A sparse point cloud accuracy assessment method without feature extraction
By using CSF and KD-Tree technology, elevation can be directly obtained from sparse point clouds without feature extraction. Combined with the control point correction process, the automation and accuracy issues of sparse point cloud evaluation are solved, and efficient and stable evaluation results are achieved.
Patent Information
- Application Number
- CN202511055274.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-30
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2045-07-30
AI Technical Summary
The sparse point clouds output by existing SLAM-based vehicle-mounted laser scanning systems are difficult to evaluate with high precision, speed, and automation. Traditional methods rely on feature extraction, which has problems such as large errors, high computational complexity, and strong inapplicability.
The cloth simulation filter (CSF) algorithm is used to extract ground point clouds. Combined with KD-Tree and median filtering techniques, the checkpoint elevation is obtained through neighborhood point search. The control point correction process is used to eliminate outliers and calculate the height difference error index to achieve accuracy evaluation without feature extraction.
It realizes full automation of elevation accuracy assessment of sparse point clouds, reduces labor costs and computational complexity, enhances the stability and consistency of evaluation results, and can truly reflect the point cloud measurement performance.
Smart Images
Figure CN120563510B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of three-dimensional laser scanning, and in particular to a sparse point cloud accuracy assessment method without feature extraction. Background Art
[0002] With the continuous evolution of 3D laser scanning technology, vehicle-mounted mobile laser scanning systems have become important equipment in fields such as road surveying and mapping, urban 3D modeling, and bridge and tunnel structure inspection, due to their "vehicle follows the road, scanning without blind spots" characteristics. Compared with traditional measurement methods, vehicle-mounted systems can simultaneously obtain high-density 3D point clouds of the road and surrounding environment during the measurement process, significantly reducing the workload of manually setting measurement points, laying out markers, and subsequent data collation. At the same time, the vehicle can continuously collect data while driving, significantly improving measurement efficiency and safety. Compared with airborne mobile laser scanning systems, vehicle-mounted systems have a shorter distance between the laser and the ground target, and the scanning resolution can reach several centimeters or even higher. They have lower installation and debugging costs and do not have high requirements for the flight environment. They can flexibly enter complex scenes such as urban roads, alleys, and under bridges that are difficult or restricted for airborne aircraft to cover.
[0003] Point clouds are the core output of vehicle-mounted mobile laser scanning systems. Their accuracy and integrity directly impact subsequent applications such as 3D modeling, road geometry extraction, and pipeline inspection. Currently, mainstream technologies for assessing the accuracy of vehicle-mounted point clouds can be divided into the following two categories:
[0004] 1. Accuracy assessment method based on control points
[0005] This method deploys high-precision control points with known coordinates within the survey area (usually using a total station or RTK technology to obtain reference coordinates). After scanning, the coordinates of the control points in the corresponding area are extracted from the point cloud and their coordinate deviations are compared with the measured reference values. Common evaluation metrics include root mean square error (RMSE), maximum deviation, and average deviation to quantify the absolute accuracy and measurement consistency of the point cloud in space. This method is simple and intuitive, and can directly reflect the overall measurement accuracy of the system. However, it is highly dependent on the number of control points deployed, the uniformity of their distribution, and on-site measurement conditions (such as signal obstruction and varying lighting).
[0006] 2. Accuracy evaluation method based on feature extraction
[0007] This method automatically or semi-automatically extracts geometrically significant feature elements from the point cloud, such as road curbs, tops of marker posts, pipeline tops, or building facades. By fitting the extracted point cloud features and comparing them with design data, measurement data, or known feature models, it assesses elevation and position deviations. Compared to the control point method, the feature extraction method offers greater flexibility, allowing for local or global accuracy assessments of different landform types without the need for additional control points. However, this method places higher demands on algorithm complexity, robustness of feature extraction, and point cloud quality under varying lighting and occlusion conditions.
[0008] In summary, control point-based and feature extraction-based evaluation methods each have their own advantages: the former has a mature measurement process and is easy to verify; the latter relies on automated algorithms and scene features, but can achieve a wider range of calibration-free evaluation.
[0009] In recent years, to meet the real-time performance requirements of vehicle-mounted mobile laser scanning systems in high-speed driving scenarios, a growing number of studies have incorporated simultaneous localization and mapping (SLAM) technology into the point cloud data acquisition process. SLAM algorithms enable systems to construct maps and locate themselves in real time while driving. However, due to the dual considerations of point cloud data volume and computational efficiency in the SLAM process, it is often necessary to first downsample the original laser point cloud or retain only a portion of keyframe data before using it for odometry estimation and map updates. While this approach can speed up positioning and mapping, it inevitably results in a reduced density of the point cloud map output by SLAM, a loss of detail, and an inability to fully reproduce the microscopic undulations and small-scale features of the road.
[0010] Traditional control point-based accuracy assessment methods for low-density SLAM-based point clouds face the following major limitations: First, due to the sparse point cloud, it is often difficult to accurately extract the echo coordinates of corresponding control points from the point cloud data, resulting in a lack of reliability in the assessment results. Second, to ensure the identifiable control points, a sufficiently dense number of additional artificial landmarks must be deployed on-site, increasing deployment costs and the difficulty of on-site operations. Furthermore, while feature extraction-based methods can assess low-density point clouds, the feature extraction process itself relies on point cloud quality and algorithm robustness. For example, the Chinese patent "CN108447126B: Laser Point Cloud Accuracy Assessment Method for Mobile Measurement Systems Based on Reference Planes" uses the plane equation of a reference plane as a basis. Using the point cloud data of this reference plane acquired by the MMS system, the mean square error (MSE) of the point cloud data in the E, N, and U directions is calculated for assessment. When point clouds are sparse, texture information is insufficient, or there are occlusions, feature extraction algorithms are prone to mismatching or misidentification, resulting in increased assessment errors. In addition, the feature extraction process is computationally complex and has high implementation costs, and often requires adjustment of algorithm parameters and secondary development for different scenarios or landform types, which reduces the consistency and automation level of the evaluation process.
[0011] Therefore, while the current SLAM-assisted vehicle-mounted laser scanning system ensures real-time performance, the output point cloud is difficult to directly apply to the traditional high-precision evaluation process, and the existing feature extraction evaluation method is also difficult to balance simplicity and universality. This makes the accurate, fast and automated evaluation of the elevation accuracy of point clouds in high-speed and high-dynamic road conditions a technical problem that needs to be solved urgently. Summary of the Invention
[0012] The purpose of the present invention is to provide a sparse point cloud accuracy assessment method without feature extraction. It does not extract any feature points, lines, or surfaces. On the basis of reducing the errors that may be caused by feature extraction and matching, it reduces the labor cost and computational complexity as much as possible, thereby providing a highly consistent and automated assessment method for the accuracy assessment of sparse point clouds.
[0013] The technical solutions of the present invention are as follows:
[0014] Step 1: Use the cloth simulation filtering algorithm to extract ground points from the original point cloud to obtain the ground point cloud ;Deploy control points and check points in the point cloud measurement area;
[0015] Step 2: Calculate the local neighborhood radius of the ground point cloud ,in, is the area of the 2D bounding box of the ground point cloud, is the total number of points in the ground point cloud;
[0016] Step 3: Calculate the elevation of the checkpoint in the ground point cloud. Use the plane coordinates of the checkpoint as the center of the circle and the local neighborhood radius as the radius. The points found in the ground point cloud are used as the neighborhood points corresponding to the current checkpoint. The median of the neighborhood point elevations is used as the elevation of the checkpoint in the ground point cloud.
[0017] Step 4: Correct the system height difference between the checkpoint and the ground point cloud using control points;
[0018] Step 5: Calculate the height difference and output the statistical indicators of root mean square error and maximum absolute error; obtain the height difference results of all checkpoints , The elevation of the checkpoint after correction is consistent with the ground point cloud benchmark. To check the elevation of the point in the ground point cloud, the root mean square error and the maximum absolute error are calculated respectively.
[0019] The original point cloud is collected by the mobile measurement system, and control points are arranged in the point cloud measurement area. With checkpoints ; The i-th control point , 、 They represent the plane coordinates of the control points under Gaussian three-dimensional projection, Indicates the elevation of the control point relative to the quasi-geoid; the i-th checkpoint , 、 They represent the plane coordinates of the checkpoint under Gaussian three-dimensional projection, Indicates the elevation of the checkpoint relative to the quasi-geoid; the elevations of the control point and the checkpoint are both obtained through leveling, and both the control point and the checkpoint are located on the road surface in the survey area.
[0020] ,in, For ground point cloud The maximum value in the direction, For ground point cloud The minimum value in the direction, is the maximum value of the ground point cloud in the Y direction, It is the minimum value of the ground point cloud in the Y direction.
[0021] The step three is as follows:
[0022] Construct KD-Tree for ground point cloud;
[0023] Traverse each checkpoint , according to its plane coordinates , search the ground point cloud for the checkpoint Planar distance is less than the local neighborhood radius All points as the current checkpoint Neighborhood point set ;
[0024]
[0025] is a single ground point in the ground point cloud, represents the ground point cloud, Represents the three-dimensional coordinates of a single ground point in the ground point cloud;
[0026] Calculate the median elevation of the neighborhood points as the measured elevation of the checkpoint in the ground point cloud;
[0027]
[0028] The elevation of all checkpoints in the ground point cloud is composed into vectors , For checkpoints The measured elevation in the ground point cloud and the original leveling elevation of all checkpoints are recorded as .
[0029] The step 4 is specifically as follows:
[0030] Use the same method to obtain the elevation of the checkpoint to obtain the elevation of the control point in the ground point cloud :
[0031]
[0032] The corresponding vector of the original leveling elevation of the control point is ;
[0033] Calculate the original height difference vector of the control point :
[0034]
[0035] right Remove outliers and obtain the cleaned height difference set ;in, is the number of height differences in the height difference set after cleaning;
[0036] calculate The arithmetic mean of As system height difference:
[0037]
[0038] Correct the system error of the original level elevation of the checkpoint to obtain the checkpoint elevation consistent with the ground point cloud benchmark .
[0039] The outliers are removed by using the triple standard deviation method; the median of the original height difference vector is calculated; , and all and The absolute mean error between ;
[0040] Eliminate all satisfaction The original height difference sample points are recorded as the cleaned height difference set. .
[0041] The absolute mean error calculation formula is as follows:
[0042]
[0043] in, Represents calculation of the median.
[0044] Beneficial effects of the present invention: The core of the present invention is to completely abandon the traditional geometric feature extraction steps, and directly use neighborhood point search and median filtering technology to efficiently and robustly obtain checkpoint elevation values from sparse ground point clouds. First, the CSF (Cloth Simulation Filter) algorithm is used to automatically separate ground points, and the adaptive neighborhood radius is calculated based on the horizontal bounding box of the point cloud and the number of points, without the need to manually set parameters; then, the point cloud set around each checkpoint is quickly located through KD-Tree, and the elevation is calculated in a median manner to maximize resistance to interference from noise and isolated points. In addition, the present invention introduces a control point correction process, performs the same median elevation acquisition on the control points under the same neighborhood criterion, and calculates the global system elevation difference after eliminating outliers through the triple mean error method, which is used to accurately convert the leveling measurement elevation to the point cloud elevation benchmark, effectively eliminating the overall offset caused by the benchmark difference.
[0045] The present invention realizes a fully automated process for evaluating the elevation accuracy of low-density sparse point clouds output by SLAM, and no longer relies on feature extraction or densely deployed markers on site, thereby minimizing labor costs and algorithm complexity. The combination of median filtering and outlier removal not only greatly enhances the ability to suppress point cloud noise and occasional errors, but also ensures the stability and consistency of the evaluation results under high-speed driving and complex environments. Through system height difference correction, the deviation of different measurement benchmarks is eliminated, so that the output error index can truly reflect the measurement performance of the point cloud itself. Finally, the root mean square error and maximum absolute error are used to perform dual quantitative evaluation of the overall and extreme cases, providing reliable and comprehensive technical support for subsequent system optimization, algorithm iteration and application scenario expansion. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a flow chart of the main steps of the present invention. DETAILED DESCRIPTION
[0047] Figure 1 It is the main flow chart of the technical solution of the present invention, such as Figure 1 As shown, the sparse point cloud accuracy assessment method without feature extraction proposed in the present invention includes the following steps:
[0048] First, the input data includes the original point cloud output by the mobile measurement system, the control points arranged in the point cloud measurement area, and the With checkpoints , where the i-th control point , the i-th checkpoint The elevations of control points and check points are obtained through leveling. At the same time, both control points and check points are located on the road surface in the survey area.
[0049] Step 1: Use CSF algorithm to extract ground points from the original point cloud to obtain the ground point cloud .
[0050] Step 2: Calculate the local neighborhood radius of the ground point cloud ,in is the area of the point cloud 2D bounding box, ,in and The ground point clouds are The maximum and minimum values in the direction, and are the maximum and minimum values in the Y direction, is the total number of points in the ground point cloud;
[0051] Step 3: Calculate the elevation of the checkpoint in the ground point cloud. Since the point cloud output by the SLAM-based vehicle-mounted mobile measurement system is relatively sparse, and the point cloud map itself also has measurement errors, it is impossible to effectively select the coordinates of the checkpoint in the output point cloud by relying solely on control points without using feature extraction. Therefore, the present invention uses the median elevation of the neighborhood points as the elevation of the checkpoint in the point cloud. The specific steps are as follows:
[0052] (1) First, construct a KD-Tree for the ground point cloud;
[0053] (2) Next, traverse each checkpoint , according to its plane coordinates , search the ground point cloud for the point The plane distance is less than the local neighborhood radius All points as the current checkpoint The neighborhood point set of .
[0054]
[0055] (3) Then calculate the median elevation of the neighborhood points as the measured elevation of the checkpoint in the ground point cloud.
[0056]
[0057] (4) Finally, the elevations of all checkpoints in the point cloud are combined into vectors , For checkpoints The measured elevation in the ground point cloud; at the same time, the original leveling elevation of all checkpoints is recorded as .
[0058] Step 4: Correct the system height difference between the checkpoint and the point cloud through the control point. It uses leveling measurement, but it is often difficult to connect to the advanced leveling control network, so most of the time it is a local elevation benchmark, and the point cloud elevation obtained by GNSS / RTK is the result of the SLAM system. It is based on a quasi-geoid of a specific ellipsoid (such as WGS-84). There is a systematic height difference between the two in an absolute sense. If the height difference is not corrected, it is directly used and Subtracting the above values will introduce a large overall deviation, which will lead to distortion in the subsequent root mean square error and maximum absolute error evaluation. Therefore, the present invention designs the following "system height difference correction process":
[0059] (1) Obtain the elevation of the control point in the point cloud in the same way as in the above steps. :
[0060]
[0061] (2) The vector corresponding to the original elevation of the control point is ;
[0062] (3) Calculate the original height difference vector of the control point :
[0063]
[0064] (4) For the errors introduced by field measurement, point cloud generation and neighborhood search, it is necessary to To remove outliers, the present invention uses the triple standard deviation method to screen: calculate the median of the original height difference , and all and The absolute mean error between ; Eliminate all The remaining data is recorded as the height difference set after cleaning ;
[0065] (5) Then calculate The arithmetic mean of is taken as the system height difference:
[0066]
[0067] (6) Finally, the original leveling elevation of the checkpoint is corrected for systematic errors to obtain the checkpoint elevation that is consistent with the point cloud benchmark. .
[0068] Step 5: Calculate the height difference and output the statistical indicators of root mean square error and maximum absolute error. First, get the height difference results of all checkpoints Next, we calculate the root mean square error (RMSE) and maximum absolute error (MAE). The RMSE reflects the overall elevation accuracy of the sparse point cloud. A smaller value indicates that the elevation differences of most checkpoints are concentrated within a smaller range. The MAE reflects the residual magnitude of the worst individual point.
Claims
1. A sparse point cloud accuracy assessment method without feature extraction, characterized in that: The following steps are involved: Step 1: Use the cloth simulation filtering algorithm to extract ground points from the original point cloud to obtain the ground point cloud ;Deploy control points and check points in the point cloud measurement area; Step 2: Calculate the local neighborhood radius of the ground point cloud ,in, is the area of the 2D bounding box of the ground point cloud, is the total number of points in the ground point cloud; Step 3: Calculate the elevation of the checkpoint in the ground point cloud. Use the plane coordinates of the checkpoint as the center of the circle and the local neighborhood radius as the radius. The points found in the ground point cloud are used as the neighborhood points corresponding to the current checkpoint. The median of the neighborhood point elevations is used as the elevation of the checkpoint in the ground point cloud. Step 4: Correct the system height difference between the checkpoint and the ground point cloud using control points; Step 5: Calculate the height difference and output the statistical indicators of root mean square error and maximum absolute error; obtain the height difference results of all checkpoints , The elevation of the checkpoint after correction is consistent with the ground point cloud benchmark. To check the elevation of the point in the ground point cloud, the root mean square error and the maximum absolute error are calculated respectively.
2. The sparse point cloud accuracy assessment method without feature extraction according to claim 1 is characterized in that: The original point cloud is collected by the mobile measurement system, and control points are arranged in the point cloud measurement area. With checkpoints ; The i-th control point , 、 They represent the plane coordinates of the control points under Gaussian three-dimensional projection, Indicates the elevation of the control point relative to the quasi-geoid; the i-th checkpoint , 、 They represent the plane coordinates of the checkpoint under Gaussian three-dimensional projection, Indicates the elevation of the checkpoint relative to the quasi-geoid; the elevations of the control point and the checkpoint are both obtained through leveling, and both the control point and the checkpoint are located on the road surface in the survey area.
3. The sparse point cloud accuracy assessment method without feature extraction according to claim 1, characterized in that: ,in, For ground point cloud The maximum value in the direction, For ground point cloud The minimum value in the direction, is the maximum value of the ground point cloud in the Y direction, It is the minimum value of the ground point cloud in the Y direction.
4. The sparse point cloud accuracy assessment method without feature extraction according to claim 1, characterized in that: The step three is as follows: Construct KD-Tree for ground point cloud; Traverse each checkpoint , according to its plane coordinates , search the ground point cloud for the checkpoint Planar distance is less than the local neighborhood radius All points as the current checkpoint Neighborhood point set ; ； is a single ground point in the ground point cloud, represents the ground point cloud, Represents the three-dimensional coordinates of a single ground point in the ground point cloud; Calculate the median elevation of the neighborhood points as the measured elevation of the checkpoint in the ground point cloud; ； The elevation of all checkpoints in the ground point cloud is composed into vectors , For checkpoints The measured elevation in the ground point cloud; at the same time, the original leveling elevation of all checkpoints is recorded as .
5. The sparse point cloud accuracy assessment method without feature extraction according to claim 1, characterized in that: The step 4 is specifically as follows: Use the same method to obtain the elevation of the checkpoint to obtain the elevation of the control point in the ground point cloud : ； The corresponding vector of the original leveling elevation of the control point is ; Calculate the original height difference vector of the control point : ； right Remove outliers and obtain the cleaned height difference set ;in, is the number of height differences in the height difference set after cleaning; calculate The arithmetic mean of As system height difference: ； Correct the system error of the original level elevation of the checkpoint to obtain the checkpoint elevation consistent with the ground point cloud benchmark .
6. The sparse point cloud accuracy assessment method without feature extraction according to claim 5, characterized in that: The outliers are removed by using the triple standard deviation method; the median of the original height difference vector is calculated; , and all and The absolute mean error between ; Eliminate all satisfaction The original height difference sample points are recorded as the cleaned height difference set. .
7. The sparse point cloud accuracy assessment method without feature extraction according to claim 6, characterized in that: The absolute mean error calculation formula is as follows: ； in, Represents calculation of the median.