Adaptive point cloud denoising method and system for severe weather

Abstract

The application discloses a self-adaptive point cloud denoising method for severe weather, which comprises the following steps: step 1, constructing a spatial index for input point cloud, obtaining the nearest neighbor distance of each point in the point cloud based on the spatial index, and determining the optimal neighborhood search number value; step 2, performing physical domain calibration processing on distance features and intensity features; step 3, based on the calibrated distance features and intensity features, constructing a fusion measurement factor to describe the structural consistency between a point and its neighborhood; in different weather scenes, different statistical models are used for threshold estimation to realize the adaptation to various severe weather conditions and complete the preliminary screening of noise points; step 4, based on the spatial connectivity and neighborhood structural consistency, performing secondary analysis on the boundary region points in the preliminary screening result; and step 5, outputting the point cloud data after final denoising, which is used for subsequent perception and decision tasks. The application realizes high-precision identification of severe weather noise points and effective reservation of real structure points.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional point cloud data processing technology, specifically relating to an adaptive point cloud denoising method and system for severe weather. Background Technology

[0002] As a core environmental perception sensor in autonomous driving and intelligent transportation systems, LiDAR can acquire high-precision three-dimensional spatial information. However, in adverse weather conditions such as rain, snow, and fog, a large number of suspended particles scatter and absorb the laser beam, resulting in abnormal echo signals and thus forming a large number of outlier noise points in the point cloud data.

[0003] These noisy points severely impact point cloud quality, thereby reducing the accuracy and stability of downstream tasks such as target detection, obstacle recognition, localization and mapping, and semantic segmentation. Therefore, effective denoising of point cloud data under adverse weather conditions is a crucial step in ensuring the safety and reliability of the perception system.

[0004] Existing point cloud denoising methods mainly include statistical rule-based methods and deep learning-based methods. Statistical rule-based methods typically require manually setting neighborhood scale or distance threshold parameters. When point cloud density changes or environmental conditions change, fixed parameters are often difficult to adapt, easily leading to the false deletion of real structural points or the missed detection of noise points. Deep learning-based methods, represented by Heinzler et al.'s "CNN-based Lidar Point Cloud De-Noising in Adverse Weather" (IEEE RA-L, 2020), proposed the WeatherNet network, which uses convolutional neural networks to perform semantic segmentation on point cloud projection maps to remove noise. Although this method performs well on specific datasets, its drawbacks are: firstly, model training relies on massive amounts of labeled data, and point-by-point labeling under adverse weather conditions is extremely costly; secondly, the model contains a large number of parameter operations and convolutional layers, resulting in high memory consumption and large inference latency, making it difficult to deploy in real time on computationally limited automotive embedded platforms.

[0005] Therefore, how to achieve adaptive, high-precision, and low-computational-cost denoising processing of point cloud data in severe weather without requiring manual parameter setting is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, the present invention aims to provide an adaptive point cloud denoising method and system for severe weather. By constructing a density-driven neighborhood scale adaptive estimation mechanism, a physical domain joint calibration mechanism for distance and intensity, a dynamic threshold discrimination mechanism based on fused statistics, and a cascaded refined recall mechanism, the method achieves high-precision identification of noise points in severe weather and effective preservation of real structure points.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An adaptive point cloud denoising method for severe weather includes the following steps; Step 1: Construct a spatial index and estimate the neighborhood scale: Step 1: Construct a spatial index for the input point cloud. Based on the spatial index, obtain the nearest neighbor distance for each point in the point cloud. Using the mean of the neighborhood distances as a reference, construct sparse neighborhood distance sets and dense neighborhood distance sets respectively. Further, obtain the local sparsity index of the point cloud. Through the ratio of the local sparsity index, the local density distribution characteristics of the point cloud can be obtained. Then, perform statistical analysis on the local density distribution characteristics of the point cloud, and use a density comparison mechanism to estimate a suitable neighborhood search scale to determine the optimal number of neighborhoods to search. value; The neighborhood scale is automatically determined by the statistical characteristics of the point cloud, without the need for manual preset of fixed parameters.

[0008] Step 2: Dual-domain physical calibration based on distance and intensity domains: Based on the number of optimal neighborhood searches For each point in the point cloud, its neighborhood distance feature and neighborhood reflection intensity feature are calculated and then calibrated in the physical domain. Considering the range attenuation of the lidar and the physical measurement characteristics, the distance feature and intensity feature are calibrated in the physical domain to compensate for the influence of measurement bias on the statistical results. Step 3: Dynamic adaptive threshold discrimination based on fused statistical metrics; Based on the calibrated distance and intensity features, a fusion metric is constructed to describe the structural consistency between a point and its neighborhood. Based on the global distribution characteristics of the fusion metric factors, different statistical models are used for threshold estimation under different weather scenarios to achieve adaptation to various severe weather conditions and complete the initial screening of noise points. The fusion metric factor simultaneously reflects spatial geometric distribution characteristics and physical reflection properties, thereby improving the reliability of noise discrimination.

[0009] Step 4: Cascaded Refined Recall Processing: Step 4: For the boundary region points in the initial screening results, a secondary analysis is performed based on spatial connectivity and neighborhood structure consistency; By using a cascaded refined recall mechanism, real structural points that may have been mistakenly deleted can be restored, thereby reducing the false deletion rate and improving the integrity of the point cloud. Step 5: Output the denoising results The final denoised point cloud data is output for subsequent perception and decision-making tasks.

[0010] In step one, a KD-tree spatial index is constructed, and the original point cloud is input. Extract the 3D coordinates and original reflection intensity of each point in the original point cloud. ; After extraction, the three-dimensional coordinates are used to calculate the nearest neighbor distance of each point in the point cloud, and the original reflection intensity is used for subsequent intensity calibration and fusion metric factor calculation.

[0011] The process of determining the optimal neighborhood search quantity based on the KD-tree spatial index in step one includes: Step 1: Process the original point cloud Nearest neighbor search to obtain the distance between the initial point pairs. ; Step 2: Set the distance between the initial point and the target point. Perform distance correction to obtain the corrected neighborhood distance. ,in , The distance from the point to the lidar. This represents taking the maximum value; Represents the normalization symbol; Step 3: Divide the global point set obtained through nearest neighbor search into sparse neighborhood distance sets: = Characterizing noise dispersion; and dense neighborhood distance sets: = , respectively characterize noise dispersion and target surface characteristics; calculate the local sparsity index of the point cloud. and : ; The average neighborhood scale of sparse points; This represents the average neighborhood scale of dense points. This represents taking the average value; Step 4: Based on the local sparsity index of the point cloud, perform density comparison, calculate the expected ratio of the sparse neighborhood distance set and the dense neighborhood distance set, and obtain the optimal number of neighborhoods to search. ; .

[0012] Step two specifically involves: Using the optimal neighborhood search number determined in step one The value is used to perform a neighborhood search to obtain the distance between each target point and its neighbors. The average neighborhood distance can be obtained using the following formula. ; Then, the average neighborhood distance of the target point is calculated separately. and original reflection intensity Perform calibration.

[0013] The distance calibration specifically involves addressing the inherent physical characteristics of lidar, where the distance between adjacent point clouds increases with distance. Consequently, the overall point cloud density exhibits a near-high density and far-low density characteristic. Distance calibration compensates for this inherent "near-high density, far-low density" physical characteristic of lidar, mapping neighborhood features at different distances to a unified statistical measurement space. The average neighborhood distance of the target point after calibration is... Represented as ; The intensity calibration specifically involves eliminating the interference of atmospheric extinction and distance energy attenuation on echo intensity, thereby enhancing the distinction between real targets and meteorological noise in the energy dimension. A global normalization strategy compensates for distance attenuation, enhancing the discriminative power of intensity characteristics without requiring manual selection of reference points; the calibrated single-point intensity... The expression is This indicates taking the maximum value. This represents the intensity of a single point obtained from the initial data acquisition. Then, by comparing the target point with its... The average strength of each neighboring point after calibration is taken to obtain a more stable average strength. : .

[0014] This represents the number of neighboring points of the current target point. Indicates the order of neighboring points.

[0015] Step three specifically involves: Differentiated treatments are implemented based on the physical scattering characteristics of different types of meteorological noise; Fog noise exhibits a small-scale, high-frequency diffuse distribution, while rain and snow noise is characterized by large size, spatial sparseness, and low-intensity outlier features. Constructing fusion metrics: For foggy weather: Because foggy noise echoes exhibit high-frequency, dense characteristics, their echo intensity is relatively high, making it difficult to distinguish between effective points. However, the distinction between effective points is more pronounced in terms of echo point distribution distance and neighborhood spatial characteristics. Therefore, an inverse distance weighting scheme is used to penalize points with large near-field neighborhood radii, resulting in a fusion metric expressed as: For rainy and snowy days: Unlike foggy days, besides the distance of echo point distribution and the spatial characteristics of neighboring points, the intensity distribution of noise points is generally much smaller than the average intensity of the neighborhood introduced by the effective points. Therefore, the low reflectivity and spatial sparseness of noise points can be used to expand the discrimination score, and its fusion metric factor is expressed as: The dynamic adjustment coefficient is generated by statistically averaging the fusion metric factors. The dynamic adjustment coefficient can then be obtained by using the ratio of each measurement factor to the statistical value. This allows for adaptive adjustment of the filtering boundary. Dynamic adjustment coefficient. The expression is as follows Adaptive threshold truncation: In foggy scenarios, an average value benchmark is selected; In rain and snow scenarios, the median is selected as the benchmark; Thresholds were obtained using the Mean-Std model in foggy weather. : .

[0016] For fusion metric factors standard deviation This is a dynamic adjustment coefficient. This represents the average value of the fusion metric.

[0017] Thresholds are obtained using the Median-MAD model in rainy or snowy weather. : From the fusion metric Find the middle ground. For about The absolute median difference; then for the final result The points can be partitioned into interior point sets. .

[0018] Step four specifically involves: Recalling mistakenly deleted sparse points; details are as follows: To calculate the neighborhood consistency score, we further calculate the score for each point using the set of inliers obtained in step three. The number of interior points in the nearest neighbor distribution is denoted as . ; Define boundary points as points that are initially filtered out but whose neighborhood contains interior points, and calculate their average connectivity benchmark score. It can be calculated using the following formula: The number of boundary points. This is based on the average value of the interior points in the boundary points. For each boundary point The number of nearest-neighbor interior points; Based on reliability gain factor Re-evaluate the deleted points and restore the target points with strong structural connectivity. The filtering expression is as follows: and It is about The mean and standard deviation are less than Those will be recalled as internal points.

[0019] An adaptive point cloud denoising system for severe weather includes a data preprocessing module, a neighborhood scale estimation module, a physical domain calibration module, a dynamic adaptive filtering module, a cascaded refined recall module, and a result output module. The data preprocessing module is used to construct a KD-tree spatial index and extract the 3D coordinates and reflection intensity of the point cloud; the extracted 3D coordinates and reflection intensity of the point cloud are input into the neighborhood scale estimation module; providing a foundation for subsequent searches.

[0020] The neighborhood scale estimation module automatically calculates the optimal neighborhood search quantity through a density comparison mechanism. Eliminate the randomness of manual parameter tuning; [This is incomplete and requires further context to translate accurately.] The value serves as the input parameter for the physical domain calibration module; The physical domain calibration module includes distance calibration and intensity calibration sub-modules, which are used to compensate for lidar measurement bias and unify statistical characteristics under different ranges; The dynamic adaptive filtering module performs initial screening of statistical features, selects the corresponding statistical model (Mean-Std or Median-MAD) based on the current weather category, realizes dynamic truncation of the discrimination threshold, and generates an interior point set and a noise point set. The cascaded refined recall module corrects the internal point set and the noise point set to form a closed-loop feedback; based on spatial connectivity, it performs secondary discrimination on the boundary points to restore the real structural points that were mistakenly deleted, thus ensuring the integrity of the point cloud. The output module outputs a denoised, clean point cloud, which is then connected to downstream sensing and mapping tasks.

[0021] The beneficial effects of this invention are: This invention constructs a density comparison mechanism based on KD-trees to autonomously estimate the number of optimal neighborhood search terms by utilizing the statistical differences in local density of point clouds. This mechanism enables the algorithm to automatically adjust the search scale according to the point cloud distribution density at different locations, solving the problem of "false deletion of distant points and missed detection of nearby points" caused by the traditional denoising method due to the manual setting of fixed neighborhood parameters.

[0022] By introducing a joint calibration mechanism in the range and intensity domains, the energy loss caused by density attenuation due to the spread of the lidar emission angle is compensated. This calibration scheme based on physical characteristics enables the fused statistics to more accurately characterize the difference between the target and the noise, significantly improving the feature discrimination capability under various complex weather conditions such as rain, snow, and fog.

[0023] Unlike filtering strategies that use a uniform threshold, this invention employs a "mean-standard deviation" model for the diffuse distribution of foggy weather and a "median-absolute deviation" model for the discrete characteristics of rainy and snowy weather, combined with dynamic adjustment coefficients for boundary truncation. This strategy allows the algorithm to autonomously adjust the filtering sensitivity according to the current weather category, efficiently filtering out dynamic meteorological noise while preserving as many low-reflectivity real structural points as possible, such as road boundaries and utility poles.

[0024] By introducing a cascaded refined recall mechanism, the boundary region is re-identified, which effectively reduces the erroneous deletion of real structural points and improves the integrity of the point cloud.

[0025] This method is based on statistical analysis and local neighborhood computation, does not rely on the deep neural network training process, has low computational resource requirements, and is suitable for vehicle-mounted edge computing platforms. Attached Figure Description

[0026] Figure 1 A schematic diagram of the adaptive point cloud denoising method.

[0027] Figure 2 This is a schematic diagram illustrating the estimation of the number of adaptive neighborhood point searches.

[0028] Figure 3 This is a schematic diagram illustrating the variation of neighborhood point cloud density with distance.

[0029] Figure 4 This is a distribution chart showing the filtering effect of different thresholds under different weather conditions.

[0030] Figure 5 A schematic diagram for filtering boundary points.

[0031] Figure 6 A schematic diagram of the actual filtering effect. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings.

[0033] An adaptive point cloud denoising method for severe weather conditions is proposed. This method employs four core steps: neighborhood scale adaptive estimation, dual-domain calibration, dynamic adaptive filtering based on spatial-intensity fusion metric, and cascaded refinement based on connectivity. Figure 1 As shown, it achieves parameter-free, efficient, accurate and robust point cloud denoising under severe weather conditions.

[0034] The specific implementation is as follows: Initialization and preprocessing: First, construct a KD-tree spatial index to accelerate neighborhood queries - input the original point cloud. Extract the 3D coordinates and original reflection intensity of each point in the point cloud. .

[0035] Step 1: Adaptive estimation of neighborhood scale: Aimed at solving the problem of neighborhood scale in traditional methods This addresses the issue of relying on manual experience for setting the parameters and the inability to dynamically adjust them according to point cloud density. By introducing a local density comparison mechanism, a suitable neighborhood search scale is automatically found for each point in the point cloud.

[0036] Step 1: Perform Nearest neighbor search to obtain the distance between the initial point pairs. .

[0037] Step 2: Set the distance between the initial point and the target point. Perform distance correction to obtain the corrected neighborhood distance. ,in , The distance from the point to the lidar. This represents taking the maximum value; This represents the normalization marker.

[0038] Step 3: Divide the global point set obtained through nearest neighbor search into a sparse set: = Characterizing noise dispersion; and dense clustering: = Characterize the surface properties of the target; calculate the local sparsity index of the point cloud. and : ; The average neighborhood scale of sparse points; This represents the average neighborhood scale of dense points. This represents taking the average value.

[0039] Step 4: Calculate the expected ratio of sparse and dense sets to determine the optimal neighborhood search quantity. ; .

[0040] The above steps are illustrated in the diagram. Figure 2 As shown; Step 2: Dual-domain calibration - Joint physical domain calibration mechanism for distance and intensity: Using the optimal neighborhood search number determined in step one The value is used to perform a neighborhood search to obtain the distance between each target point and its neighbors. The average neighborhood distance can be obtained using the following formula. ; Then, the average neighborhood distance of the target point is calculated separately. and original reflection intensity Perform calibration.

[0041] Distance calibration: Due to the inherent physical characteristics of lidar, the distance between adjacent point clouds increases with increasing distance. Therefore, the overall point cloud distribution density exhibits a characteristic of larger density near points and smaller density far points. Figure 3 As shown. Distance calibration can compensate for the inherent "closer proximity, more sparser distance" physical characteristic of lidar, mapping neighborhood features at different distances to a unified statistical measure space. The average neighborhood distance of the target point after calibration is... Represented as ; Intensity calibration: Intensity calibration can eliminate the interference of atmospheric extinction effect and distance energy attenuation on echo intensity, enhancing the distinction between real targets and meteorological noise in the energy dimension. It compensates for distance attenuation through a global normalization strategy, enhancing the discriminative power of intensity characteristics without requiring manual selection of reference points; its calibrated single-point intensity... The expression is This indicates taking the maximum value. This represents the intensity of a single point obtained from the initial data acquisition. Then, by comparing the target point with its... The average strength of each neighboring point after calibration is taken to obtain a more stable average strength. : ; Step 3: Weather Specificity Measurement and Dynamic Filtering Differentiated treatment is implemented based on the physical scattering characteristics of different types of meteorological noise. Fog noise typically exhibits a small-scale, high-frequency diffuse distribution, while rain and snow noise is characterized by large size, spatial sparseness, and low-intensity outlier features.

[0042] Constructing fusion metrics: By constructing targeted metric factors, the algorithm can more accurately capture noise characteristics under specific weather conditions. Compared with general denoising algorithms, this specific design can achieve higher classification accuracy with lower computational overhead, realizing deep adaptation to complex meteorological environments.

[0043] For foggy weather: Because foggy noise echoes exhibit high-frequency, dense characteristics, their echo intensity is relatively high, making it difficult to distinguish between effective points. However, the distinction between effective points is more pronounced in terms of echo point distribution distance and neighborhood spatial characteristics. Therefore, an inverse distance weighting scheme is used to penalize points with large near-field neighborhood radii, resulting in a metric factor expressed as: For rainy or snowy days: Unlike foggy days, besides the distance of echo point distribution and the spatial characteristics of neighboring points, the intensity distribution of noise points is generally much smaller than the average intensity of the neighborhood introduced by the effective points. Therefore, the low reflectivity and spatial sparseness of noise points can be comprehensively utilized to expand the discrimination score, and its metric factor is expressed as follows: The dynamic adjustment coefficient is generated by statistically averaging the fusion metric factors. The dynamic adjustment coefficient can then be obtained by using the ratio of each measurement factor to the statistical value. This allows for adaptive adjustment of the filtering boundary. Dynamic adjustment coefficient. The expression is as follows Adaptive threshold truncation: Depend on Figure 4 It can be seen that the filtering effect varies depending on the statistical model used for different weather conditions. Figure 4Images (a1)-(a3) show the noise and clean point segmentation results using the average value as the threshold, and images (b1)-(b3) show the noise and clean point segmentation results using the median value as the threshold. It can be seen that under the average value threshold, although the noise removal rate is slightly reduced, clean points are largely preserved. In foggy scenes, both the average value and the median value are very effective at removing noise, and the focus should be on preserving clean points; therefore, the average value threshold is more suitable for foggy scenes. In rainy or snowy scenes, the emphasis should be on noise removal, making the median value threshold more suitable.

[0044] Based on these statistical results, the following detailed framework is designed.

[0045] Thresholds were obtained using the Mean-Std (mean-standard deviation) model in foggy weather. : .

[0046] For fusion metric factors standard deviation This is a dynamic adjustment coefficient. This represents the average value of the fusion metric.

[0047] Thresholds are obtained using the Median-MAD (median-absolute median difference) model on rainy or snowy days. : From the fusion metric Find the middle ground. For about The absolute median difference. Then, ultimately for... The points can be partitioned into interior point sets. .

[0048] Step 4: Cascaded Fine-grained Recovery (CCR) The third-stage statistical filtering may unintentionally delete some sparse, valid points. To address this issue, a connectivity-based refinement module is introduced. This module utilizes neighborhood consistency to recall the mistakenly deleted sparse points. Its specific implementation is as follows: To calculate the neighborhood consistency score, we further calculate the score for each point using the set of inliers obtained in step three. The number of interior points in the nearest neighbor distribution is denoted as . .

[0049] Define boundary points as points that are initially filtered out but whose neighborhood contains interior points, and calculate their average connectivity benchmark score. It can be calculated by the following formula. The number of boundary points. This is based on the average value of the interior points in the boundary points. For each boundary point The number of interior points in the nearest neighbor distribution.

[0050] Based on reliability gain factor Re-evaluate the deleted points and restore the target points with strong structural connectivity. The filtering expression is as follows: and It is about The mean and standard deviation are less than Those will be recalled as internal points.

[0051] In the cascaded refinement process, the boundary point selection diagram is as follows: Figure 5 As shown, points 1-4 are the interior points identified in step three. Points 5-6 are the noise points filtered out in step three, but since these noise points are within the neighborhood of the interior points, they are considered as boundary points.

[0052] Finally, combining the results of steps three and four, the denoised complete point cloud is output.

[0053] This embodiment uses the internationally authoritative Winter Adversarial Dataset (WADS) as the validation benchmark. This dataset, published by Kurup et al. in "WADS: Winter Adversarial Dataset for Lidar Segmentation and Object Detection in the Snow" (IEEE ITSC, 2022), is one of the most comprehensive benchmarks for LiDAR perception in snowy weather currently available. A schematic diagram of the actual filtering effect is shown below. Figure 6 As shown.

[0054] Experimental results show that, without manual threshold adjustment, the present invention achieves an average noise recognition accuracy of 95.55% and an average noise removal rate of 96.62% on the WADS dataset. Especially when snowfall levels change dynamically, the present invention can achieve consistently stable high-precision noise reduction without requiring manual parameter resetting.

Claims

1. A self-adaptive point cloud denoising method for severe weather, characterized in that, Includes the following steps; Step 1: Construct a spatial index for the input point cloud. Based on the spatial index, obtain the nearest neighbor distance for each point in the point cloud. Using the mean of the neighborhood distances as a reference, construct sparse neighborhood distance sets and dense neighborhood distance sets respectively. Further, obtain the local sparsity index of the point cloud. Through the ratio of the local sparsity index, the local density distribution characteristics of the point cloud can be obtained. Then, perform statistical analysis on the local density distribution characteristics of the point cloud, and use a density comparison mechanism to estimate a suitable neighborhood search scale to determine the optimal number of neighborhoods to search. value; Step 2: Based on the number of optimal neighborhood searches For each point in the point cloud, calculate its neighborhood distance feature and neighborhood reflection intensity feature, and perform physical domain calibration. Step 3: Based on the calibrated distance and intensity features, construct a fusion metric to describe the structural consistency between a point and its neighborhood; Based on the global distribution characteristics of the fusion metric factors, different statistical models are used to estimate thresholds under different weather scenarios to achieve adaptation to various severe weather conditions and complete the initial screening of noise points. Step 4: For the boundary region points in the initial screening results, a secondary analysis is performed based on spatial connectivity and neighborhood structure consistency; Step 5: Output the final denoised point cloud data for subsequent perception and decision-making tasks.

2. The adaptive point cloud denoising method for severe weather as described in claim 1, characterized in that, In step one, a KD-tree spatial index is constructed, and the original point cloud is input. Extract the 3D coordinates and original reflection intensity of each point in the original point cloud. ; After extraction, the three-dimensional coordinates are used to calculate the nearest neighbor distance of each point in the point cloud, and the original reflection intensity is used for subsequent intensity calibration and fusion metric factor calculation.

3. The adaptive point cloud denoising method for severe weather as described in claim 2, characterized in that, The process of determining the optimal neighborhood search quantity based on the KD-tree spatial index in step one includes: Step 1: Process the original point cloud Nearest neighbor search to obtain the distance between the initial point pairs. ; Step 2: Set the distance between the initial point and the target point. Perform distance correction to obtain the corrected neighborhood distance. ,in , The distance from the point to the lidar. This represents taking the maximum value; Represents the normalization symbol; Step 3: Divide the global point set obtained through nearest neighbor search into sparse neighborhood distance sets: = Characterizing noise dispersion; and dense neighborhood distance sets: = , respectively characterize noise dispersion and target surface characteristics; calculate the local sparsity index of the point cloud. and : ； The average neighborhood scale of sparse points; This represents the average neighborhood scale of dense points. This represents taking the average value; Step 4: Based on the local sparsity index of the point cloud, perform density comparison, calculate the expected ratio of the sparse neighborhood distance set and the dense neighborhood distance set, and obtain the optimal number of neighborhoods to search. ; 。 4. The adaptive point cloud denoising method for severe weather as described in claim 3, characterized in that, Step two specifically involves: Using the optimal neighborhood search number determined in step one The value is used to perform a neighborhood search to obtain the distance between each target point and its neighbors. The average neighborhood distance can be obtained using the following formula. : ； Then, the average neighborhood distance of the target point is calculated separately. and original reflection intensity Perform calibration.

5. The adaptive point cloud denoising method for severe weather as described in claim 4, characterized in that, The distance calibration specifically involves mapping neighborhood features at different distances to a unified statistical measure space; the average neighborhood distance of the target point after calibration. Represented as: 。 6. The adaptive point cloud denoising method for severe weather as described in claim 5, characterized in that, The single-point strength after strength calibration The expression is: This indicates taking the maximum value. The intensity of a single point is obtained from the original data acquisition, and then the intensity of the target point is compared with its... The average intensity is obtained by taking the average intensity of each neighboring point after calibration. : ； This represents the number of neighboring points of the current target point. Indicates the order of neighboring points.

7. The adaptive point cloud denoising method for severe weather as described in claim 6, characterized in that, Step three specifically involves: Differentiated treatments are implemented based on the physical scattering characteristics of different types of meteorological noise; For foggy weather: An inverse distance weighting scheme is used to penalize points with large near-field neighborhood radii, resulting in a fusion metric factor expressed as: For rainy and snowy days: The discrimination score is increased by leveraging the low reflectivity and spatial sparseness of noise points, resulting in a fusion metric factor expressed as follows: The dynamic adjustment coefficient is generated by statistically averaging the fusion metric factors. Then, the dynamic adjustment coefficient is obtained by using the ratio of each fusion metric factor to its statistical value. This enables adaptive adjustment of the filtering boundary and dynamic adjustment of the coefficient. The expression is as follows Thresholds were obtained using the Mean-Std model in foggy weather. : ； For fusion metric factors standard deviation This is a dynamic adjustment coefficient. This represents the average of the fusion metric factors; Thresholds are obtained using the Median-MAD model in rainy or snowy weather. : From the fusion metric Find the middle ground. For about The absolute median difference; then for the final result The points are divided into interior point sets. .

8. The adaptive point cloud denoising method for severe weather as described in claim 7, characterized in that, Step four specifically involves: The sparse points that were mistakenly deleted were retrieved, as detailed below: To calculate the neighborhood consistency score, we further calculate the score for each point using the set of inliers obtained in step three. The number of interior points in the nearest neighbor distribution is denoted as . ; Define boundary points as points that are initially filtered out but whose neighborhood contains interior points, and calculate their average connectivity benchmark score. It can be calculated using the following formula: The number of boundary points. This is based on the average value of the interior points in the boundary points. For each boundary point The number of nearest-neighbor interior points; Based on reliability gain factor Re-evaluate the deleted points and restore the target points with strong structural connectivity. The filtering expression is as follows: and It is about The mean and standard deviation are less than Those will be recalled as internal points.

9. An adaptive point cloud denoising system for severe weather conditions, used to implement the method according to any one of claims 1-8, characterized in that, It includes a data preprocessing module, a neighborhood scale estimation module, a physical domain calibration module, a dynamic adaptive filtering module, a cascaded refined recall module, and a result output module; The data preprocessing module is used to construct a KD-tree spatial index and extract the 3D coordinates and reflection intensity of the point cloud; the extracted 3D coordinates and reflection intensity of the point cloud are input into the neighborhood scale estimation module; The neighborhood scale estimation module automatically calculates the optimal neighborhood search quantity through a density comparison mechanism. Eliminate the randomness of manual parameter tuning; determine The value serves as the input parameter for the physical domain calibration module; The physical domain calibration module includes distance calibration and intensity calibration sub-modules, which are used to compensate for lidar measurement bias and unify statistical characteristics under different ranges; The dynamic adaptive filtering module performs initial screening of statistical features, selects the corresponding statistical model based on the current weather category, realizes dynamic truncation of the discrimination threshold, and generates an interior point set and a noise point set. The cascaded refined recall module corrects the internal point set and the noise point set to form a closed-loop feedback; Based on spatial connectivity, a secondary discrimination is performed on boundary points to restore the real structural points that were mistakenly deleted, thus ensuring the integrity of the point cloud. The output module outputs a denoised, clean point cloud, which is then connected to downstream sensing and mapping tasks.

Images