the clustering cluster corresponding to the clustering result to obtain elite point information; the elite point is the center or the most representative data point of each clustering cluster.

[0107] A denoising network is constructed based on the clustering result and the elite point information.

[0108] The training set of the target point cloud data is input into the denoising network for training, and the network parameters are adjusted through the validation set of the target point cloud data to generate a denoising model.

[0109] In the embodiments of the application, the point cloud data is divided into different clustering clusters according to local density and structural features, and the local characteristics of the point cloud are accurately identified to provide a basis for subsequent differential denoising. The application uses a hierarchical clustering algorithm based on density, such as HDBSCAN, to aggregate points with similar densities into clusters by calculating the mutual reachable distance between points, and to retain the hierarchical structure of the clusters, such as a large cluster containing a small cluster. For example, in a three-dimensional object point cloud, the algorithm can automatically "aggregate surface smooth areas into one cluster, edge and corner areas into another cluster, and sparse noise areas into small clusters. A plurality of clustering clusters with different densities and structural features are obtained, as well as core points of each cluster, and the core points represent the central area of the cluster to provide a structured input for subsequent path decision-making.

[0110] The most representative points in the clustering results are screened out as anchor points for subsequent network construction and path optimization, avoiding noise points or secondary points from interfering with decision-making, and improving the accuracy of path selection in extreme scenarios. The elite point selection logic includes: selecting the individual with the best performance from each clustering cluster as the elite point, which is usually the center of the cluster, such as the point with the highest density, or the point that best represents the characteristics of the cluster, such as the vertex in the edge cluster. For example, in a smooth area cluster, the elite point may be the geometric center of the area; in an edge cluster, the elite point may be the vertex with the maximum curvature. As core reference points, subsequent network expansion and path selection will prioritize around these points, ensuring that the model focuses on the key structures of the point cloud and reduces ineffective learning of noise areas.

[0111] A network is constructed that can dynamically adjust the denoising strategy according to the local features of the point cloud, solving the problem of poor adaptability of traditional single paths in complex scenarios. Differentiated denoising path sub-networks are designed for different clustering clusters, such as smooth areas, sharp edges, and sparse areas. For example, high-complexity paths are configured for sharp edge clusters (to preserve details), and low-complexity paths are configured for smooth area clusters (to improve efficiency).

[0112] The denoising block is composed of a shared sub-network (extracting common features of each cluster) and a denoising path sub-network, and multiple denoising blocks are connected according to the clustering hierarchy to form a complete network. The elite point information is used to optimize the network weights, giving higher weights to the features of the elite points and their neighbors, enhancing the network's learning ability for core structures.

[0113] The denoising network constructed in this application has dynamic adaptation capability and can automatically select the optimal denoising path according to the local features of the input point cloud, especially in extreme scenarios. Through data-driven optimization of network parameters, the model has stable denoising ability and good generalization on unseen data.

[0114] Optionally, the training process includes:

[0115] The target point cloud data is divided into a training set (point cloud with noise) and input into the denoising network, which outputs the denoised result through dynamic path processing. During training, the model performance is evaluated periodically using a validation set, such as quantifying the denoising effect using PSNR and SSIM indicators, and monitoring whether it is overfitting, such as good training set but poor validation set. According to the validation results, adjust the hyperparameters such as learning rate, network layers, batch size, and optimize the network performance. Finally, the generalization ability of the model is verified using an independent test set, and the final denoising model is determined after evaluation. This application enables the network to learn the denoising rules of different noise types and point cloud structures by learning a large amount of labeled data, ensuring that high-quality denoised results are output in actual applications.

[0116] The application solves the problems of blind path decision and low denoising accuracy of traditional methods in extreme cases through a collaborative mechanism of clustering feature division, elite point anchoring core, dynamic network adaptation scene and data-driven optimization parameter, and finally generates a denoising model that can accurately adapt to complex point cloud structures.

[0117] Optionally, a hierarchical clustering algorithm based on density is used to cluster the target point cloud data to obtain a clustering result, including:

[0118] The mutual reachable distance of each point in the target point cloud data is calculated; the mutual reachable distance is the maximum value of the core distance of a first position point, the core distance of another reference second position point, and the distance between the first position point and the second position point; the core distance is the distance between the first position point and the kth nearest sample point;

[0119] Based on the mutual reachable distance, a minimum cluster sample point number is set, and a minimum tree with the mutual reachable distance as the weight is generated;

[0120] The weights of the minimum tree are sorted in ascending order, and a hierarchical clustering algorithm based on density is used to preferentially merge the cluster with the smallest weight, and the merging operation is iteratively performed until all clusters are merged to the highest weight state, and a hierarchical structure of the clusters is constructed;

[0121] A minimum cluster threshold is set, and the hierarchical structure is traversed, and a judgment is made at each split; if the number of sample points contained in the new cluster after splitting is less than the minimum cluster threshold, the new cluster is marked as an outlier;

[0122] Based on the reciprocal of the mutual reachable distance, the stability of each cluster is calculated, and the cluster and the corresponding core point that meet the preset condition are extracted as the clustering result.

[0123] Optionally, based on the clustering result and the elite point information, a denoising network is constructed, including:

[0124] For the local features of each cluster in the clustering result, the position of the elite point information is combined to configure a corresponding denoising path subnetwork;

[0125] A denoising block composed of a shared subnetwork and the denoising path subnetwork is constructed; the shared subnetwork is used to extract common features of each cluster; the denoising path subnetwork is used to perform differential denoising operations for different cluster characteristics;

[0126] Based on the elite point information, the network feature weight of the denoising network to be constructed is optimized;

[0127] The plurality of denoising blocks are sequentially connected according to a clustering hierarchy to form a complete denoising network; wherein the output features of a previous denoising block are taken as the input features of a subsequent denoising block, and a displacement vector is regressed through a max-pooling layer and a fully connected layer to correct and output the point cloud coordinates.

[0128] In the embodiments of the present application, through the logic of feature adaptation, module combination, weight optimization and network integration, dynamic adaptation to different point cloud structures is realized, and the problem of insufficient processing capacity of a single path in a traditional network is solved. For the local features of each cluster, a denoising path subnetwork is configured in combination with the position of the elite point, and a dedicated denoising strategy is designed for regions with different structural characteristics in the point cloud, so as to avoid feature loss or incomplete denoising caused by one-size-fits-all processing. Based on the clustering results, the local features of each cluster are identified, such as determining whether it is a sharp region or a smooth region by curvature calculation, and determining whether it is a dense core region or a sparse transition region by density distribution. In combination with the position of the elite point, the center or representative point of each cluster, the spatial range of the feature region is accurately positioned. The denoising path subnetwork corresponding to different features is configured for the cluster. The network has the ability to process specifically, and lays a foundation for subsequent differentiated denoising.

[0129] The denoising block composed of the shared subnetwork and the denoising path subnetwork can balance common feature extraction and individual feature processing, improve denoising accuracy while ensuring processing efficiency. The shared subnetwork is used to represent the basic feature extraction module shared by all clusters, which is responsible for extracting the general features of the point cloud, avoiding repeated calculation and reducing network complexity. The denoising path subnetwork is a dedicated module configured for each cluster, which performs differentiated denoising operations on the individual features of the cluster based on the common features output by the shared subnetwork.

[0130] Each denoising block is a combination of a shared subnetwork and a specific denoising path subnetwork, which serves as a basic unit for network processing and realizes complete processing of a type of feature cluster.

[0131] Optimizing the feature weights of the denoising network based on the elite point information is used to enhance the attention of the network to the core feature region, avoid the interference of noise or secondary regions on model learning, and improve the decision accuracy in extreme scenarios. The elite point is the "core representative" of each cluster, and the region where it is located is usually the key structure of the point cloud. During network training and feature extraction, higher weights are given to the features in the neighborhood range of the elite point, so that the network prioritizes learning and retaining effective information in these regions, and weakens the feature influence of noise regions. The present application guides the network to focus on the core structure of the point cloud, and improves the sensitivity and denoising stability of the model to key features.

[0132] Connecting multiple denoising blocks forms a complete denoising network, realizes coordinate correction output, integrates the functions of each module, forms an end-to-end denoising process, and finally outputs the denoised point cloud coordinates.

[0133] According to the hierarchical structure of clustering, first process large-scale clusters, then process subdivided small clusters, and sequentially connect multiple denoising blocks, the output features of a previous denoising block are taken as the input of a subsequent denoising block, to realize progressive optimization of the features. After processing through all the denoising blocks, the feature dimension is compressed through a max-pooling layer, and then the displacement vector is regressed through two fully connected layers, to obtain the correction amount of each point cloud coordinate. The displacement vector obtained by regression is superimposed on the original point cloud coordinates to obtain the denoised point cloud data, and the displacement of the noise points is corrected, and the features of the effective points are retained. Through multi-level processing and coordinate correction, complete conversion from the noisy point cloud to the clean point cloud is realized, to ensure that the denoising result is accurate and retains the original structural features.

[0134] Further, the training set divided from the target point cloud data is input into the denoising network for training, and the network parameters are adjusted through the validation set divided from the target point cloud data to generate a denoising model, including:

[0135] The training set divided from the target point cloud data is input into the denoising network, and the features of the noisy point cloud are extracted and processed through the denoising network to output a preliminary denoising result.

[0136] During the training process, the validation set divided from the target point cloud data is input into the network at regular intervals to calculate the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) to determine the evaluation result of the validation set.

[0137] According to the evaluation result of the validation set, the network hyperparameters are adjusted until the denoising effect of the model on the validation set reaches a preset standard to generate a denoising model. The network hyperparameters at least include the learning rate, the training batch size, the number of denoising blocks, and the number of network layers.

[0138] In the embodiments of the present application, the process of this step is the training and optimization phase of the denoising model. Through the closed-loop process of data input, effect evaluation, and parameter optimization, the model is ensured to have good generalization ability while fitting the training data, and finally a stable and reliable denoising model is generated.

[0139] The training set is input into the denoising network to output a preliminary denoising result, so that the network learns the noisy point cloud data and masters the noise features and denoising rules to form a preliminary denoising ability. In the present application, the training set divided from the target point cloud data, which contains point cloud data with different types and intensities of noise, is input into the constructed denoising network. The denoising network extracts common features of the point cloud through the shared sub-network, and then performs differential processing on different clusters through each denoising path sub-network, such as retaining details in sharp areas and efficiently denoising in smooth areas, to finally output the preliminary denoised point cloud data. In this way, through the training of a large amount of noisy data, the network parameters gradually approach the optimal solution to form the ability of recognizing and correcting noise.

[0140] The training effect of the monitoring model is monitored to timely find the overfitting (the model only fits the training data and performs poorly on new data) or underfitting (the model does not learn the data law sufficiently) problem. In the training process, the validation set (noise point cloud not involved in training) divided from the target point cloud data is input into the network at a fixed period, such as every 10 rounds of training, to obtain the denoising result of the validation set.

[0141] Here, two core indicators can be selected: PSNR (peak signal-to-noise ratio): quantifying the pixel difference between the denoising result and the real clean point cloud, the higher the value, the more complete the noise removal; SSIM (structural similarity index): measures the structural consistency of the denoising result and the real point cloud, the closer the value to 1, the more complete the original features retained. The application evaluates the generalization ability of the model through objective indicators to provide a basis for subsequent parameter adjustment.

[0142] According to the evaluation result, the hyperparameters are adjusted, which can be used to optimize the network performance, so that the generated denoising model balances between denoising effect (PSNR or SSIM) and generalization ability. Among them, if the PSNR or SSIM of the validation set continues to rise and tends to be stable, it indicates that the model training is normal; if the indicators stagnate or decrease, the network hyperparameters need to be adjusted: the step size of parameter update, such as the learning rate, is too large, which can be reduced to 1 / 10 of the original; the learning rate is too small, which can be appropriately increased; the number of samples input into the network each time, such as the batch size, is too large, which can be reduced; the batch size is too small, which can be increased; the complex scene needs to increase the denoising block or the number of layers to improve the feature extraction ability, and the simple scene can reduce it to improve the efficiency.

[0143] Repeat the adjustment until the PSNR and SSIM of the validation set reach the preset standard, and the model at this time is the final denoising model.

[0144] Through hyperparameter optimization, the application solves the problems of overfitting and slow convergence in model training, and ensures that the model can efficiently denoise and retain key structural features of the point cloud in actual application. Through the cyclic mechanism of training, evaluation and optimization, the denoising network learns from the beginning to gradually optimize, and finally generates a model that performs well in noise removal effect and structure retention ability, providing reliable support for subsequent actual point cloud denoising tasks.

[0145] Referring to Figure 3 The specific process of constructing the denoising model includes:

[0146] Point cloud training set construction: first, the point cloud training set is prepared to provide a data basis for subsequent processing.

[0147] Point cloud denoising network construction: Based on the training set, a point cloud denoising network is constructed for removing noise interference in point cloud data.

[0148] Path decision: Use the denoised point cloud data for path decision to generate an initial path scheme.

[0149] Optimization model judgment: Evaluate the current path decision result to determine whether the model needs to be optimized.

[0150] If optimization is needed: Enter the "optimization model" link to adjust and optimize the model parameters, structure, etc.; after optimization, enter the "end" node to complete the process. If optimization is not needed: Directly enter the "end" node and terminate the process.

[0151] Update optimization: If the denoising network or process logic needs to be iteratively optimized after the "optimization model" link, it can be traced back to the "point cloud denoising network construction" through the "update optimization" link to realize closed-loop iteration of the process.

[0152] This process takes "point cloud data processing, path decision, and model optimization" as the core link, realizes iterative optimization through the "optimization model judgment" node, ensures the accuracy and robustness of path decision, and finally completes the process closed loop.

[0153] Specifically, the present application optimizes and constructs the denoising network based on HDBSCAN-NSGA to solve the path decision selection optimization problem in extreme cases in the point cloud denoising process. First, obtain the point cloud data with added noise to form a training set and a test set of noise point cloud, then construct the denoising network HDNCN-NSGA, and optimize the point cloud selection strategy by adding an elite strategy (NSGA) in the denoising network.

[0154] In the process of constructing the denoising network, the density clustering algorithm HDBSCAN is used to preliminarily cluster the data set, and the technology of extracting planar clusters is used to identify the main core points and clustering clusters; after clustering, the best individuals are selected to ensure that they are recorded as "elite points" in the next generation. These points may represent the center or most representative data points of each cluster.

[0155] HDBSCAN-NSGA can handle clustering problems with different densities and increase the elite selection strategy for optimization. It uses a mutual reachable distance metric, after generating a minimum tree, uses a hierarchical clustering method to aggregate clusters, and uses a compressed cluster tree method to filter outliers in the later stage. The steps are as follows: 1) Transform the density space. By constructing the core distance as the mutual reachable distance of each sample point as the core point, the low-density points are dispersed. The mutual reachable distance of points to points is defined as: d mreach-k (a,b)=max{corek (a),core k (b),d(a,b)};where core k (a) is the distance between the kth nearest sample point to the core point a, and d(a,b) is the distance between a and b. This method allows the denser samples to be clustered together more easily, while separating the clusters further apart. By taking the maximum of the three, the mutual reachable distance balances the local density and the actual distance between two points, avoiding the clustering bias caused by the extremely small distance between points in dense areas.

[0156] 2) Generate the minimum tree. Two points can determine an edge, and the mutual reachable distance is the edge of the sample points in the space, which is also defined as the weight between any two vertices. Set the number of sample points in the minimum cluster k, and use the Prim algorithm to determine the minimum tree.

[0157] 3) Hierarchical construction of the cluster hierarchy. For each minimum tree, the same strategy as the minimum distance hierarchical clustering method is used. By sorting the weights in ascending order, the nearest minimum tree is merged, and this process is iterated until the highest weight is merged.

[0158] 4) Compress the clustering tree. Use the horizontal line to split the clustering cluster shown in the figure to determine the minimum cluster size k. Traverse the hierarchy and determine at each split: whether the size of the new cluster split will be smaller than the predefined k, if so, mark these sample points as outlier points.

[0159] 5) Extract the cluster. Before selecting the cluster, the reciprocal of the mutual reachable distance is used to measure the stability of the cluster. For each cluster, the stability can be defined as: ∑pEcluster(λ p -λ birth );where λ p is the λ value when the point is removed from the cluster. λ birth is the λ value after the current cluster is successfully split and becomes a sub-cluster, λ = 1 / d mreach-k In the subsequent clustering expansion, the data points adjacent to the elite points are given priority, thereby speeding up the clustering process and avoiding excessive expansion of noise points.

[0160] 5) Denoising block. According to the characteristics of the cluster, different complexity denoising paths are provided. For example, points in sharp areas need to retain local sharp features when denoising; while point clusters in segmented smooth areas can be correspondingly reduced. The purpose of the denoising block is to provide different complexity denoising paths for each point cluster. Specifically, the mth denoising block consists of a shared subnetwork and a denoising path subnetwork. The mth denoising block g m is constructed as follows:

[0161]

[0162] Among them G m and G m-1 These are the input and output features of the m-th denoised block, respectively. Subnetwork g m and They are all KANs (Kolmogorov-Arnold Networks), g m This represents the path selected from the denoising path subnetwork. Once the output features of the t-th denoised block are obtained (1≤t≤T), the displacement vector is regressed by applying max pooling and two fully connected layers. Then, the network outputs the denoised point cloud by adding the regressed displacement vector to the original point cloud.

[0163] (4) Generate the denoising network. Input the training dataset into the denoising network for evaluation and processing. During training, periodically evaluate the model's performance using a validation set to monitor for overfitting. Calculate metrics such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) to quantify the denoising effect. Adjust hyperparameters such as learning rate, batch size, and number of network layers based on the evaluation results to improve model performance. Evaluate the final model using an independent test set to ensure its generalization ability on unseen data.

[0164] (5) Point cloud denoising. Input the new noisy image into the trained denoising network to obtain the image after denoising according to the current denoising network. Further process the output result as needed, such as adjusting brightness and contrast.

[0165] In one specific embodiment, reference is made to... Figure 4 The schematic diagram of the overall point cloud denoising process shown includes:

[0166] Point Cloud Data Acquisition: As the starting point of the process, the initial acquisition of point cloud data is completed, providing the basic data source for subsequent processing. Local Outlier Removal: The acquired point cloud data is preprocessed to identify and remove local outliers (such as noise points, erroneous acquisition points, etc.), reducing interference factors in subsequent processing. Denoising Optimization Model Construction: Based on the preprocessed data, an optimized model specifically for point cloud denoising is designed and constructed, providing the algorithmic framework and parameter basis for subsequent model generation. Denoising Model Generation: Using the constructed optimized model, the denoising model is generated, enabling it to perform denoising processing on the point cloud data. Point Cloud Denoising: In the final stage, the generated denoising model is used to perform denoising operations on the point cloud data, achieving noise removal and quality optimization. This process systematically completes the denoising processing of point cloud data through five steps: acquisition, preprocessing, model construction, model generation, and denoising, ensuring the accuracy and usability of the point cloud data.

[0167] The point cloud denoising method provided in the application firstly acquires point cloud data through point cloud data collection, and secondly removes abnormal points through a local outlier factor algorithm; then a new denoising network is constructed and optimized, a suitable denoising strategy is dynamically selected according to the density and local features of the point cloud, and a denoising model is generated, so as to effectively improve the quality of the point cloud data and be suitable for most point cloud denoising scenes.

[0168] In summary, the denoising network construction provided in the application uses the density clustering algorithm HDBSCAN-NSGA to optimize and construct the denoising network, and solves the path decision selection optimization problem in the extreme case of the point cloud denoising process. First, the point cloud data with added noise is acquired to form a training set and a test set of noise point clouds, and then a denoising network HDNCN-NSGA is constructed, and the point cloud selection strategy is optimized by adding an elite strategy (NSGA) in the denoising network.

[0169] The local outlier point removal provided in the application uses the local outlier factor algorithm (LOF algorithm) to assign each data point a local outlier factor LOF dependent on the neighborhood density, and introduces a more accurate and comprehensive Mahalanobis distance to judge the distance between points and points, calculate the local density and distance deviation of each data point to detect abnormal points, find the outlier points in the point cloud data and remove them. The application removes the abnormal points in advance by using the local outlier factor algorithm, reduces the influence of the edge nodes on the denoising network, and proposes a new density clustering algorithm for denoising of abnormal point cloud data, and then different denoising schemes are proposed according to different abnormal scenes, so as to achieve better denoising effect.

[0170] The application based on the HDBSCAN algorithm introduces the elite strategy of the non-dominated sorting genetic algorithm (NSGA) in the denoising network, optimizes the selection strategy of the point cloud, effectively improves the clustering performance and denoising ability of the algorithm, and realizes more accurate and robust point cloud denoising effect.

[0171] The above introduces various methods of the embodiments of the application. The following will further provide a device for implementing the above method.

[0172] Please refer to Figure 5 The embodiments of the application also provide a point cloud denoising device, which comprises:

[0173] The acquisition module 51 is configured to acquire target point cloud data for abnormal point removal;

[0174] The first processing module 52 is configured to construct a denoising network based on a density-based hierarchical clustering algorithm and an elite selection strategy according to the target point cloud data, and generate a denoising model based on the denoising network;

[0175] The second processing module 53 is configured to input the to-be-de-noised point cloud data into the de-noising model, and output de-noised point cloud data.

[0176] Optionally, the acquisition module 51 comprises:

[0177] The acquisition unit is configured to acquire original point cloud data.

[0178] The first processing unit is configured to add Gaussian noise and / or random noise to the original point cloud data to generate noise point cloud data; the Gaussian noise is generated by moving points in the original point cloud according to a Gaussian distribution rule, without adding new data points; the random noise is generated by generating random numbers uniformly distributed in a preset range and superimposing the random numbers on the coordinates of the original point cloud, so that random disturbance is generated around each point.

[0179] The first acquisition unit is configured to perform outlier point elimination on the noise point cloud data based on a local outlier factor algorithm, and acquire target point cloud data after outlier point elimination.

[0180] Optionally, the first acquisition unit is specifically configured to:

[0181] Calculate the Mahalanobis distance between any two data points in the noise point cloud data; the Mahalanobis distance is calculated by a difference vector of the two data points, a transpose of the difference vector, and an inverse matrix of a multi-dimensional random variable covariance matrix of the noise point cloud data;

[0182] Determine a k-distance neighborhood of each data point, and solve a k-reachable distance of each point in the k-distance neighborhood based on the Mahalanobis distance; the k-reachable distance is a maximum value of the k-distance of a first target point and the Mahalanobis distance between the first target point and a second target point;

[0183] Calculate a local reachable density of each data point according to the k-reachable distance; the local reachable density is a reciprocal of an average value of the k-reachable distances of all points in the k-distance neighborhood of the data point;

[0184] Calculate a local outlier factor of each data point based on the local reachable density;

[0185] Select n data points with the largest local outlier factors as outlier points to form an outlier point set, remove the outlier point set from the noise point cloud data, and acquire the target point cloud data after outlier point elimination.

[0186] Optionally, the first processing module 52 comprises:

[0187] The second acquisition unit is configured to perform clustering on the target point cloud data by using a density-based hierarchical clustering algorithm, and acquire a clustering result.

[0188] The second processing unit is configured to introduce an elite strategy of the operation selection strategy, select an individual with the best performance as an elite point from a cluster corresponding to the clustering result, and obtain elite point information; the elite point is a center or a most representative data point of each cluster;

[0189] The construction unit is configured to construct a denoising network based on the clustering result and the elite point information.

[0190] The third processing unit is configured to input a training set of the target point cloud data into the denoising network for training, adjust network parameters through a validation set of the target point cloud data, and generate a denoising model.

[0191] Optionally, the second obtaining unit is specifically configured to:

[0192] Calculate a mutual reachable distance of each point in the target point cloud data; the mutual reachable distance is a maximum value of a core distance of a first position point, a core distance of a second position point of another reference, and a distance between the first position point and the second position point; the core distance is a distance between the first position point and a kth nearest sample point;

[0193] Based on the mutual reachable distance, set a minimum cluster sample point number, and generate a minimum tree with the mutual reachable distance as a weight;

[0194] Sort the weights of the minimum tree in an ascending order, perform a preferential merging on a cluster with the smallest weight by using a density hierarchical clustering algorithm, and iteratively perform a merging operation until all clusters are merged into a state with the highest weight, to construct a hierarchical structure of the clusters;

[0195] Set a minimum cluster threshold, traverse the hierarchical structure, and perform a judgment at each time of splitting; if a new cluster split out contains a sample point number less than the minimum cluster threshold, the new cluster is marked as an outlier;

[0196] Based on a reciprocal of the mutual reachable distance, calculate a stability of each cluster, and extract a cluster and a corresponding core point with a stability satisfying a preset condition as the clustering result.

[0197] Optionally, the construction unit is specifically configured to:

[0198] For a local feature of each cluster in the clustering result, combine a position of the elite point information, and configure a corresponding denoising path subnetwork;

[0199] Construct a denoising block composed of a shared subnetwork and the denoising path subnetwork; the shared subnetwork is configured to extract common features of the clusters; and the denoising path subnetwork is configured to perform a differential denoising operation according to characteristics of different clusters.

[0200] based on the elite point information, optimizing network feature weights of a denoising network required to be constructed;

[0201] a plurality of the denoising blocks are connected in turn according to the clustering hierarchy to form a complete denoising network; wherein, output features of a previous denoising block are taken as input features of a subsequent denoising block, a displacement vector is regressed through a max-pooling layer and a fully connected layer, and a point cloud coordinate is corrected and output.

[0202] Optionally, the third processing unit is specifically used for:

[0203] the training set of the target point cloud data is input into the denoising network, and noisy point cloud features are extracted and processed through the denoising network to output a preliminary denoising result;

[0204] in the training process, the verification set of the target point cloud data is input into the network at regular intervals, a peak signal-to-noise ratio (PSNR) and a structural similarity index (SSIM) are calculated, and an evaluation result of the verification set is determined;

[0205] according to the evaluation result of the verification set, network hyperparameters are adjusted until a denoising effect of the model on the verification set reaches a preset standard, and a denoising model is generated; the network hyperparameters at least include a learning rate, a training batch size, a number of denoising blocks, and a number of network layers.

[0206] It should be noted that the device in this embodiment corresponds to the point cloud denoising method described above, and the implementation modes in the above embodiments are applicable to the embodiments of the device and can achieve the same technical effects. The device described above provided in the embodiments of the present application can implement all method steps achieved by the method embodiments and can achieve the same technical effects. Therefore, the same parts and beneficial effects in the method embodiments will not be described in detail here.

[0207] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program. When the computer program is executed by a processor, each process of the point cloud denoising method embodiments described above is implemented, and the same technical effects can be achieved. To avoid repetition, this will not be described in detail here. The computer readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0208] The embodiments of the present application also provide a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, each process of the point cloud denoising method embodiments described above is implemented, and the same technical effects can be achieved. To avoid repetition, this will not be described in detail here.

[0209] It should be noted that, in the present document, the terms "comprising", "containing" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not only include those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitations, an element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0210] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned example methods can be realized by means of software and a necessary general hardware platform, and of course, they can also be realized by hardware, but in many cases, the former is a better implementation. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product in essence or in the part that contributes to the prior art, and the computer software product is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk), and includes a plurality of instructions for causing a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in the various embodiments of the present application.

[0211] The embodiments of the present application are described above in combination with the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative and not restrictive, and those of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope protected by the claims.

Claims

1. A point cloud denoising method, characterized in that, include: Obtain the target point cloud data for outlier removal; Based on the target point cloud data, a denoising network based on density-based hierarchical clustering algorithm and business selection strategy is constructed, and a denoising model is generated based on the denoising network. Input the point cloud data to be denoised into the denoising model, and output the denoised point cloud data.

2. The method according to claim 1, characterized in that, Obtain the target point cloud data for outlier removal, including: Collect raw point cloud data; Gaussian noise and / or random noise are added to the original point cloud data to generate noisy point cloud data; wherein, the Gaussian noise is generated by moving the points in the original point cloud according to the Gaussian distribution law, without adding new data points; the random noise is generated by generating uniformly distributed random numbers within a preset range and superimposing them on the coordinates of the original point cloud to create random disturbances around each point. Based on the local anomaly factor algorithm, outlier points are removed from the noisy point cloud data to obtain the target point cloud data for outlier removal.

3. The method according to claim 2, characterized in that, Based on the local anomaly factor algorithm, outlier removal is performed on the noisy point cloud data to obtain target point cloud data for outlier removal, including: Calculate the Mahalanobis distance between any two data points in the noisy point cloud data; the Mahalanobis distance is calculated using the difference vector between the two data points, the transpose of the difference vector, and the inverse of the covariance matrix of the multidimensional random variables in the noisy point cloud data. Determine the k-distance neighborhood of each data point, and based on the Mahalanobis distance, solve the k-reachable distance of each point within the k-distance neighborhood; the k-reachable distance is the maximum value of the k-distance of the first target point and the Mahalanobis distance between the first target point and the second target point; Based on the k-reachable distance, calculate the local reachability density of each data point; the local reachability density is the reciprocal of the average k-reachable distance of all points in the k-distance neighborhood of the data point; Based on the local reachability density, the local outlier factor for each data point is calculated; The n data points with the largest local outlier factors are selected as outliers to form an outlier set. The outlier set is removed from the noisy point cloud data to obtain the target point cloud data for outlier removal.

4. The method according to claim 1, characterized in that, Based on the target point cloud data, a denoising network based on a density-based hierarchical clustering algorithm and a business selection strategy is constructed. Based on the denoising network, a denoising model is generated, including: The target point cloud data is clustered using a density-based hierarchical clustering algorithm to obtain the clustering results; An elite strategy that introduces business selection strategies selects the best-performing individual from the clusters corresponding to the clustering results as elite points and obtains elite point information; the elite point is the center or the most representative data point of each cluster. Based on the clustering results and the elite point information, a denoising network is constructed; The training set of the target point cloud data is input into the denoising network for training, and the network parameters are adjusted using the validation set of the target point cloud data to generate a denoising model.

5. The method according to claim 4, characterized in that, A density-based hierarchical clustering algorithm is used to cluster the target point cloud data, and the clustering results are obtained, including: Calculate the mutual reachability distances of all points in the target point cloud data; the mutual reachability distances are the maximum value among the core distance of a first location point, the core distance of a second reference location point, and the distance between the first location point and the second location point; the core distance is the distance between the first location point and the k-th nearest sample point; Based on the mutual reachability distance, a minimum number of cluster sample points is set, and a minimum tree with mutual reachability distance as the weight is generated; The weights of the minimum tree are sorted in ascending order. A density-based hierarchical clustering algorithm is used to merge the clusters with the smallest weights first. The merging operation is performed iteratively until all clusters are merged to the state with the highest weight, thus constructing a hierarchical structure of clusters. Set a minimum cluster threshold, traverse the hierarchical structure, and make a judgment at each split. If the number of sample points contained in the new cluster is less than the minimum cluster threshold, then mark the new cluster as an outlier. Based on the reciprocal of the mutually reachable distance, the stability of each cluster is calculated, and clusters whose stability meets the preset conditions and their corresponding core points are extracted as clustering results.

6. The method according to claim 4, characterized in that, Based on the clustering results and the elite point information, a denoising network is constructed, including: For the local features of each cluster in the clustering results, and in combination with the location of the points, a corresponding denoising path sub-network is configured; A denoising block is constructed, consisting of a shared subnetwork and a denoising path subnetwork; the shared subnetwork is used to extract common features of each cluster; the denoising path subnetwork is used to perform differentiated denoising operations based on the characteristics of different clusters. Based on the elite point information, optimize the network feature weights of the denoising network to be constructed; Multiple denoising blocks are sequentially connected according to a clustering hierarchy to form a complete denoising network; wherein, the output features of the previous denoising block are used as the input features of the next denoising block, and the displacement vector is regressed through a max pooling layer and a fully connected layer to correct the output of the point cloud coordinates.

7. The method according to claim 4, characterized in that, The training set of the target point cloud data is input into the denoising network for training, and the network parameters are adjusted using the validation set of the target point cloud data to generate a denoising model, including: The training set of the target point cloud data is input into the denoising network. The denoising network extracts and processes the features of the noisy point cloud and outputs the preliminary denoising results. During training, the validation set of the target point cloud data is periodically input into the network to calculate the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) to determine the evaluation results of the validation set. Based on the evaluation results of the validation set, the network hyperparameters are adjusted until the denoising effect of the model on the validation set reaches the preset standard, and a denoised model is generated; the network hyperparameters include at least the learning rate, training batch size, number of denoised blocks, and number of network layers.

8. A point cloud noise reduction device, characterized in that, include: The acquisition module is used to acquire the target point cloud data for anomaly removal. The first processing module is used to construct a denoising network based on a density-based hierarchical clustering algorithm and a business selection strategy according to the target point cloud data, and to generate a denoising model based on the denoising network. The second processing module is used to input the point cloud data to be denoised into the denoising model and output the denoised point cloud data.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images