Data filtering and noise reduction method based on laser point cloud

Abstract

The invention discloses a data filtering and noise reduction method based on laser point cloud. According to the method, the multi-scale features are intelligently screened through the random forest algorithm, and the system can automatically generate the optimal filtering rule according to the local geometric characteristics of the point cloud without depending on manual parameter presetting. In a complex urban scene, a fine structure of a high-rise vertical face can be accurately reserved through curvature features, and sparse noise of a vegetation area is rapidly eliminated through a density threshold value. The data-driven adaptive mechanism not only reduces the manual intervention cost, but also greatly improves the scene generalization ability of the filtering strategy, so that the filtering strategy can still keep stable output in a dynamic environment. The dynamic interference is accurately positioned by combining the consistency verification of the motion trails of the multi-frame point clouds and utilizing the time sequence correlation analysis, so that the data reliability in a complex dynamic scene is remarkably improved, and a solid foundation is provided for high-precision three-dimensional reconstruction and real-time perception.

Description

Technical Field

[0001] The present invention belongs to the technical field of data noise reduction, and specifically relates to a data filtering and noise reduction method based on laser point cloud. Background Art

[0002] LiDAR technology is widely used in fields such as autonomous driving and surveying and mapping. The point cloud data it generates is massive and contains significant noise. With the increasing adoption of high-precision LiDAR, point cloud data processing has become a critical factor limiting its effectiveness. Currently, point cloud filtering algorithms are primarily categorized into three main types: statistical, geometric, and deep learning-based. These methods exhibit significant trade-offs between real-time performance and accuracy. In recent years, the development of Level 4 autonomous driving technology has placed even higher demands on the real-time and accuracy of point cloud data processing.

[0003] However, traditional methods only filter out noise through local geometric features or fixed thresholds, and cannot effectively distinguish between instantaneous noise introduced by moving targets and static valid point clouds, resulting in increased false deletion and missed detection rates in dynamic scenarios; at the same time, the filtering parameters need to be manually preset and cannot be dynamically adjusted. It is difficult to balance real-time performance and accuracy in complex environments, which seriously restricts the reliable application of high-precision lidar data. Summary of the Invention

[0004] The purpose of the present invention is to provide a data filtering and denoising method based on laser point cloud in order to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: a data filtering and noise reduction method based on laser point cloud, the method comprising the following steps: S1: Collect the original laser point cloud data and normalize it to the spatial coordinate system to eliminate the error caused by the sensor posture deviation.

[0006] S2: A multi-scale pyramid structure is used to perform hierarchical feature analysis on the point cloud, extracting local density, curvature and normal change features at scales of 0.5m, 1.0m and 2.0m, and generating a multi-dimensional feature matrix.

[0007] S3: Based on the random forest algorithm, the importance of the feature matrix is ​​evaluated, the parameters with feature importance threshold higher than 0.7 are screened, and adaptive filtering rules are dynamically generated.

[0008] S4: Execute coarse filtering processing to quickly remove outlier noise points at a retention rate of 60% to 80% according to the adaptive rule, and retain potential valid point clouds.

[0009] S5: Combined with the motion trajectory of continuous frame point cloud data, the displacement consistency of point clouds between adjacent frames is analyzed through the dynamic time warping (DTW) algorithm, and inconsistent points are marked as dynamic noise and eliminated twice.

[0010] S6: Perform fine filtering processing, build a Gaussian mixture model based on local geometric features, and finely filter out residual noise with a retention rate of 85% to 95%.

[0011] S7: Perform topological connectivity detection on the filtered point cloud and use the region growing algorithm to complete the key structural points lost due to filtering.

[0012] S8: Output the denoised point cloud data and feed back the spatiotemporal consistency verification results to the adaptive filtering parameter generation module in real time to optimize the data processing rules for the next frame.

[0013] S9: Generate a visual filtering report, record noise distribution, parameter adjustment records and processing efficiency indicators, and support subsequent algorithm iterative optimization.

[0014] In a preferred embodiment, in step S1, the system uses an iterative nearest point algorithm to align the point cloud collected from multiple perspectives to the global coordinate system based on the sensor calibration parameters. The specific process is to extract corner points or edge points in the point cloud as key feature points, iteratively optimize the rotation matrix and translation vector by minimizing the Euclidean distance between pairs of feature points, and eliminate the coordinate offset caused by the deviation of the lidar installation posture or the vibration of the motion platform. In the autonomous driving scenario, vehicle bumps will cause the point cloud to jitter. The algorithm realizes dynamic compensation through multi-frame registration to ensure the spatial consistency of subsequent feature analysis. This step solves the problem of feature extraction distortion caused by misaligned coordinate systems in traditional methods, and provides high-precision input for multi-scale analysis.

[0015] In a preferred embodiment, in step S2, the system constructs a three-scale pyramid with scales of 0.5 meters, 1.0 meters and 2.0 meters respectively. In each layer, the data density is simplified by voxel downsampling, and the neighborhood search is accelerated based on the spatial index structure. Taking the 0.5-meter fine layer as an example, the algorithm calculates the local density of each point, which is defined as the number of points within a sphere with a radius of 0.3 meters; at the same time, the curvature is calculated, and the degree of surface curvature is characterized by the minimum eigenvalue of the covariance matrix; and the normal change rate is measured by the angle between adjacent normal vectors. The coarse-scale layer focuses on the statistics of large-scale point cloud distribution patterns, such as detecting the boundary area between the ground and obstacles through density gradients. After weighted fusion, the multi-scale feature matrix comprehensively characterizes the local details and global structural characteristics of the point cloud.

[0016] In a preferred embodiment, in step S3, the system inputs the multi-scale feature matrix into an integrated learning model composed of multiple decision trees, and each tree selects the optimal feature splitting node based on the Gini coefficient. After training, the algorithm outputs a feature importance ranking, for example, the curvature weight accounts for 35% and the density accounts for 28%, and filters parameters with importance higher than the preset threshold to construct filtering rules. For areas with high curvature importance, such as vehicle edges, the system automatically lowers the curvature filtering threshold to retain sharp edges; for flat areas such as the ground, the density threshold is increased to quickly eliminate sparse noise. This step replaces manual parameter adjustment with data-driven, significantly improving the scenario adaptability of the filtering strategy.

[0017] In a preferred embodiment, in step S4, the system searches for neighboring points within a sphere with a radius of 0.5 meters with each point as the center, and calculates the distance mean and standard deviation. If the average distance between a point and its neighboring points exceeds the mean plus two times the standard deviation, it is determined to be an outlier and removed. The coarse filter retention rate is dynamically controlled at 60% to 80%. For example, it is reduced to 65% in vegetation-covered areas to efficiently remove branch and leaf noise, and increased to 75% in regular scenes such as building facades to avoid excessive deletion. This algorithm quickly distinguishes significant noise through local statistical characteristics, retaining a high-quality candidate point set for subsequent processing.

[0018] In a preferred embodiment, in step S5, dynamic noise is secondary eliminated through a spatiotemporal consistency verification mechanism, and the specific process includes: S5-1. Acquire multiple frames of laser point cloud data and extract the spatial coordinates of key feature points in each frame. Next, S5-2. Perform alignment and matching, calculate the cumulative difference in their displacement paths, and determine whether there is inconsistent displacement based on a preset dynamic threshold. Points whose displacement direction or amplitude in adjacent frames significantly deviates from the overall motion trend (such as moving objects such as pedestrians and vehicles) are marked as dynamic noise points. S5-3. Finally, secondary filtering is used to remove these inconsistent points while retaining the valid point cloud for static scenes. This mechanism effectively distinguishes dynamic noise from static objects by fusing temporal information with spatial displacement correlations. This overcomes the drawbacks of traditional single-frame filtering methods, which are prone to mistakenly deleting valid points or missing transient noise in dynamic scenes.

[0019] The calculation formula for correcting the calculation error by weighted summing the spatial displacement of the matching feature points in two consecutive frames and combining the time factor is: Dpath represents the cumulative difference between the point cloud matching paths of adjacent frames, which is used to measure the motion consistency between two frames. The smaller the value, the more consistent the displacement.

[0020] pit represents the three-dimensional coordinate vector of the i-th feature point in the t-th frame.

[0021] Indicates the t+1 frame through the DTW path and The coordinate vector of the matched feature points, j(i) represents the index mapping relationship after dynamic alignment.

[0022] w(Δt) represents the time decay weight function, Δt is the time interval between frames, and is used to reduce the probability of misjudgment caused by uneven sensor sampling rate or sudden change in motion speed.

[0023] In a preferred embodiment, in step S6, the point cloud is divided into multiple Gaussian-distributed clusters. The number of clusters is automatically determined using the Bayesian Information Criterion, and the cluster mean, covariance matrix, and weight coefficients are iteratively optimized using the expectation-maximization algorithm. The fine filtering stage calculates the posterior probability of each point belonging to each cluster. If the maximum probability is less than 0.9, the point is marked as noise. For gravel on the road surface, which has similar curvature to the ground but an abnormal density, the model separates and removes it by using covariance matrix differences, while retaining valid points in smooth areas of the road surface. This step replaces hard thresholding with a probabilistic model, effectively resolving the problem of overlapping spatial distributions of noise and valid points.

[0024] In a preferred embodiment, in step S7, curvature extreme points or density core points are used as seed points, and adjacent points that meet curvature continuity and density similarity are gradually merged. Curvature continuity is defined as the angle between the normals of adjacent points being less than 15 degrees, and density similarity requires a difference of less than 20%. For wall edges that are broken due to filtering, seed points are grown from the corners and extended along the normal direction to the break, automatically completing the missing points based on curvature consistency. This process not only restores the geometric integrity of the point cloud but also prevents the inadvertent repair of isolated noise points through connectivity constraints.

[0025] In a preferred embodiment, in step S8, the system analyzes the spatiotemporal consistency verification results in real time, such as the dynamic noise ratio and false negative rate, and updates the adaptive rules using an exponentially weighted moving average algorithm. If the curvature false negative rate in a region exceeds a threshold across multiple consecutive frames, the curvature filtering threshold is automatically lowered; conversely, if the noise miss rate increases, the density threshold weight is increased. Feedback data is stored in a structured format and strictly synchronized with the point cloud frame sequence to ensure the timeliness and traceability of parameter adjustments. This step marks the transition of the algorithm from static rules to dynamic adaptation.

[0026] In a preferred embodiment, in step S9, a point cloud noise distribution heat map is created using graphics rendering technology, with high-frequency noise areas (such as moving object trajectories) marked in red, and static valid points (such as blue). A timing analysis tool is also used to generate a line chart showing the inter-frame variation trends of indicators such as filter retention rate and processing latency. The report supports interactive operations, such as focusing on a specific area to view a comparison of local filtering effects. This step not only provides a quantitative assessment of algorithm performance but also provides an intuitive basis for scene understanding and model iteration.

[0027] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, a random forest algorithm is used to intelligently filter multi-scale features. The system automatically generates optimal filtering rules based on the local geometric properties of the point cloud, eliminating the need for manually preset parameters. In complex urban scenes, the fine structure of high-rise facades can be accurately preserved using curvature features, while sparse noise in vegetation areas can be quickly eliminated using density thresholds. This data-driven adaptive mechanism not only reduces the cost of manual intervention but also significantly improves the filtering strategy's ability to generalize across scenarios, enabling it to maintain stable output even in dynamic environments.

[0028] 2. In the present invention, efficient recognition and suppression of dynamic noise are achieved by integrating spatiotemporal correlation. Traditional methods are limited to single-frame data analysis, making it difficult to distinguish between instantaneous noise introduced by moving targets and static valid point clouds. This method combines the motion trajectory consistency verification of multi-frame point clouds and uses temporal correlation analysis to accurately locate dynamic interference. In autonomous driving scenarios, the system can track target displacement between consecutive frames in real time, eliminate abnormal points caused by vehicle bumps or obstacle movement, and ensure the integrity of static structures such as road boundaries and traffic signs. This spatiotemporal collaborative filtering strategy significantly improves data reliability in complex dynamic scenes, providing a solid foundation for high-precision three-dimensional reconstruction and real-time perception. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the process principle of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0031] Reference Figure 1 , A data filtering and noise reduction method based on laser point cloud, the method comprising the following steps: S1: Collect the original laser point cloud data and normalize it to the spatial coordinate system to eliminate the error caused by the sensor posture deviation.

[0032] S2: A multi-scale pyramid structure is used to perform hierarchical feature analysis on the point cloud, extracting local density, curvature and normal change features at scales of 0.5m, 1.0m and 2.0m, and generating a multi-dimensional feature matrix.

[0033] S3: Based on the random forest algorithm, the importance of the feature matrix is ​​evaluated, the parameters with feature importance threshold higher than 0.7 are screened, and adaptive filtering rules are dynamically generated.

[0034] S4: Execute coarse filtering processing to quickly remove outlier noise points at a retention rate of 60% to 80% according to the adaptive rule, and retain potential valid point clouds.

[0035] S5: Combined with the motion trajectory of continuous frame point cloud data, the displacement consistency of point clouds between adjacent frames is analyzed through the dynamic time warping (DTW) algorithm, and inconsistent points are marked as dynamic noise and eliminated twice.

[0036] S6: Perform fine filtering processing, build a Gaussian mixture model based on local geometric features, and finely filter out residual noise with a retention rate of 85% to 95%.

[0037] S7: Perform topological connectivity detection on the filtered point cloud and use the region growing algorithm to complete the key structural points lost due to filtering.

[0038] S8: Output the denoised point cloud data and feed back the spatiotemporal consistency verification results to the adaptive filtering parameter generation module in real time to optimize the data processing rules for the next frame.

[0039] S9: Generate a visual filtering report, record noise distribution, parameter adjustment records and processing efficiency indicators, and support subsequent algorithm iterative optimization.

[0040] In step S1, the system uses an iterative closest point algorithm based on sensor calibration parameters to align the point cloud collected from multiple perspectives to the global coordinate system. The specific process is to extract corner points or edge points in the point cloud as key feature points, iteratively optimize the rotation matrix and translation vector by minimizing the Euclidean distance between pairs of feature points, and eliminate coordinate offsets caused by deviations in the LiDAR installation posture or vibrations of the motion platform. In autonomous driving scenarios, vehicle bumps can cause point cloud jitter. The algorithm achieves dynamic compensation through multi-frame registration to ensure spatial consistency in subsequent feature analysis. This step solves the problem of feature extraction distortion caused by misaligned coordinate systems in traditional methods, providing high-precision input for multi-scale analysis.

[0041] In step S2, the system constructs a three-level scale pyramid with scales of 0.5 meters, 1.0 meters, and 2.0 meters respectively. In each layer, the data density is simplified by voxel downsampling, and the neighborhood search is accelerated based on the spatial index structure. Taking the 0.5-meter fine layer as an example, the algorithm calculates the local density of each point, defined as the number of points within a sphere with a radius of 0.3 meters; at the same time, it calculates the curvature, which characterizes the degree of surface curvature through the minimum eigenvalue of the covariance matrix; and the normal change rate, which measures the geometric mutation through the angle between adjacent normal vectors. The coarse-scale layer focuses on statistically analyzing the distribution pattern of large-scale point clouds, such as detecting the boundary area between the ground and obstacles through density gradients. After weighted fusion, the multi-scale feature matrix comprehensively characterizes the local details and global structural characteristics of the point cloud.

[0042] In step S3, the system inputs the multi-scale feature matrix into an ensemble learning model composed of multiple decision trees. Each tree selects the optimal feature splitting node based on the Gini coefficient. After training, the algorithm outputs a feature importance ranking, for example, with curvature weighting accounting for 35% and density accounting for 28%. Parameters with importance above a preset threshold are selected to construct filtering rules. For areas with high curvature importance, such as vehicle edges, the system automatically lowers the curvature filtering threshold to preserve sharp edges. For flat areas, such as the ground, the density threshold is increased to quickly remove sparse noise. This step replaces manual parameter adjustment with data-driven optimization, significantly improving the scenario adaptability of the filtering strategy.

[0043] In step S4, the system searches for neighboring points within a sphere with a radius of 0.5 meters, centered at each point, and calculates the mean and standard deviation of the distance. If the average distance between a point and its neighbors exceeds the mean plus two standard deviations, it is identified as an outlier and removed. The coarse filter retention rate is dynamically controlled between 60% and 80%. For example, it is reduced to 65% in vegetation-covered areas to effectively remove branch and foliage noise, while it is increased to 75% in regular scenes such as building facades to avoid excessive deletion. This algorithm quickly distinguishes significant noise through local statistical characteristics, retaining a high-quality candidate point set for subsequent processing.

[0044] In step S5, dynamic noise is secondary eliminated through a spatiotemporal consistency verification mechanism. The specific process includes: S5-1. Acquiring multiple frames of laser point cloud data and extracting the spatial coordinates of key feature points in each frame; secondly, S5-2. Performing alignment and matching, calculating the cumulative difference of their displacement paths, and determining whether there is inconsistent displacement based on a preset dynamic threshold; for points in adjacent frames whose displacement direction or amplitude significantly deviates from the overall motion trend (such as moving objects such as pedestrians and vehicles), these points are marked as dynamic noise points; S5-3. Finally, secondary filtering is used to remove these inconsistent points while retaining the valid point cloud for static scenes. This mechanism effectively distinguishes dynamic noise from static objects by fusing temporal information with spatial displacement correlations. This overcomes the drawbacks of traditional single-frame filtering methods, which are prone to mistakenly deleting valid points or missing transient noise in dynamic scenes.

[0045] The calculation formula for correcting the calculation error by weighted summing the spatial displacement of the matching feature points in two consecutive frames and combining the time factor is: Dpath represents the cumulative difference between the point cloud matching paths of adjacent frames, which is used to measure the motion consistency between two frames. The smaller the value, the more consistent the displacement.

[0046] pit represents the three-dimensional coordinate vector of the i-th feature point in the t-th frame.

[0047] Indicates the t+1 frame through the DTW path and The coordinate vector of the matched feature points, j(i) represents the index mapping relationship after dynamic alignment.

[0048] w(Δt) represents the time decay weight function, Δt is the time interval between frames, and is used to reduce the probability of misjudgment caused by uneven sensor sampling rate or sudden change in motion speed.

[0049] In step S6, the point cloud is divided into multiple Gaussian-distributed clusters. The number of clusters is automatically determined using the Bayesian Information Criterion. The cluster mean, covariance matrix, and weight coefficients are iteratively optimized using the expectation-maximization algorithm. The fine filtering stage calculates the posterior probability of each point belonging to each cluster. Points with a maximum probability less than 0.9 are marked as noise. Gravel on the road surface, which has similar curvature to the ground but an abnormal density, is separated and removed by the model using covariance matrix differences, while retaining valid points in smooth areas. This step replaces hard thresholding with a probabilistic model, effectively resolving the problem of overlapping spatial distributions of noise and valid points.

[0050] In step S7, using curvature extreme points or density core points as seed points, neighboring points that meet curvature continuity and density similarity are gradually merged. Curvature continuity is defined as the angle between the normals of adjacent points being less than 15 degrees, while density similarity requires a difference of less than 20%. For wall edges that are broken due to filtering, seed points are grown from the corners and extended along the normal direction to the break. Missing points are automatically completed based on curvature consistency. This process not only restores the geometric integrity of the point cloud but also prevents the inadvertent repair of isolated noise points through connectivity constraints.

[0051] In step S8, the system analyzes the spatiotemporal consistency verification results in real time, such as the dynamic noise ratio and false negative rate, and updates the adaptive rules using an exponentially weighted moving average algorithm. If the curvature false negative rate in a certain area exceeds the threshold across multiple consecutive frames, the curvature filtering threshold is automatically lowered. Conversely, if the noise miss rate increases, the density threshold weight is increased. Feedback data is stored in a structured format and strictly synchronized with the point cloud frame sequence to ensure the timeliness and traceability of parameter adjustments. This step marks the transition of the algorithm from static rules to dynamic adaptation.

[0052] In step S9, a heat map of the point cloud noise distribution is created using graphics rendering technology. High-frequency noise areas, such as moving object trajectories, are highlighted in red, while static, valid points are indicated in blue. A time series analysis tool is also used to generate a line chart showing the inter-frame trends of metrics such as filter retention and processing latency. The report supports interactive operations, such as focusing on specific areas to compare local filtering effects. This step not only provides a quantitative assessment of algorithm performance but also provides an intuitive basis for scene understanding and model iteration.

[0053] From the above we can know: In this invention, a random forest algorithm is used to intelligently filter multi-scale features. The system automatically generates optimal filtering rules based on the local geometric properties of the point cloud, eliminating the need for manually preset parameters. In complex urban scenes, the fine structure of high-rise facades can be accurately preserved using curvature features, while sparse noise in vegetation areas can be quickly removed using density thresholds. This data-driven adaptive mechanism not only reduces the cost of manual intervention but also significantly improves the filtering strategy's ability to generalize across scenarios, enabling it to maintain stable output even in dynamic environments.

[0054] In the present invention, efficient recognition and suppression of dynamic noise are achieved by integrating spatiotemporal correlation. Traditional methods are limited to single-frame data analysis and it is difficult to distinguish between instantaneous noise introduced by moving targets and static valid point clouds. This method combines the motion trajectory consistency verification of multi-frame point clouds and uses temporal correlation analysis to accurately locate dynamic interference. In autonomous driving scenarios, the system can track target displacement between consecutive frames in real time, eliminate abnormal points caused by vehicle bumps or obstacle movement, and ensure the integrity of static structures such as road boundaries and traffic signs. This spatiotemporal collaborative filtering strategy significantly improves data reliability in complex dynamic scenes, providing a solid foundation for high-precision three-dimensional reconstruction and real-time perception.

[0055] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A data filtering and denoising method based on laser point cloud, characterized by: The method comprises the following steps: S1: Collect the original laser point cloud data and normalize it to the spatial coordinate system to eliminate the error caused by the sensor posture deviation; S2: A multi-scale pyramid structure is used to perform hierarchical feature analysis on the point cloud, extracting local density, curvature, and normal change features at scales of 0.5m, 1.0m, and 2.0m, and generating a multi-dimensional feature matrix. S3: Based on the random forest algorithm, the importance of the feature matrix is ​​evaluated, the parameters with feature importance threshold higher than 0.7 are screened, and adaptive filtering rules are dynamically generated; S4: Execute coarse filtering to quickly remove outlier noise points at a retention rate of 60% to 80% according to the adaptive rule, and retain the potential valid point cloud; S5: Combine the motion trajectory of continuous frame point cloud data and analyze the displacement consistency of point clouds between adjacent frames through the dynamic time warping (DTW) algorithm. Mark inconsistent points as dynamic noise and remove them twice. S6: Perform fine filtering, build a Gaussian mixture model based on local geometric features, and finely filter out residual noise with a retention rate of 85% to 95%; S7: Perform topological connectivity detection on the filtered point cloud and use the region growing algorithm to complete the key structural points lost due to filtering; S8: Output the denoised point cloud data and feed back the spatiotemporal consistency verification results to the adaptive filtering parameter generation module in real time to optimize the data processing rules for the next frame; S9: Generate a visual filtering report, record noise distribution, parameter adjustment records and processing efficiency indicators, and support subsequent algorithm iterative optimization.

2. The method for filtering and denoising data based on laser point cloud according to claim 1, characterized in that: In step S1, the system aligns the point cloud collected from multiple perspectives to the global coordinate system using an iterative closest point algorithm based on sensor calibration parameters. The specific process is to extract corner points or edge points in the point cloud as key feature points, iteratively optimize the rotation matrix and translation vector by minimizing the Euclidean distance between pairs of feature points, and eliminate coordinate offsets caused by deviations in the laser radar installation posture or vibrations of the motion platform.

3. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S2, the system constructs a three-level scale pyramid with scales of 0.5 meters, 1.0 meters, and 2.0 meters, respectively. In each layer, the data density is simplified by voxel downsampling, and the neighborhood search is accelerated based on the spatial index structure. Taking the 0.5-meter fine layer as an example, the algorithm calculates the local density of each point, which is defined as the number of points within a sphere with a radius of 0.3 meters. At the same time, the curvature is calculated, and the degree of surface curvature is characterized by the minimum eigenvalue of the covariance matrix; as well as the normal change rate, which measures the geometric mutation by the angle between adjacent normal vectors.

4. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S3, the system inputs the multi-scale feature matrix into an integrated learning model composed of multiple decision trees, and each tree selects the optimal feature splitting node through the Gini coefficient; After training, the algorithm outputs a feature importance ranking, for example, curvature weight accounts for 35% and density accounts for 28%. Parameters with importance above a preset threshold are selected to construct filtering rules. For areas with high curvature importance, such as vehicle edges, the system automatically lowers the curvature filtering threshold to preserve sharp edges. For flat areas such as the ground, the density threshold is increased to quickly remove sparse noise.

5. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S4, the system searches for neighboring points within a sphere with a radius of 0.5 meters with each point as the center, and calculates the distance mean and standard deviation; if the average distance between a point and its neighboring points exceeds the mean plus two standard deviations, it is determined to be an outlier and is removed.

6. The method for filtering and denoising data based on laser point cloud according to claim 1, characterized in that: In step S5, dynamic noise is removed twice through a spatiotemporal consistency verification mechanism. The specific process includes: S5-1. Obtain continuous multi-frame laser point cloud data and extract the spatial coordinates of key feature points in each frame; secondly, S5-2. Perform alignment matching, calculate the cumulative difference of its displacement path, and determine whether there is inconsistent displacement based on a preset dynamic threshold; for points in adjacent frames where the displacement direction or amplitude significantly deviates from the overall motion trend, mark them as dynamic noise points; S5-3. Finally, these inconsistent points are removed through secondary filtering while retaining the valid point cloud of the static scene; The calculation formula for correcting the calculation error by weighted summing the spatial displacement of the matching feature points in two consecutive frames and combining the time factor is: Dpath represents the cumulative difference between the point cloud matching paths of adjacent frames, which is used to measure the motion consistency between two frames. The smaller the value, the more consistent the displacement. pit represents the three-dimensional coordinate vector of the i-th feature point in the t-th frame; Indicates the t+1 frame through the DTW path and The coordinate vector of the matched feature points, j(i) represents the index mapping relationship after dynamic alignment; w(Δt) represents the time decay weight function, Δt is the time interval between frames, and is used to reduce the probability of misjudgment caused by uneven sensor sampling rate or sudden change in motion speed.

7. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S6, the point cloud is divided into multiple Gaussian distribution clusters, the number of clusters is automatically determined by the Bayesian Information Criterion, and the cluster mean, covariance matrix and weight coefficient are iteratively optimized by the expectation maximization algorithm; The fine filtering stage calculates the posterior probability of each point belonging to each cluster. If the maximum probability is less than 0.9, it is marked as noise. For gravel on the road, its curvature is close to that of the ground but its density is abnormal. The model separates and eliminates it through the difference in the covariance matrix, while retaining valid points in the smooth area of ​​the road.

8. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S7, the curvature extreme point or density core point is used as the seed point, and the adjacent points that meet the curvature continuity and density similarity are gradually merged; curvature continuity is defined as the angle between the normals of adjacent points is less than 15 degrees, and the density similarity requires the difference to be less than 20%; for the wall edge that is broken due to filtering, the seed point starts to grow from the corner of the wall and extends along the normal direction to the break, and the missing points are automatically filled in by judging the curvature consistency.

9. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S8, the system analyzes the spatiotemporal consistency verification results in real time, such as the dynamic noise ratio and the false rejection rate, and updates the adaptive rules through the exponentially weighted moving average algorithm; if the curvature false rejection rate in a certain area exceeds the threshold in multiple consecutive frames, the curvature filtering threshold is automatically lowered; conversely, if the noise missed detection rate increases, the density threshold weight is increased.

10. The method for filtering and denoising data based on laser point cloud according to claim 1, wherein: In step S9, a point cloud noise distribution heat map is drawn using graphics rendering technology, with high-frequency noise areas such as moving target trajectories marked in red and static valid points in blue; and a line graph is generated using a timing analysis tool.

Images