Adaptive joint filtering method based on local feature and spatial distance relationship

Abstract

The application discloses an adaptive joint filtering method based on local feature and spatial distance relationship, belongs to the technical field of point cloud denoising, and is used for point cloud denoising, comprising the following steps: after filtering out large-scale noise by using adaptive statistical filtering, adaptive bilateral filtering is used for noise suppression; the adaptive bilateral filtering comprises the following steps: searching for neighborhood points of the point cloud, calculating local curvature and local density, adjusting the spatial domain and attribute domain parameters of the bilateral filtering based on the local curvature and the local density, introducing the local feature into a weight parameter, normalizing and adjusting the weight parameter, and updating the geometric position of a point. The application effectively identifies and removes noise of different densities and distances, makes the algorithm have stronger robustness, enhances the denoising effect, and can retain more detail information.

Description

Technical Field

[0001] This invention discloses an adaptive joint filtering method based on the relationship between local features and spatial distance, belonging to the field of point cloud denoising technology. Background Technology

[0002] With the accelerated development of intelligent and connected new energy vehicles, autonomous driving technology has become the core of industrial competition. LiDAR, with its advantages of high precision and 3D environmental perception, has gradually become one of the important sensors in autonomous driving technology. It generates 3D point cloud data by scanning the surrounding environment, providing spatial information support for vehicles. However, in practical applications, point cloud acquisition is easily affected by various factors. Interference from severe weather such as rain and snow, and vehicle vibrations during driving, can lead to a large amount of outlier noise and drift noise in the point cloud, directly affecting the accuracy of environmental perception and the safety and rationality of subsequent environmental planning. How to effectively filter out point cloud noise in real driving environments is an important task in current intelligent driving. Existing point cloud denoising methods are often single-method denoising, which often struggles to balance efficiency and accuracy. Fixed-parameter filtering methods also cannot adapt to multi-scenario filtering. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive joint filtering method based on the relationship between local features and spatial distance, so as to solve the problems in the prior art where a single denoising method is difficult to balance efficiency and accuracy, and fixed parameter filtering methods cannot adapt to filtering in multiple scenarios.

[0004] An adaptive joint filtering method based on the relationship between local features and spatial distance includes using adaptive statistical filtering to filter out large-scale noise and then using adaptive bilateral filtering for noise suppression.

[0005] The adaptive bilateral filtering includes searching for neighboring points in the point cloud, calculating local curvature and local density, adjusting the spatial domain and attribute domain parameters of the bilateral filter based on the local curvature and local density, introducing local features into the weight parameters, normalizing and adjusting the weight parameters, and updating the geometric position of the points.

[0006] The search of the neighborhood points of the point cloud includes, using a fixed neighborhood method combined with KD-Tree, the search point cloud of the th Points The most recent There are neighborhood points, and the set of points within the neighborhood is . , In the formula, yes The Construct the covariance matrix of the neighborhood points. :

[0007] ;

[0008] ;

[0009] In the formula, express The mean, Indicates the first Neighboring points, express The 1 neighboring point.

[0010] Calculating local curvature includes, for Perform eigenvalue decomposition, three features The feature vectors corresponding to the three features are , , , It is the local plane normal vector The estimated value, local curvature The ratio of the smallest eigenvalue to the sum of eigenvalues:

[0011] .

[0012] Calculating local density includes calculating Euclidean distance to neighboring points :

[0013] ;

[0014] Local density The average Euclidean distance between points in the neighborhood The reciprocal:

[0015] .

[0016] Adjusting the spatial domain and attribute domain parameters of the bilateral filter includes adjusting the spatial domain parameters. With property domain parameters for:

[0017] ;

[0018] ;

[0019] In the formula, , These represent the initial parameters, , It is a regulating factor.

[0020] Incorporating local features into the weight parameters includes , The difference in normal vectors is incorporated into the weights of the bilateral filter, resulting in improved bilateral filter weights. for:

[0021] ;

[0022] In the formula, Point With neighboring points The difference in normal vectors.

[0023] Normalizing the weight parameters includes:

[0024] ;

[0025] In the formula, This represents the normalized weight value.

[0026] Geometric position of update point include:

[0027] .

[0028] Compared with the prior art, the present invention has the following advantages: it effectively identifies and removes noise of different densities and distances, making the algorithm more robust and enhancing the denoising effect while retaining more detailed information. Attached Figure Description

[0029] Figure 1 It is the original point cloud of the KITTI dataset;

[0030] Figure 2 This is the original point cloud of the KITTI dataset after adding noise;

[0031] Figure 3 This is a visualization of the statistical filtering results from the KITTI dataset;

[0032] Figure 4 This is a visualization of the radius filtering results from the KITTI dataset;

[0033] Figure 5 This is a visualization of the filtering results of this invention on the KITTI dataset;

[0034] Figure 6 It compares the accuracy metrics of various filtering algorithms in different scenarios;

[0035] Figure 7 It compares the recall metrics of various filtering algorithms in different scenarios;

[0036] Figure 8 It compares the origin retention rate of various filtering algorithms in different scenarios;

[0037] Figure 9 This is a visualization of the bias results from statistical filtering;

[0038] Figure 10 This is a visualization of the deviation results from the radius filtering;

[0039] Figure 11 This is a visualization result of the filtering deviation of this invention;

[0040] Figure 12 This is the original point cloud from the Boreas dataset;

[0041] Figure 13 This is the statistical filtering denoising effect on the Boreas dataset;

[0042] Figure 14 This is the radius filtering denoising effect on the Boreas dataset;

[0043] Figure 15 This is the filtering and denoising effect of the present invention on the Boreas dataset. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0045] An adaptive joint filtering method based on the relationship between local features and spatial distance includes using adaptive statistical filtering to filter out large-scale noise and then using adaptive bilateral filtering for noise suppression.

[0046] The adaptive bilateral filtering includes searching for neighboring points in the point cloud, calculating local curvature and local density, adjusting the spatial domain and attribute domain parameters of the bilateral filter based on the local curvature and local density, introducing local features into the weight parameters, normalizing and adjusting the weight parameters, and updating the geometric position of the points.

[0047] The search of the neighborhood points of the point cloud includes, using a fixed neighborhood method combined with KD-Tree, the search point cloud of the th Points The most recent There are neighborhood points, and the set of points within the neighborhood is . , In the formula, yes The Construct the covariance matrix of the neighborhood points. :

[0048] ;

[0049] ;

[0050] In the formula, express The mean, Indicates the first Neighboring points, express The 1 neighboring point.

[0051] Calculating local curvature includes, for Perform eigenvalue decomposition, three features The feature vectors corresponding to the three features are , , , It is the local plane normal vector The estimated value, local curvature The ratio of the smallest eigenvalue to the sum of eigenvalues:

[0052] .

[0053] Calculating local density includes calculating Euclidean distance to neighboring points :

[0054] ;

[0055] Local density The average Euclidean distance between points in the neighborhood The reciprocal:

[0056] .

[0057] Adjusting the spatial domain and attribute domain parameters of the bilateral filter includes adjusting the spatial domain parameters. With property domain parameters for:

[0058] ;

[0059] ;

[0060] In the formula, , These represent the initial parameters, , It is a regulating factor.

[0061] Incorporating local features into the weight parameters includes , The difference in normal vectors is incorporated into the weights of the bilateral filter, resulting in improved bilateral filter weights. for:

[0062] ;

[0063] In the formula, Point With neighboring points The difference in normal vectors.

[0064] Normalizing the weight parameters includes:

[0065] ;

[0066] In the formula, This represents the normalized weight value.

[0067] Geometric position of update point include:

[0068] .

[0069] This invention first performs adaptive statistical filtering, which calculates the average distance and standard deviation between all points to obtain a distance threshold. Then, it calculates the average distance between each point and its K nearest neighbors, comparing the distance threshold with the average distance of the K points. Points outside the threshold range are considered noise. The denoising effect of the algorithm is affected by the preset K value and the threshold standard deviation coefficient. An overly strict threshold leads to a large number of false positives, while an overly lenient threshold fails to effectively remove noise, especially in scenarios with uneven data density distribution, such as road traffic. Global statistics cannot adequately describe local conditions, resulting in the ineffective preservation of low-density data and insufficient denoising of high-density data. To address these shortcomings, a local density and distance compensation mechanism is introduced into the statistical filtering algorithm to adaptively adjust the algorithm threshold, mitigating the problem of false positives due to low density.

[0070] First, a KD-Tree point cloud index is constructed, followed by local density calculation. To achieve adaptive threshold adjustment based on density during denoising, the distribution of points within their local space needs to be obtained. Calculating local density quantifies the sparsity of the point cloud in that region and incorporates it into the threshold standard deviation coefficient of the statistical filter. This allows the threshold to adaptively increase or decrease based on the local density of the current point, thereby generating a differentiated filtering threshold for each point. This ensures that high-density point cloud noise is removed while effectively preserving low-density points. This invention employs a fixed-radius neighborhood method, statistically analyzing the neighborhood of each point within a specified radius. The number of neighboring points within, will At a fixed radius The number of points in a neighborhood is defined as the local density. :

[0071] ;

[0072] ;

[0073] Based on local density Calculate the global average density of the point cloud. :

[0074] ;

[0075] To better adapt the threshold standard deviation coefficient to local sampling conditions, based on local density Global average density Define the threshold standard deviation adaptive coefficient. :

[0076] ;

[0077] In the formula, Represents the basic threshold coefficient. This represents the density compensation index, used to adjust the compensation level. When... This indicates a low-density area, through The ratio result is amplified to increase the corresponding removal threshold, preventing the false deletion of low-density valid points. Conversely, when... , The ratio can be reduced to make the corresponding rejection conditions more stringent, thereby reducing the number of missed detections of high-density noise.

[0078] Based on the scanning characteristics of LiDAR (LiDAR system), where the point cloud density is high near the sensor and sparse far away, LiDAR can often scan a range of tens or even hundreds of meters in traffic scenarios. Nearby vehicles and road signs are densely covered by the point cloud, while distant objects are only partially covered. The number of points decreases sharply with increasing distance. A distance compensation mechanism is designed to improve the uneven distribution of point cloud data between near and far distances. A stricter filtering strategy is applied to the high-density areas near the sensor, while the filtering requirements are appropriately relaxed for the low-density areas at greater distances, in order to retain more effective information. To achieve this strategy, a compensation factor is introduced. To achieve dynamic adjustment of point clouds at different distances, the following definition is used:

[0079] ;

[0080] in, This represents the Euclidean distance from the point to the sensor origin. Indicates all the median of Indicates the adjustment factor. When hour This indicates that the point is far from the origin, the point cloud density in the region is low, and the filtering threshold needs to be appropriately increased; conversely, when... hour This indicates that the point is close to the origin, and the point cloud density in the region is high, requiring an appropriate reduction in the filtering threshold. The adaptive threshold formula, combining local density and distance compensation mechanisms, is designed as follows:

[0081] ;

[0082] The specific process of adaptive statistical filtering is as follows:

[0083] (1) Using KD-Tree to search for a fixed radius Inner nearest neighbor ;

[0084] (2) Calculate the query point With Neighboring points Between :

[0085] ;

[0086] (3) Calculate the mean of the local average distances of all N points in the point cloud. and standard deviation :

[0087] ;

[0088] ;

[0089] (4) Calculate the local density at each point With global average density Then calculate the points Distance to sensor origin And construct a compensation factor based on the distance to the median. Design adaptive filtering threshold ;

[0090] (5) Comparison With the filter threshold, when At that time, it was believed For a valid point, if If it is not within this range, then it is considered a point. Remove noise.

[0091] Even after filtering out large-scale noise, residual small-scale noise in the point cloud can still affect the fine structure of the data and the accuracy of subsequent processing. Bilateral filtering designs weight functions for two domains, considering both spatial and attribute weights. Points in the local neighborhood that are close to the target point in both location and attributes are given higher weights. Then, based on these weights, the noise points are moved along the normal vector direction, thus smoothing noise while preserving local geometric details. However, traditional bilateral filtering algorithms use fixed spatial and attribute scales, making them difficult to adapt to traffic road scenarios with large density variations, leading to unstable filtering results. By introducing local feature information such as local curvature, density, and normal vectors, bilateral filtering algorithms can adaptively adjust parameters, enabling more precise noise suppression and feature preservation.

[0092] The curvature index can reflect the flatness of the local surface of a point cloud. When the curvature index is large, it indicates that the point may be located in a region with sharp surface or obvious edge features. Conversely, it indicates that the point may be located in a region with smooth surface.

[0093] Density reflects the density of a local point cloud distribution. In the statistical filtering stage, a density calculation method based on the number of neighborhood points is used. To detect outliers. In the bilateral filtering stage, the distribution of points affects the filtering strength. Therefore, a density calculation method based on the reciprocal of the average distance of neighboring points is used to adaptively adjust the filtering parameters.

[0094] To verify the effectiveness of the algorithm of this invention, experiments were conducted using the KITTI and Boreas datasets. The KITTI dataset, collected from real traffic scenarios, is one of the evaluation datasets for autonomous driving and computer vision. The Boreas dataset is a multi-seasonal autonomous driving dataset, which collects data by driving the same fixed route at different times, capturing real driving scenarios under different seasons and adverse weather conditions, thus possessing rich environmental noise samples.

[0095] To evaluate the algorithm's effectiveness in denoising, this invention uses denoising accuracy. Recall rate Origin retention rate R o A quantitative evaluation and comparison were conducted, and the calculation formulas for each evaluation indicator are as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] TP represents the number of points that were correctly identified as noise and removed, FP represents the number of points that were actually not noise but were incorrectly removed, TN represents the number of points that were actually not noise and were correctly retained, and FN represents the number of points that were actually noise but were incorrectly retained.

[0100] To analyze the denoising effect of the algorithm in complex scenarios, Gaussian noise with a mean of 0m, a standard deviation of 0.2m, and a quantity ratio of 30% was added to a real intersection scene data from the KITTI dataset. Simultaneously, to better simulate rain and snow, 10% uniformly distributed random spatial noise was superimposed. Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 As shown, Gaussian noise mainly forms high-density drift noise along the normal vector direction of the effective point cloud, while random noise manifests as outlier noise uniformly distributed across the entire domain. Looking at the whole image, it can be seen that each algorithm has a good effect on removing uniformly distributed random spatial noise. However, observing the details in the image clearly shows that a large number of noise points remain near trees and vehicles at close to medium distances after statistical filtering and radius filtering. This is because statistical filtering and radius filtering rely on preset parameters. When facing uneven point cloud distribution, to ensure that enough points are retained in distant areas, the filtering parameters are often set relatively loosely, resulting in the inability to effectively remove noise from high-density point cloud areas at close range. In contrast, the algorithm of this invention considers local features and distance compensation, can adapt to point clouds of different densities, flexibly change the filtering threshold, and effectively filter out point clouds of different densities. Observing the rear of a nearby vehicle, it can be seen that the application of adaptive bilateral filtering can more effectively preserve vehicle edge features and also has a significant effect on smoothing noise.

[0101] To achieve an objective evaluation of algorithm performance, this invention uses denoising precision, recall, and origin retention rate for quantitative evaluation. Precision reflects the proportion of actual noise points among the removed points, recall measures the proportion of all noise points successfully removed, and origin retention rate measures the proportion of original points retained during the denoising process. Analysis of Table 1 shows that when the origin retention rate is around 90%, the precision and recall of this invention's algorithm are significantly higher than those of statistical filtering and radius filtering. Statistical filtering has the lowest precision, indicating its weak ability to identify and remove noise. Radial filtering has a much lower recall rate than the other two algorithms, indicating that it retains a large amount of noise without removing it in order to preserve valid points. In contrast, with a similar origin retention rate, this invention's algorithm improves precision by 11.08% and 10.48% compared to statistical filtering and radius filtering, respectively, while improving recall by 13.90% and 22.49%, respectively. This demonstrates that the algorithm can reduce noise interference while ensuring higher data integrity.

[0102] Table 1. Comparison of three metrics for various filtering algorithms;

[0103] ;

[0104] To verify the denoising effect of the algorithm in different scenarios, KITTI point cloud data from different scenarios were taken, and Gaussian noise with a mean of 0m, a standard deviation of 0.2m, and a quantity ratio of 30% was added. Similarly, 10% uniformly distributed random spatial noise was also superimposed. Point cloud denoising was performed without changing the algorithm parameters. The precision, recall, and origin retention rates of each algorithm were compared. Figure 6 , Figure 7 , Figure 8 As shown in the figure, the analysis reveals results similar to those in Table 1. When the original point retention rate is similar across different scenarios, the precision and recall of the algorithm presented in this invention are significantly higher than those of statistical filtering and radius filtering in all scenarios. Compared to statistical filtering, precision and recall are improved by approximately 5.8% and 15.9%, respectively; compared to radius filtering, they are improved by approximately 6.3% and 24.8%, respectively. This indicates that the algorithm can more accurately distinguish between valid and noise points, and exhibits higher robustness and adaptability in multiple scenarios.

[0105] To provide a more objective evaluation of the algorithm, point cloud data of a vehicle was selected from the aforementioned traffic scenario. Error analysis was performed between the extracted point cloud and the original point cloud. The results are as follows: Figure 9 , Figure 10 , Figure 11 As shown in Table 2, the statistical filtering and radius filtering algorithms have far more high-bias points than the algorithm of this invention, indicating that the algorithm of this invention can better filter out outliers, smooth noise to keep the bias within a small range, and improve the expressive power of edge features.

[0106] Analysis of Table 2 shows that the maximum distance deviation of the algorithm of this invention is reduced by 6.48% and 9.10% respectively compared with the two traditional filtering algorithms. It is observed that the maximum distance error is significantly higher than the average distance error, indicating that a very small number of points with high deviation are still not removed by the three algorithms. In terms of average distance error, the algorithm of this invention reduces it by 54.56% and 55.90% respectively compared with statistical filtering and radius filtering, indicating that the overall point cloud accuracy is significantly improved. In terms of standard deviation, the error dispersion is reduced by 35.35% and 37.06% respectively, indicating that the algorithm of this invention has a significant advantage in error stability.

[0107] Table 2 shows the model deviations of each filtering algorithm.

[0108] ;

[0109] To verify the robustness and denoising performance of the algorithm under actual rain and snow conditions, this invention also selected the Boreas dataset collected under real rain and snow conditions as experimental data. The results are as follows: Figure 12 , Figure 13 , Figure 14 , Figure 15 As shown in the figure, it is easy to see that all three filtering algorithms can filter out most of the rain and snow noise. However, for high-density noise point cloud regions, the filtering effect of the algorithm of this invention is significantly better than that of traditional statistical filtering and radius filtering. The drift noise near the vehicle is also significantly reduced, and the noise reduction effect is better.

[0110] This invention proposes an adaptive joint filtering point cloud denoising algorithm. By considering local density and spatial distance, the filtering threshold can be flexibly adjusted according to neighborhood density and spatial distance, effectively preserving distant sparse point clouds while suppressing nearby high-density noisy point clouds. Local features are introduced to improve the weight parameters of the bilateral filtering algorithm, smoothing noise while enhancing edge feature representation. Experimental results show that the algorithm of this invention has better denoising performance than traditional statistical filtering and radius filtering. In multi-traffic scenarios with uneven density, it can more effectively preserve local information, with precision and recall improving by an average of approximately 5.8% and 15.9% compared to statistical filtering, and by approximately 6.3% and 24.8% compared to radius filtering; the error dispersion is reduced by 35.35% and 37.06%, respectively. In complex noisy environments, the stable denoising performance of the algorithm of this invention provides a solid and reliable foundation for subsequent environmental perception and path planning.

[0111] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive joint filtering method based on local feature and spatial distance relationship, characterized in that, The adaptive bilateral filtering includes searching neighborhood points of the point cloud, calculating local curvatures and local densities, adjusting spatial domain and attribute domain parameters of the bilateral filtering based on the local curvatures and the local densities, introducing local features into a weight parameter, and normalizing and adjusting the weight parameter, and updating a geometric position of the point. The normalizing and adjusting of the weight parameter includes: The searching of the neighborhood points of the point cloud comprises searching the nearest neighborhood points of the first point in the point cloud by using a fixed neighborhood method combined with a KD-Tree, the neighborhood point set being , , , , , , , , , wherein the first neighborhood point of the first point is ； ； wherein denotes the mean of denotes the neighborhood point, denotes the neighborhood point; Computing the local curvature comprises, for each point of the set of points, performing an eigenvalue decomposition, three eigenvalues , and three eigenvectors corresponding to the three eigenvalues , , , is an estimate of the local plane normal , and the local curvature is the ratio of the smallest eigenvalue to the sum of the eigenvalues. ； Computing the local density comprises computing the Euclidean distance to each point in the neighborhood : ； local density the inverse of the average of the euclidean distances of the points in the neighborhood to the center point ； Adjusting the spatial domain and attribute domain parameters for bilateral filtering includes adjusting the spatial domain parameters and the attribute domain parameters are: ； ； wherein , respectively represent initial parameters, , is an adjustment factor; Introducing local features into the weight parameters includes introducing , , the normal vector difference into the weight of bilateral filtering, and the improved bilateral filtering weight is: ； In the formula, representing points and the normal vector difference of the neighborhood points ; ​ ； In the formula, denotes the normalized weight value; Geometric position of the update point comprising: 。

Images