Density-based iterative voxel downsampling method

Abstract

The invention provides an improved density-based iterative voxel downsampling method, and more flexible and adaptive point cloud downsampling is realized by introducing point cloud density information and iterative voxel size. According to a traditional voxel filtering method, a fixed voxel size is used, important details can not be effectively reserved in an area with large density change, and excessive downsampling can be caused in a sparse area. According to the method, the point cloud number of each voxel is calculated, and the voxels are dynamically segmented and subdivided according to the preset threshold value, so that more point cloud information is kept in a high-density region, redundant data is reduced in a low-density region, and the downsampling precision and the calculation efficiency are improved. Specifically, when the number of points in a voxel exceeds a set threshold value and the size of the voxel is larger than the minimum size, the algorithm divides the voxel into eight sub-voxels, and recursive subdivision continues until conditions are met. The method is high in adaptability, can dynamically adjust the sampling strategy according to the density change of the point cloud, and remarkably improves the precision and efficiency of point cloud processing. Compared with a traditional voxel filtering method, the method has the advantages that details of a dense region can be better reserved, redundant points of a sparse region are reduced, and an efficient and accurate point cloud downsampling solution is provided.

Description

Technical Field

[0001] This invention belongs to the technical field of point cloud processing and data optimization, specifically to a density-based iterative voxel downsampling method. This method is primarily applied to point cloud data downsampling, improving the efficiency and accuracy of point cloud data processing by adaptively adjusting voxel size and density estimation. By introducing recursive voxel subdivision and parallel computing techniques, this invention enables efficient point cloud downsampling in complex environments, optimizing the accuracy and processing speed of point cloud data, and providing efficient and accurate data support for subsequent point cloud analysis and applications. Background Art

[0002] With the rapid development of intelligent technologies such as autonomous driving, robotic navigation, and smart cities, point cloud data has been widely used in various fields as an important data source for environmental perception and target detection. Point cloud data, primarily acquired by sensors such as LiDAR (Light Detection and Ranging) or depth cameras, provides rich three-dimensional spatial information and plays a vital role in environmental modeling, obstacle detection, and path planning. However, due to the high dimensionality and large size of point cloud data, the efficiency of processing this data has become a major bottleneck. This is especially true in application scenarios such as autonomous driving and drone cruising, where real-time requirements are high. How to efficiently and accurately process point cloud data has become an unresolved issue.

[0003] To address this challenge, point cloud data downsampling technology has been proposed as a solution. Through downsampling, the amount of point cloud data can be reduced, thereby alleviating the computational burden of subsequent processing and improving the efficiency of data processing. Traditional voxel filtering methods divide the point cloud space into uniform voxel blocks and use the centroid of each voxel block as the representative point of the block to reduce redundant information in the point cloud, thereby achieving downsampling. This method can effectively reduce the volume of point cloud data and reduce computational complexity to a certain extent. However, traditional voxel filtering methods have certain limitations, especially when faced with areas with uneven density in point cloud data, their effect is not ideal. In high-density areas, due to the limitation of fixed voxel size, detailed information may be lost, affecting the accuracy of subsequent processing; while in low-density areas, the excessively large voxel size may make it impossible to effectively reduce the amount of data, and may instead introduce redundant data.

[0004] To address these issues, this paper proposes a density-based iterative voxel downsampling method. This method estimates the density of points within each voxel block and dynamically adjusts the voxel size to preserve more detailed information in high-density areas while increasing the voxel size to effectively reduce redundant data in low-density areas. This method effectively reduces computational effort while preserving critical information, improving the efficiency and accuracy of point cloud data processing and meeting the needs of applications with high real-time requirements. Summary of the Invention

[0005] The present invention aims to provide a density-based iterative voxel downsampling method specifically designed for efficiently and accurately downsampling point cloud data in complex environments. This method adaptively adjusts voxel size to handle both high- and low-density areas, preserving detail in high-density regions while reducing redundant point cloud data in low-density areas, thereby optimizing point cloud processing efficiency. Furthermore, the present invention incorporates a recursive voxel subdivision mechanism, enabling it to handle areas with large density variations, ensuring both accuracy and efficiency during the downsampling process.

[0006] The density-based iterative voxel downsampling method of the present invention comprises the following steps:

[0007] Step 1: Point cloud data collection and region of interest demarcation. Obtain raw point cloud data through lidar or depth camera, and filter the region of interest by setting the spatial range. Perform preliminary processing on the point cloud data, filter out invalid points that are too far or too high, and retain data that is meaningful for subsequent analysis.

[0008] Step 2: Density estimation and voxel size adjustment. Calculate the number of point clouds within each voxel to obtain the density information of the voxel. Dynamically adjust the voxel size based on the density value: reduce the voxel size in high-density areas to retain more details; increase the voxel size in low-density areas to reduce redundant data.

[0009] Step 3: Voxel segmentation and recursive subdivision. In high-density areas, if the number of point clouds within a voxel exceeds the preset threshold and the voxel size is larger than the minimum size requirement, the voxel is segmented into 8 sub-voxels. The sub-voxels are recursively subdivided until the minimum size or point cloud number requirement is met;

[0010] Step 4: Voxel data integration and optimized output. Analyze the density information of all voxels and adaptively adjust the size of the voxels based on the density. Output the optimized point cloud data, remove redundant point clouds, retain important spatial information, and provide a concise and useful point cloud dataset.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] 1. High matching accuracy: By dynamically adjusting the voxel size, details in high-density areas are preserved, redundant data in low-density areas is reduced, and the accuracy of point cloud downsampling is improved;

[0013] 2. Significant improvement in computational speed: Through improved density estimation methods and dynamic voxel resizing, unnecessary computational steps are reduced, making data processing more efficient. Especially when processing point cloud data with large density variations, the new method can intelligently adjust the computational strategy based on data characteristics, avoiding over-computation and further improving processing speed.

[0014] 3. Strong adaptability: Adaptive voxel resizing and recursive subdivision mechanisms are used to process point cloud data in areas of different densities, enhancing the robustness and adaptability of the algorithm in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Density-based iterative voxel filtering structure diagram

[0016] Figure 2 Flowchart of density-based iterative voxel filtering program

[0017] Figure 3 Comparison of the average downsampling running time before and after optimization DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings.

[0019] The core idea of ​​this invention is to gradually subdivide the voxel area to achieve point cloud downsampling. By counting the number of points in each voxel, when the number of points exceeds a set threshold, the algorithm subdivides the voxel into smaller sub-voxels until a minimum size requirement or point count condition is met.

[0020] The system architecture diagram of the recursive voxel segmentation method based on voxel density estimation is as follows Figure 1 As shown, it includes the following steps:

[0021] Step 1: After acquiring the point cloud data, the system first calculates the number of point clouds for each voxel and then evaluates the density of each voxel. Assume that there are N points in the voxel V. points points, the voxel size is V x ,V y ,V z , where V x ,V y ,V z Represents the length of the voxel in the x, y, and z directions respectively. The maximum number of points is N. max , the default minimum voxel size is L x ,L y ,L z If the voxel block satisfies N points >N max and V x ,V y ,V z Greater than L x ,L y ,L z , then the voxel needs to be segmented. The calculation process is evaluated by the following formula:

[0022]

[0023] Step 2: Voxel segmentation. For each voxel that exceeds the threshold, first calculate the center point of the voxel and generate 8 smaller sub-voxels. Assuming the boundaries of the current voxel are: (Xmin, Ymin, Xmax, Xmin), calculate the voxel midpoint using the following formula:

[0024]

[0025] Then, the current voxel is divided into 8 sub-voxels, and the boundaries of each sub-voxel are as follows:

[0026] Voxel1(X min ,mid x ,Y min ,mid y ,Z min ,mid z )

[0027] Voxel2(mid x ,X max ,Y min ,mid y ,Z min ,mid z )

[0028] Voxel3(X min ,mid x ,mid y ,Y max ,Z min ,mid z )

[0029] Voxel4(mid x ,X max ,mid y ,Y max ,Z min ,mid z )

[0030] Voxel5(X min ,mid x ,Y min ,mid y ,mid z ,Z max )

[0031] Voxel6(mid x ,X max ,Y min ,mid y ,mid z ,Z max )

[0032] Voxel7(X min ,mid x ,mid y ,Y max ,mid z ,Z max )

[0033] Voxel8(mid x ,X max ,mid y ,Y max ,mid z ,Z max )

[0034] You can get 8 sub-voxels that are half the size of the original voxel.

[0035] Step 3: Recursive subdivision

[0036] For each segmented subvoxel, if the number of points N in the subvoxel points Still greater than the maximum number of points N max , and the voxel block size is larger than the preset minimum size L x ,L y ,L z , then continue to subdivide the voxels until either of the two conditions is not met. The program flow chart is as follows Figure 2 shown.

[0037] In order to verify the advantages of the density-based iterative voxel filtering algorithm proposed in this patent in preserving and processing efficiency of medium and short-range non-ground point clouds, a test set containing 446 frames of point cloud data was selected under the same experimental environment. For medium and short-range point clouds within the range of [0, 20] meters, Grid_sampling, Random_sampling, Voxel_filter and the algorithm in this paper were used for comparative experiments. The proportion of medium and short-range point clouds and the proportion of non-ground point clouds in this range of each algorithm were counted, as shown in the following figure: Figure 3 The experimental results show that the proposed algorithm accounts for 87.52% of the medium and short-range point clouds, of which the non-ground point cloud accounts for 24.86%, which is significantly better than Grid_sampling (85.43% of medium and short-range points, 8.00% of non-ground points), Random_sampling (95.23%, 18.80%) and Voxel_filter (81.14%, 19.57%). This shows that the proposed algorithm can better retain the key medium and short-range non-ground point cloud information, providing richer and more effective data support for subsequent clustering processing.

Claims

1. A density-based iterative voxel downsampling method, characterized in that The following steps are involved: Step 1: Use multi-line LiDAR to collect and process point cloud data. The point cloud data acquired by the LiDAR is divided into regions of interest. On this basis, the point cloud data is divided into several voxels to prepare for subsequent voxel downsampling and optimization processing. Step 2: Voxel density estimation and size adjustment: Calculate the number of point clouds within each voxel to obtain the density information of each voxel. The voxel size is dynamically adjusted based on the density value of the voxel: the voxel size is reduced in high-density areas to ensure that details are not lost; the voxel size is increased in low-density areas to reduce unnecessary calculations and the number of point clouds. Step 3: Voxel Segmentation and Recursive Subdivision. In high-density areas, when the number of points within a voxel exceeds a set threshold and the voxel size is greater than the minimum size, the algorithm splits the current voxel into eight sub-voxels. These sub-voxels are recursively subdivided until the minimum size requirement or point cloud quantity requirement is met. Step 4: Voxel Data Integration and Output. A comprehensive analysis of the density information of all voxels is performed to ensure that the number of point clouds in each area meets the specified requirements. In high-density areas, more detailed information is retained; in low-density areas, redundant point cloud data is reduced. Finally, the optimized point cloud data is output, ensuring that the data volume is effectively reduced while retaining important details.

2. The method for dividing a region of interest according to claim 1, wherein: The following steps are included: Step 1: Set the spatial extent of the region of interest. The spatial extent of the region of interest is determined based on the LiDAR's detection range and the characteristics of the study scenario. The scenario studied in this invention involves a low-speed unmanned vehicle in a semi-enclosed environment, traveling at speeds no greater than 5 km / h. Therefore, it is crucial to focus on obstacles in the medium to short distances in front of the vehicle. Step 2: Use the passthrough filtering algorithm to perform preliminary point cloud filtering. Using the passthrough filtering algorithm, the point cloud data is processed within the specified X, Y, and Z axis ranges. The algorithm determines whether the coordinate value of each point falls within the preset numerical range and only retains the point cloud data that meets the conditions. Point cloud data that falls outside the range is automatically discarded to remove redundant information and noise, reducing unnecessary computational burden. Specific settings are: X-axis range is [0.0, 50.0] meters, Y-axis range is [-5.0, 5.0] meters, and Z-axis range is [-1.0, 2.0] meters.

3. The voxel density estimation and size adjustment step according to claim 1, characterized in that: The following steps are included: Step 1: Calculate the number of point clouds within each voxel. By calculating the number of point clouds within each voxel, we can obtain the density information of the voxel, which provides a basis for subsequent voxel adjustment. The number of point clouds within each voxel reflects the point cloud density of the area and provides a basis for dynamic adjustment of voxel size. Step 2: Dynamically adjust voxel size. In denser areas, voxel size is reduced to preserve detail. This ensures that important spatial details are not lost in dense regions. In less dense areas, voxel size is increased, reducing the computational effort required to calculate redundant point cloud data and ensuring processing efficiency. Step 3: Density-Adaptive Voxel Resizing. Based on the calculated voxel density information, the voxel size is dynamically adjusted to accommodate the point cloud density in different regions. This process ensures that high-density areas have more sample points, while low-density areas reduce unnecessary redundant data by increasing the voxel size.

4. The voxel segmentation and recursive subdivision step according to claim 1, characterized in that: The following steps are included: Step 1: Perform voxel segmentation on each high-density region. When the number of points within a voxel exceeds a set threshold and the voxel size exceeds a minimum set size, the algorithm will segment the voxel. Specifically, the voxel will be split into 8 sub-voxels, each half the size of the original voxel. Step 2: Recursively subdivide the subvoxels. Within each segmented subvoxel, if the number of points still exceeds the threshold and the subvoxel size exceeds the minimum size requirement, recursive subdivision continues. This process continues until the minimum size requirement or point cloud number requirement is met.

5. The voxel data integration and output step according to claim 1, characterized in that: The following steps are included: Step 1: Density Analysis and Adaptive Voxel Size Adjustment. The density information of each voxel is analyzed and the voxel size is adaptively adjusted based on the density. This step ensures that high-density areas retain more details, while low-density areas reduce redundant point cloud data, optimizing the representation of the point cloud data. Step 2: Output optimized point cloud data. By optimizing the voxel data, the optimized point cloud data is ultimately output. This data removes redundant point clouds while retaining key spatial information, providing a concise and useful point cloud dataset.

Images