Dynamic point cloud filtering method based on comprehensive confidence coefficient and reference parameter adaptive adjustment

Abstract

The invention provides a dynamic point cloud filtering method based on comprehensive confidence and reference parameter adaptive adjustment. The dynamic point cloud filtering method comprises the following implementation steps: calculating a complexity score and a time sequence factor of three-dimensional point cloud data of each frame; performing adaptive adjustment on the reference parameter data set based on the complexity score and the time sequence factor; and obtaining a dynamic point cloud filtering result. According to the method, the dynamic and static attributes of the object are judged through semantic, time sequence and geometric information in the comprehensive confidence of each piece of three-dimensional point cloud data, so that the defect that a dynamic object of which the moving speed is lower than a threshold value is easily misjudged as a static object is overcome, and the probability of filtering the dynamic point leakage is reduced; and the reference parameter data set is adaptively adjusted based on the complexity score and the time sequence factor, so that the reference parameters are adaptively adjusted according to the complexity of the scene in the complex scene with changeable dynamic object density and motion modes, the method has relatively high adaptability to the complex environment, and the filtering precision and robustness are improved.

Description

A dynamic point cloud filtering method based on adaptive adjustment of comprehensive confidence and baseline parameters. Technical Field

[0001] This invention belongs to the field of robotics technology and relates to a dynamic point cloud filtering method based on adaptive adjustment of comprehensive confidence and benchmark parameters, which can be applied to fields such as autonomous driving, robot navigation, and environmental perception. Background Technology

[0002] Mapping and localization in fields such as autonomous driving, robot navigation, and environmental perception rely on point cloud data provided by LiDAR to represent 3D environmental information. However, in complex environments containing dynamic objects such as vehicles and pedestrians, these objects introduce significant artifacts, severely impacting the robustness of mapping and localization systems. Effectively filtering out dynamic objects from point cloud data is crucial for improving localization accuracy.

[0003] Dynamic point cloud filtering can be categorized into motion detection-based methods, semantic segmentation-based methods, and clustering and tracking-based methods. The clustering and tracking-based dynamic point cloud filtering method first preprocesses the point cloud, then uses a clustering algorithm to segment spatially adjacent point clouds into independent object instances, extracting geometric features for each instance. Next, a multi-object tracking algorithm is used to associate detected object instances in consecutive frames and predict their motion trajectories, thereby determining their dynamic nature based on the object's motion state. Finally, point clouds belonging to dynamic objects are removed based on the determination results.

[0004] To improve the generalization of dynamic point cloud filtering methods based on clustering and tracking, for example, patent application CN115861968 A discloses a dynamic obstacle removal method based on real-time point cloud data. This invention first acquires object information to form real-time 3D point cloud data; through preprocessing, the target point cloud is divided into sparse and dense point clouds; the sparse point cloud is clustered using the density-based spatial clustering algorithm DBSCAN with noise, dividing the point cloud data into several clusters, and bounding box fitting is performed on each cluster to obtain the geometric feature information of each individual cluster; based on the Hungarian and Kalman filtering algorithms, multi-target matching and tracking are performed on reasonable clusters to distinguish between dynamic and static points; point cloud matching is performed on the dense point cloud using KD-Tree; the dynamic and static attributes of the cluster target are determined based on the dynamic and static ratio of points in the cluster, and dynamic targets are filtered out. This invention improves the generalization ability of dynamic obstacle detection without prior models. However, it is prone to misclassifying dynamic objects with a moving speed below the threshold as static objects. Furthermore, the dynamic and static attributes of clustering targets are determined by the ratio of dynamic to static objects, which affects the further improvement of filtering accuracy and robustness in complex scenarios with varying dynamic object density and motion patterns. Summary of the Invention

[0005] The purpose of this invention is to overcome the defects of the prior art and propose a dynamic point cloud filtering method based on adaptive adjustment of comprehensive confidence and benchmark parameters, which is used to solve the technical problems of poor filtering accuracy and robustness in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Initialize parameters:

[0008] Initialization includes a 3D spatial scene containing various dynamic and static objects. Each frame contains A set of point cloud data in three dimensions and its corresponding data including The set of semantic labels for each semantic category, and the baseline parameter data set are... ,in , , , No. Frame number 3D point cloud data The corresponding semantic tags are , For filtration intensity, This is the penalty coefficient for dynamic objects. This is the confidence scaling factor;

[0009] (2) Calculate the complexity score and temporal factor for each frame of 3D point cloud data:

[0010] Dynamic object proportions based on each frame of 3D point cloud data Semantic diversity Motion complexity and spatial homogeneity Calculating the complexity score for each frame of 3D point cloud data At the same time, through the front Cumulative number of frames of 3D point cloud data Calculate the temporal factor of each frame of 3D point cloud data ;

[0011] (3) Adaptive adjustment of the baseline parameter dataset based on complexity score and time series factor:

[0012] Complexity score based on each frame of 3D point cloud data and time series factors For the benchmark parameter data set Perform adaptive adjustments to obtain the adjusted baseline parameter data set. ;

[0013] (4) Obtain the dynamic point cloud filtering results:

[0014] Using the adjusted baseline parameter data set Calculate each 3D point cloud data Overall confidence level and will With the pre-set dynamic judgment threshold satisfy The dynamic 3D point cloud data is removed to obtain a static point cloud set.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] (1) The present invention judges the dynamic and static attributes of objects by using the semantic, temporal and geometric information in the comprehensive confidence of each three-dimensional point cloud data. This avoids the defect of the prior art that it is easy to misjudge dynamic objects with a moving speed below the threshold as static objects, reduces the probability of missing dynamic points, and effectively improves the filtering accuracy.

[0017] (2) Based on complexity score and time factor, the present invention adaptively adjusts the benchmark parameter data set, realizing the adaptive adjustment of benchmark parameters according to the complexity of the scene in complex scenes with dynamic object density and motion mode. This avoids the defect of poor adaptability caused by the use of fixed parameters in the prior art, and improves the robustness of filtering while improving the filtering accuracy. Attached Figure Description

[0018] Figure 1 is a flowchart of the implementation of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0020] Referring to Figure 1, the present invention includes the following steps:

[0021] Step 1) Initialize parameters:

[0022] Firstly, the complex environment will be scanned using a multi-line lidar to obtain data including... Each frame contains multiple original 3D point cloud data points, which are transformed into the robot's body coordinate system. Outliers are removed from the transformed original 3D point cloud data in each frame, resulting in a dataset containing... This is a collection of 3D point cloud data. Coordinate system transformation ensures that all frames of data are in the same spatial reference frame, which is a prerequisite for subsequent time-series analysis. Outlier removal filters out sparse noise points generated by sensor noise, improving point cloud quality and preventing these points from interfering with subsequent calculations such as clustering and mesh generation.

[0023] Then for the first Frame number 3D point cloud data Perform semantic segmentation and process the resulting 3D point cloud data. semantic category tags and Performing timestamp alignment yields a set of semantic labels corresponding to the point cloud set. Timestamp alignment ensures a strict temporal correspondence between the semantic labels and the point cloud data, providing an accurate foundation for dynamic determination by fusing semantic information.

[0024] The three-dimensional space containing three-dimensional point cloud data along Divide the axis to obtain A three-dimensional spatial grid is used. Gridding is the basis for calculating the uniformity of space. It discretizes a continuous space, which makes it easier to count the number of points in each grid, thereby quantifying the uniformity of the point cloud distribution.

[0025] The formula for mesh generation is:

[0026] ;

[0027] in, 3D point cloud data The corresponding grid index, For point Three-dimensional spatial coordinates, Let be the minimum coordinates of the point cloud bounding box. This is the minimum default grid side length.

[0028] Step 2) Calculate the complexity score and temporal factor for each frame of 3D point cloud data:

[0029] Dynamic object proportions based on each frame of 3D point cloud data Semantic diversity Motion complexity and spatial homogeneity Calculating the complexity score for each frame of 3D point cloud data At the same time, through the front Cumulative number of frames of 3D point cloud data Calculate the temporal factor of each frame of 3D point cloud data Its core is to quantify the dynamic complexity and processing progress of point cloud data. Among them, the proportion of dynamic objects reflects the proportion of dynamic points in each frame of point cloud; semantic diversity reflects the richness of semantic categories of point cloud; motion complexity reflects the complexity of point cloud motion; spatial uniformity is used to measure the degree of dispersion of point cloud distribution in spatial grid; temporal factor introduces the time accumulation effect, reflecting that the system's cognition of the environment tends to stabilize as data accumulates.

[0030] The proportion of dynamic objects Semantic diversity Motion complexity and spatial homogeneity and complexity score The calculation formulas are as follows:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] in, This refers to the number of points in a 3D point cloud dataset that have been labeled with dynamic semantic category tags. For the first The probability of each semantic category occurring. For the first The number of 3D point cloud data points corresponding to each semantic category It is the variance function. 3D point cloud data The velocity vector, , The total number of velocity vectors, This represents the total number of three-dimensional spatial grids. For the first The amount of 3D point cloud data contained in the mesh. for The average value of the 3D point cloud data contained therein, , , , These are the weighting coefficients.

[0039] It is a factor that gradually saturates from 0 to 1, representing the "maturity" of the system's processing. When the value is low, it indicates that the system is still in the learning and adaptation phase, and parameter adjustments should be relatively conservative; as the number of frames processed increases to exceed the empirical threshold, A value of 1 indicates that the system has fully perceived the scene and can make full use of parameter adjustments.

[0040] Time series factor The calculation formula is:

[0041] ;

[0042] in The number of frames is the empirical threshold.

[0043] Step 3) Adaptively adjust the baseline parameter dataset based on complexity scores and time series factors:

[0044] Complexity score based on each frame of 3D point cloud data and time series factors For the benchmark parameter data set Perform adaptive adjustments to obtain the adjusted baseline parameter data set. Scoring by complexity and time series factors Adaptive adjustment of the baseline parameter dataset replaces the fixed threshold; when complexity scoring... High and time series factor A larger value indicates higher scene complexity and a more stable processing state. By adaptively adjusting the baseline parameter dataset, the filtering intensity will be actively enhanced, the dependence on motion and geometric features will be strengthened, and the dynamic object penalty coefficient will be increased to impose a greater dynamic object penalty on object points that have been semantically labeled as dynamic. When the semantic information may be less reliable, the confidence scaling factor will be reduced to allow more stable information such as geometry to occupy a larger proportion.

[0045] Among them Filtration intensity Dynamic object penalty coefficient and confidence scaling factor The adaptive adjustment formulas are as follows:

[0046] ;

[0047] ;

[0048] ;

[0049] in, , , They are respectively , , The results of the adjustment , This is the proportionality coefficient.

[0050] Step 4) Obtain the dynamic point cloud filtering results:

[0051] Using the adjusted baseline parameter data set Calculate each 3D point cloud data Overall confidence level and will With the pre-set dynamic judgment threshold satisfy Dynamic 3D point cloud data is removed to obtain a static point cloud set. A comprehensive confidence level is then constructed. To address the problem of dynamic point leakage filtering, the comprehensive confidence score integrates three complementary types of information: semantic, temporal, and geometric. Among them, semantic confidence score... Identifying potential dynamic objects using prior knowledge from labels, with time-series confidence. Kalman filter residual detection of motion inconsistencies and geometric confidence The positional changes are directly captured through depth residuals, and the fusion weights of the three are not fixed, but are adaptively adjusted by the parameters from step 3. The dynamic decision-making process intelligently adjusts the weights of three different information sources. Simultaneously, the dynamic judgment threshold and filtering intensity form a positive feedback mechanism, thereby intelligently adjusting the dynamic judgment threshold. Finally, dynamic points are filtered out by comparing the comprehensive confidence level with the dynamic judgment threshold.

[0052] Overall confidence level and dynamic threshold determination The calculation formulas are as follows:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] in, , , For each 3D point cloud data Semantic confidence, temporal confidence, and geometric confidence. for Basic semantic weights, , The maximum and minimum values ​​of the weights, It is an exponential function. Let be the residual norm. For residual scaling parameters, For absolute depth residuals, Based on the threshold, This is the proportionality coefficient.

[0059] Each 3D point cloud data semantic confidence Time series confidence and geometric confidence The method to obtain it is as follows:

[0060] Semantic confidence Acquisition method: Each 3D point cloud data semantic confidence , among them Basic semantic weights It is based on each semantic category label The basic semantic weights are obtained from the preset semantic category weight mapping. The semantic confidence score for that point is then calculated using the formula for semantic confidence score. ;

[0061] Time series confidence Acquisition method: Based on the 3D point cloud data point cloud set, candidate dynamic point sets of 3D point cloud data are obtained through filtering. Then, sampling quotas are allocated based on the proportion of the 3D point cloud data. In the division obtained At most one random sample is taken from each of the grids. After traversing all grid points, the extracted points are aggregated to form a key dynamic point set. ,use frame The motion trajectory is constructed, and a state vector is established for each trajectory. Kalman filtering is used for state prediction and update, and finally, the residual norm is used. Calculate time series confidence ;

[0062] Each 3D point cloud data Time series confidence The residual norm The calculation formula is:

[0063] ;

[0064] ;

[0065] in, These are the actual observed values ​​of the dynamic points. For dimension The observation matrix The prior estimate of the state vector obtained by Kalman filtering is given. Let be the Euclidean norm of the vector. For dimension The state transition matrix.

[0066] Among them, candidate dynamic point set Sampling quota The Kalman filter prediction and update formulas are as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] in, This is a list of semantics for dynamic objects within a semantic category distribution. for Three-dimensional point cloud data belonging to the semantics of dynamic objects. The preset total number of sampling points, For error covariance, The process noise covariance matrix is... For covariance, To observe the noise covariance matrix, Kalman gain;

[0072] Geometric confidence Acquisition method: Extract each point from the 3D data point cloud set. Based on its three-dimensional coordinates and projection formula, it is projected onto a two-dimensional depth image. The corresponding pixel position is recorded, and the depth value of that pixel is also recorded. Then, with the obtained depth image The corresponding pixel position and its corresponding depth value For reference, calculate the absolute depth residual between the two. Finally, through absolute depth residuals Calculate geometric confidence level ;

[0073] Each 3D point cloud data geometric confidence The formula for calculating the absolute depth residual is as follows:

[0074] ;

[0075] in, , For each 3D point cloud data Two-dimensional depth image , The depth value of the corresponding pixel.

[0076] The projection formula for a two-dimensional depth image is:

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] in Yaw angle The pitch angle, The width / height of the 2D depth image. This represents the total field of view of the lidar. This is a downward field of view.

Claims

1. A dynamic point cloud filtering method based on adaptive adjustment of comprehensive confidence and benchmark parameters, characterized in that, The steps include: (1) Initializing parameters: Initializing the three-dimensional spatial scene including different dynamic and static objects. Each frame contains A set of point cloud data in three dimensions and its corresponding data including The set of semantic labels for each semantic category, and the baseline parameter data set are... ,in ， ， , the Frame number 3D point cloud data The corresponding semantic tags are ， For filtration intensity, This is the penalty coefficient for dynamic objects. (2) Calculate the complexity score and temporal factor of each frame of 3D point cloud data: based on the dynamic object ratio of each frame of 3D point cloud data. Semantic diversity Motion complexity and spatial homogeneity Calculating the complexity score for each frame of 3D point cloud data At the same time, through the front Cumulative number of frames of 3D point cloud data Calculate the temporal factor of each frame of 3D point cloud data ； (3) Adaptive adjustment of the baseline parameter dataset based on complexity score and time series factor: based on the complexity score of each frame of 3D point cloud data and time series factors For the benchmark parameter data set Perform adaptive adjustments to obtain the adjusted baseline parameter data set. (4) Obtain dynamic point cloud filtering results: using the adjusted baseline parameter data set Calculate each 3D point cloud data Overall confidence level and will With the pre-set dynamic judgment threshold satisfy The dynamic 3D point cloud data is removed to obtain a static point cloud set.

2. The method according to claim 1, characterized in that, The point cloud set mentioned in step (1) is obtained by: combining the point cloud data acquired through multi-line lidar scanning of a complex environment, including... Each frame contains multiple original 3D point cloud data points, which are transformed into the robot's body coordinate system. Outliers are removed from the transformed original 3D point cloud data in each frame, resulting in a dataset containing... A collection of point clouds containing three-dimensional point cloud data.

3. The method according to claim 1, characterized in that, The semantic tag set mentioned in step (1) is obtained by: [The method for obtaining the set of semantic tags is missing here, but can be left as is]. Frame number 3D point cloud data Perform semantic segmentation and assign semantic category labels to the segmented data. and Perform timestamp alignment to obtain the semantic tag set corresponding to the point cloud set.

4. The method according to claim 1, characterized in that, The dynamic object scale described in step (2) Semantic diversity Motion complexity and spatial homogeneity and complexity score and time series factors The calculation formulas are as follows: ； ； ； ； ； ； ； ;in, This refers to the number of points in a 3D point cloud dataset that have been labeled with dynamic semantic category tags. For the first The probability of each semantic category occurring. For the first The number of 3D point cloud data points corresponding to each semantic category It is the variance function. 3D point cloud data The velocity vector, ， The total number of velocity vectors, This represents the total number of grid cells after meshing in a 3D space. For the first The amount of 3D point cloud data contained in the mesh. for The average value of the 3D point cloud data contained therein, 、 、 、 These are the weighting coefficients. The number of frames is the empirical threshold.

5. The method according to claim 1, characterized in that, The reference parameter data set described in step (3) Perform adaptive adjustments, including for Filtration intensity Dynamic object penalty coefficient and confidence scaling factor The adaptive adjustment formulas are as follows: ； ； ;in, 、 、 They are respectively 、 、 The results of the adjustment 、 This is the proportionality coefficient.

6. The method according to claim 5, characterized in that, The comprehensive confidence level mentioned in step (4) and dynamic threshold determination The calculation formulas are as follows: ； ； ； ； ;in, 、 、 For each 3D point cloud data Semantic confidence, temporal confidence, and geometric confidence. for Basic semantic weights, 、 The maximum and minimum values ​​of the weights, It is an exponential function. Let be the residual norm. For residual scaling parameters, For absolute depth residuals, Based on the threshold, This is the proportionality coefficient.

7. The method according to claim 6, characterized in that, Each 3D point cloud data semantic confidence , among them Basic semantic weights It is based on each semantic category label It is obtained from the preset semantic category weight mapping.

8. The method according to claim 6, characterized in that, Each 3D point cloud data Time series confidence The residual norm The calculation formula is: ； ;in, These are the actual observed values ​​of the dynamic points. For the observation matrix, The prior estimate of the state vector obtained by Kalman filtering is given. For the Euclidean norm, Let be the state transition matrix.

9. The method according to claim 6, characterized in that, Each 3D point cloud data geometric confidence The absolute depth residual The calculation formula is: ;in, 、 The 3D point cloud data in the 2D depth image projected from each frame of 3D point cloud data. 、 corresponding pixels 、 The depth value.

Images