3D Scanning for Structural Inspections: Point Cloud Noise Removal and High-Precision Enhancement with Gaussian Filtering

Importance of Structural Inspections and the Background of 3D Scan Adoption

Structures that support social infrastructure—such as bridges, tunnels, retaining walls, and buildings—require regular inspection and maintenance due to aging. Neglecting inspections can lead to small degradations being overlooked and potentially causing large-scale accidents. For example, early detection and repair of concrete delamination and cracking, or corrosion and deformation of steel members, contributes to extending service life and ensuring safety. For this reason, national and local governments have strengthened statutory inspections and preventive maintenance efforts, and the importance of structural inspections has been growing year by year.

However, traditional inspection methods have mainly relied on human visual checks, requiring extensive preparation and effort such as erecting scaffolding or using aerial work platforms. Inspection results are also prone to operator subjectivity, making it difficult to quantitatively record conditions or track degradation over time.

Against this backdrop, the recent adoption of 3D scanning technology has attracted attention. Using 3D scanning methods such as laser scanners, smartphone-mounted LiDAR, or drone photogrammetry, the shape of structures can be captured as high-density point cloud data. Point cloud records accurately capture overall dimensions and displacements of structures and can be analyzed in detail afterward, enabling objective and reproducible inspections. Within initiatives like i-Construction and CIM promoted by the Ministry of Land, Infrastructure, Transport and Tourism, 3D scan utilization is becoming a foundational technology that supports the DX (digital transformation) of infrastructure maintenance.

Noise Issues in Point Cloud Data and Their Impact

Point cloud data acquired by 3D scanning inevitably contain noise caused by sensor errors and environmental influences. Noise refers to unnecessary points scattered at positions where no real surface exists or points displaced from the actual surface. For example, laser scanners can produce measurement errors at long distances or on low-reflectivity surfaces, creating a “fog” of points floating away from a wall surface. In photogrammetry, inaccurate points may appear when reproducing surfaces without texture or glass. Additionally, misalignments when merging multiple scans can cause noise-like variability.

If many noise points remain in a point cloud, they hinder subsequent analyses and utilization in various ways. Major problems caused by noisy point clouds include the following.

• Decreased measurement accuracy: When measuring distances, areas, or displacements on a point cloud, noise introduces larger errors than actually exist. For example, when scanning a flat bridge girder, scattered noise on the surface can make originally smooth lines appear jagged when evaluating cross-sectional shapes, preventing accurate calculation of deflection or section dimensions.

• Misinterpretation of shape: High numbers of noise points make it difficult to intuitively grasp the structure’s shape from the point cloud. Points floating at positions where none should exist can be mistaken for cracks or defects. Especially on seemingly smooth surfaces such as tunnels or retaining walls, noise-induced apparent irregularities can lead to erroneous degradation judgments.

• Interference with analysis processing: When noise remains, processes such as point cloud meshing (polygon modeling) or feature extraction (edge detection or plane detection) take longer and yield lower-quality results. During mesh generation, stray outlier points can create unnecessary subdivided meshes or cause holes. When comparing point clouds (displacement measurement), large noise makes it difficult to distinguish true changes from noise-induced false differences.

As described above, noise undermines the accuracy and reliability of point cloud data usage, so it is important to perform noise removal or reduction by some method. Various filtering techniques are used for this purpose, and among typical smoothing methods the Gaussian filter is particularly notable.

What Is a Gaussian Filter: Mathematical Basis and Application Purpose

A Gaussian filter (Gaussian filter) is a filtering process that uses weights based on the Gaussian (normal) distribution. The Gaussian distribution is represented by a bell-shaped curve, where values near the center have higher probability (weight) and those farther away decrease exponentially. The basic principle of a Gaussian filter is to average using weights so that data in a point’s neighborhood contribute strongly while distant data contribute weakly. In one-dimensional signals and image processing, it has long been known as “Gaussian blur” and is commonly used to softly average pixels for noise removal. Compared to a mean filter, the Gaussian filter excels at smoothing while preserving local information, attenuating only high-frequency noise without unnecessarily blurring the entire image—hence its widespread use as a smoothing filter.

Mathematically, the Gaussian filter weight is proportional to exp(-(d^2)/(2\\sigma^2)) with respect to distance d (\\sigma is the standard deviation of the Gaussian distribution and a parameter that adjusts the filter’s spread). The weight approaches 1 as the distance approaches zero, and decreases exponentially as distance increases. Using this weighting to compute a weighted average with surrounding points reduces the influence of outliers and random errors and replaces values with smoother ones. The Gaussian filter is a type of linear filter and has the property that applying the filter twice is equivalent to applying a stronger single filter once (the autocorrelation of a Gaussian is also Gaussian), so repeated application can achieve in steps the same effect as a single stronger pass.

Applying a Gaussian filter to point cloud data serves to average out small sensor-derived measurement errors and bring the data closer to the true surface shape. Random noise components are largely canceled by averaging, while the actual structural shape (low-frequency components) is preserved as much as possible. For these reasons, the Gaussian filter is an ideal preprocessing step before extracting meshes or dimensions from point clouds, and by effectively removing noise it achieves high-precision point cloud data.

Practical Point Cloud Processing with Gaussian Filtering (Filter Parameters, Local Processing, Continuity)

Let us look at procedures and key points when actually applying a Gaussian filter to point cloud data. The basic flow is to “compute a weighted average with neighboring points for each point and move the point to the new position” to smooth the cloud. Setting filter parameters is critical in this process.

• Setting the neighborhood range: First, decide “which range of points will be considered neighbors” for a given point. One method is to set a fixed radius and consider points within that spherical range as neighbors; another is to target a fixed number of nearest neighbors regardless of distance (k-nearest neighbors: kNN). If the chosen neighborhood is too large, points from other, distant structures may be included in the averaging, blurring the shape. If the neighborhood is too small, sufficient averaging cannot be achieved and noise remains. Select an appropriate neighborhood range according to the point density and the size of the structures in the point cloud.

• Adjusting the weight function (σ value): As noted earlier, the Gaussian filter weights are determined by the standard deviation \\sigma. A larger \\sigma gives relatively large weights to farther points, resulting in strong smoothing (blurring). Conversely, a smaller \\sigma makes only very close points contribute significantly, resulting in weak smoothing. In general, set \\sigma to match the magnitude of point cloud errors. For example, if ranging error is about ±5 mm (±0.20 in), choose a smoothing level that approximately removes noise of that magnitude. It is advisable to experimentally apply the filter and determine a \\sigma that sufficiently reduces noise without degrading the shape.

• Local processing and repeated application: The Gaussian filter is essentially a local processing method where calculations are completed within each point’s local neighborhood. Therefore, computations are easily parallelizable and do not affect the overall macroscopic shape of the point cloud. By repeating the filter several times, noise can be attenuated step by step. However, excessive repetition can round off true edges and corners, so the number of iterations and intensity should be adjusted according to the objective.

• Consideration for shape continuity: Smoothing improves the continuity and smoothness of surfaces reproduced from point clouds. Irregular slope faces or tunnel inner surfaces will have fine irregularities averaged out after a Gaussian filter, enhancing surface continuity. However, very fine discontinuities such as cracks may also be obscured, so when the goal is damage detection one must carefully consider where and how strongly to apply the filter. Conversely, if the objective is overall shape understanding or dimensional measurement, a somewhat rounded edge that yields a noise-free continuous surface can improve measurement accuracy. It is important to adjust the degree of filtering case by case to balance shape continuity and local detail.

As a preliminary step before applying a Gaussian filter, it is efficient to remove obvious outliers in advance. Isolated points that are statistically alone (with no neighboring points) can be removed by statistical outlier removal filters (statistical filters) or radius-based removal, which are implemented in many point cloud processing software packages. Performing such outlier removal before Gaussian smoothing prevents the average from being pulled by extreme anomalous points and yields a more stable noise reduction effect.

Application Examples for Civil Structures (Bridges, Slopes, Tunnels) and Building Structures

Noise removal and smoothing of point clouds using Gaussian filters have proven effective in inspection data for various structures. Below are application examples and their effects for representative civil structures—bridges, slopes, tunnels—and for buildings in the architectural field.

Application to Bridges

Bridge inspections (road bridges and railway bridges) involve many components such as girders, piers, and suspension elements assembled in complex ways. Point clouds acquired by laser scanning or photogrammetry for bridges can contain noise due to long-distance measurement errors or reflections from metallic surfaces. Applying a Gaussian filter suppresses such unwanted points by averaging and reproduces smoother component surfaces. For instance, deflection measurement of steel girders cannot achieve high accuracy with noisy point clouds, but after filtering the cross-sectional profile becomes clear and millimeter-level displacement measurements become feasible. For bridge decks and pavement irregularity measurements, smoothed point clouds enable accurate longitudinal and transverse profiles to be obtained, improving detection accuracy for steps and distortions.

While high-resolution imagery is often used to detect micro-cracks in bridge inspections, Gaussian-filtered point clouds provide data well-suited for overall shape understanding and deformation measurement.

Application to Slopes

For monitoring slopes and retaining walls, laser scanners and UAV LiDAR are used to detect signs of landslides and surface changes. Acquired slope point clouds can show point variability due to vegetation or decreased accuracy from oblique scanning angles. Applying a Gaussian filter removes such noisy points and yields a smooth terrain model for the slope. For example, micro-unevenness on a slope surface (such as erosion gullies from rainfall) can be difficult to distinguish from noise, but after smoothing the larger-scale terrain undulations become clearer, making it easier to quantitatively evaluate bulging or settlement that may be hard to notice manually.

When comparing point clouds from multiple epochs to analyze slope deformation differences, applying equivalent filtering to each epoch’s point cloud increases the reliability of difference analysis. The Gaussian filter, which removes unnecessary points while capturing terrain trends, is therefore a useful tool for slope management.

Application to Tunnels

3D scanning inside tunnels can investigate inner wall deformations, cracks, or delamination of lining concrete. Tunnel point clouds typically have less noise due to the enclosed space, but small errors still occur depending on laser incidence angles or lighting. Gaussian filtering smooths tunnel inner surface point clouds so that tunnel cross-sections approach ideal circular (or horseshoe) shapes. This enables high-precision calculation of the difference from the design section to determine the degree of sectional contraction (deformation). Especially when monitoring long-term displacements of lining concrete, radius measurements from filtered point clouds have lower fluctuation and higher reliability. Improved wall surface planarity also facilitates alignment when mapping crack or delamination images.

However, filter strength should be adjusted to avoid smoothing away actual irregularities such as delamination. Overall, Gaussian filtering in tunnel inspections is expected to enable precise understanding of sectional shapes and improved accuracy of deformation analysis.

Application to Buildings

In architecture, 3D scanning of building façades and historic structures is becoming widespread. Noise removal benefits building point clouds when evaluating wall smoothness, column plumbness, and floor flatness. For example, when scanning a concrete building façade, fine surface roughness or missing points due to sensor dropouts may occur, but Gaussian filtering can compensate for these, resulting in a smoother point cloud closer to the actual wall surface. As a result, accurate verticality and bulge measurements can be taken from undistorted wall models, assisting diagnostics such as detecting tiles detaching or structural tilting.

Indoor laser scans can also leave floor, wall, and ceiling surfaces rough with noise even after furniture removal. Using filtered point clouds allows high-precision detection of floor sagging or checking ceiling height uniformity. For historical buildings, where distinguishing intentional ornamentation from aging distortion is necessary, reducing noise makes it easier to identify differences from the true shape. Thus, Gaussian filtering in building point clouds contributes to improved reliability of dimensional measurements and clarified shape understanding.

Effects on Point Cloud Accuracy Verification, Error Reduction, and Mesh Generation

Point clouds processed with Gaussian filters offer clear advantages in terms of accuracy compared to unfiltered data. For accuracy verification, experiments applying the filter to point clouds scanned from known shapes are commonly performed. For example, when scanning a flat reference plate, statistical analysis of each point’s deviation from the reference plane may show variations on the order of ± a few millimeters (standard deviation) before filtering, whereas after filtering the standard deviation can decrease to less than half. This provides a quantitative example of the error reduction effect.

When comparing point clouds to compute differences (for structural displacement measurement or defect detection), filtered data reduces the influence of unwanted points in the difference results, allowing clearer evaluation of the amount of change.

On the other hand, be aware of the adverse effects of excessive filtering. Significant deviations from the original shape (for example, areas where structural cracks are open) can be averaged out by overly aggressive smoothing. Therefore, after filtering it is desirable to compare key areas with the original data or known measurement values to ensure critical features have not been lost. Fortunately, the Gaussian filter’s effects can be finely tuned via parameters, enabling error reduction while balancing required accuracy and detail preservation.

Gaussian filtering also has major benefits when generating meshes from point clouds. When converting point clouds into collections of triangular polygons, many small triangles and a rough model result if the input is noisy. A filtered point cloud produces smoother surfaces, allowing high-accuracy meshes with a controlled polygon count. Advanced algorithms like Poisson surface reconstruction perform more stably and produce cleaner meshes with fewer holes when input point clouds are less noisy.

Especially when integrating multiple scan datasets before meshing, small misalignments between scans can be mitigated by smoothing, enabling generation of a seamless integrated model. Combined with downsampling, filter processing can reduce point cloud size while maintaining shape accuracy. Appropriately thinning point density after noise removal not only shortens processing time by reducing data size, but also omits unnecessary fine details during mesh generation, resulting in overall model accuracy improvement.

Thus, Gaussian-filtered noise reduction enhances both measurement accuracy and post-processing quality of point clouds, yielding reliable analysis results. In infrastructure management using point clouds, applying Gaussian filters alongside accuracy verification and error assessment processes enables quantitative and high-confidence use of inspection data.

Prospects for Cloud Utilization and AR/BIM Integration

Recently, efforts to utilize high-resolution point cloud data on the cloud and to integrate it with AR (augmented reality) and BIM have been accelerating. Point clouds cleaned up by Gaussian filters have high affinity with such advanced technologies and hold potential to further drive DX in maintenance. The following points outline future prospects.

• Data sharing and analysis via cloud services: Large point cloud datasets can be uploaded to cloud platforms for easy sharing among stakeholders. Syncing scans to the cloud immediately from the field enables point cloud review and analysis from the office. Noise-reduced point clouds can be displayed smoothly in cloud viewers, and online analyses such as distance, area, and volume calculations can be performed accurately. Leveraging cloud computational power also opens the possibility of running advanced analyses like AI-based automated damage detection or time-differenced point cloud extraction. Linking field operations with the cloud will make the cycle of point cloud data utilization faster and more efficient.

• Fusion with AR technologies: Overlaying point clouds or derived 3D models onto the real world with AR enables intuitive on-site situational awareness and instructions. For example, displaying a high-precision 3D model obtained after filtering on a tablet over the actual structure allows comparison of degraded areas and deformation magnitudes with the real object. Marking detected crack positions in AR to share with repair teams or, in the future, projecting information derived from point clouds (such as highlighting areas requiring repair) onto an inspector’s view through smart glasses could be possible. Low-noise point cloud models improve registration accuracy with the real world, maximizing AR-assisted benefits.

• Integration with BIM: BIM (Building Information Modeling) is increasingly used in maintenance, and more cases are appearing where acquired point clouds are compared with existing BIM models to detect changes, or conversely used to generate BIM models of existing structures. Filtered point clouds with clear shapes and minimal noise are valuable for BIM integration. For example, when scanning a structure after construction and comparing it to the design BIM for as-built verification, noisy point clouds complicate setting error thresholds, but smoothed point clouds make it easier to clearly identify error distributions. When constructing digital twins of existing infrastructure assets, lower noise in point clouds enables more precise modeling, enhancing subsequent simulation and monitoring. In the future, it will be feasible to realize advanced maintenance such as real-time digital twin environments that fuse point clouds, BIM, and sensor data to continuously monitor structural health.

Conclusion: Use of LRTK for Easy 3D Surveying and High-Precision Scanning

This article explained the significance of 3D scanning for structural inspections and the effects of Gaussian filtering for point cloud noise removal and high-precision enhancement. Removing noise from high-density point clouds to obtain precise models and measurements directly contributes to improving the reliability of infrastructure maintenance. The Gaussian filter is a powerful tool whose usefulness has been confirmed across a wide range of structures from bridges and tunnels to buildings.

Recently, technologies like LRTK have emerged as solutions that allow anyone to perform high-precision 3D scans easily. LRTK combines high-precision positioning technology (RTK-GNSS) with 3D scanning to realize simple 3D surveying using smartphones or compact devices. Using dedicated equipment and apps and following operation procedures that can be learned with short training, anyone can readily acquire point clouds with absolute coordinates simply by walking on site. Because the acquired point clouds already contain high-precision positional information from the initial positioning stage, LRTK greatly reduces alignment work during post-processing.

By leveraging LRTK together with Gaussian filtering for noise removal, it is possible to perform high-precision scans quickly at inspection sites and immediately analyze and share results. For example, one could scan a bridge pier’s tilt on site, apply a filter to the point cloud to compute displacement, and send the results via the cloud to office specialists for instant assessment—enabling remote collaboration.

In the future, digital solutions that combine 3D scanning and point cloud processing with cloud, AR, and other technologies will become indispensable in infrastructure inspection and maintenance. By proactively adopting advanced point cloud processing techniques such as Gaussian filters and innovative measurement tools like LRTK, we can achieve safer, more efficient, and higher-precision maintenance.