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

The Importance of Structural Inspections and the Introduction of 3D Scanning

Structures that support social infrastructure—such as bridges, tunnels, retaining walls, and buildings—require regular inspection and maintenance due to aging. Neglecting inspections can allow small degradations to be overlooked, potentially leading to major accidents. For example, early detection and repair of concrete spalling or cracking, and corrosion or deformation of steel members, contribute to prolonging service life and ensuring safety. For this reason, national and local governments have been strengthening statutory inspections and preventive maintenance efforts, and the importance of structural inspection continues to grow year by year.

However, traditional inspection methods have relied mainly on human visual inspection, which requires substantial preparation and effort such as erecting scaffolding or using aerial work platforms. In addition, inspection results tend to depend on the inspector's subjective judgment, making it difficult to record quantitative data or track deterioration over time.

Against this background, the adoption of 3D scanning technology has attracted attention in recent years. Using 3D scanning—such as laser scanners, smartphone-mounted LiDAR, or drone photogrammetry—makes it possible to acquire the shape of a structure as high-density point cloud data. Recording structures as 3D point clouds captures overall dimensions and displacements accurately and allows detailed post hoc analysis, enabling objective and reproducible inspections. Within initiatives like the Ministry of Land, Infrastructure, Transport and Tourism’s promotion of i-Construction and the trend toward CIM, 3D scanning is becoming a foundational technology supporting the DX (digital transformation) of infrastructure maintenance.

Noise Issues in Point Cloud Data and Their Impact

Point cloud data acquired by 3D scanning always contain noise arising from sensor errors or environmental influences. Noise refers to unnecessary points scattered at positions where they should not exist, or points displaced from the actual surface. For example, in laser scanning, measurement errors at long distances or on low-reflectivity surfaces can create a “cloud” of points floating off a wall. In photogrammetry, inaccurate points can appear when reproducing textureless surfaces or glass. Furthermore, alignment errors when merging multiple scan datasets can also introduce noise-like variability.

If a point cloud retains a large amount of noise, it undermines subsequent analysis and utilization in various ways. The main problems caused by noisy point clouds are as follows:

• Reduced measurement accuracy: When measuring distances, areas, or displacements on a point cloud, noise introduces larger errors than the true values. For example, even when scanning a flat bridge girder, surface scatter caused by noise can make an originally smooth cross-section appear jagged, preventing accurate calculation of deflection or cross-sectional dimensions.

• Shape misrecognition: A high density of noise points makes it difficult to intuitively grasp the structure’s shape from the point cloud. Points floating where nothing should exist can be mistaken for cracks or defects. Especially on seemingly smooth surfaces like tunnels or retaining walls, noise can give the false impression of irregularities and lead to incorrect deterioration assessments.

• Interference with analysis processes: Leaving noise in the data increases the time and reduces the quality of processes such as point cloud meshing (polygon modeling) and feature extraction (edge detection, plane detection). During mesh generation, scattered outliers can cause unnecessary fine mesh subdivisions or create holes. Likewise, when comparing point clouds (for displacement measurement), heavy noise makes it difficult to distinguish real 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 some form of noise removal or reduction. Various filter processes are used for this purpose, and among common smoothing methods the Gaussian filter is especially noteworthy.

What Is a Gaussian Filter: Mathematical Basis and Purpose

A Gaussian filter 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 values decrease exponentially with distance. The basic principle of the Gaussian filter is to weight nearby data more strongly and distant data more weakly when averaging a point’s neighborhood. In one-dimensional signals and image processing, it has long been known as “Gaussian blur” and is widely used to softly average pixels for noise removal. Compared to a simple mean filter, the Gaussian filter excels at smoothing while preserving neighborhood information, reducing only high-frequency noise without excessively blurring the entire image, making it a widely used smoothing filter.

Mathematically, the Gaussian filter weight is proportional to $\exp(-(d^2)/(2\sigma^2))$ with respect to distance d (where $\sigma$ is the standard deviation of the Gaussian distribution and a parameter that controls the filter’s spread). The weight approaches 1 when the distance is near zero and decreases exponentially as distance grows. By using these weights to compute a weighted average with neighboring points, the influence of outliers and random errors is reduced and values are replaced by smoother ones. Also, since the Gaussian filter is a linear filter and the convolution of a Gaussian with itself is also Gaussian, applying the filter twice is equivalent to a single stronger application—so iterative applications can achieve progressive smoothing with theoretically similar effects as a single, appropriately tuned filter.

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 canceled out by averaging, while the actual structure’s shape (low-frequency components) is largely preserved. For these reasons, the Gaussian filter is an optimal preprocessing step before generating meshes or extracting dimensions from point clouds; by effectively removing noise it realizes high-precision point cloud data.

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

Next, let’s look at procedures and key points when actually applying a Gaussian filter to point cloud data. The basic flow is to smooth by “computing a weighted average with neighboring points for each point and moving the point to a new position.” At this time, setting the filter parameters is crucial.

• Setting the neighborhood range: First, decide “which range of points to consider as neighbors” for a given point. One method uses a fixed radius and treats points within that spherical range as neighbors; another uses a fixed number of nearest neighbors (k-nearest neighbors: kNN) regardless of distance. If the chosen neighborhood is too wide, points from other, distant structures may be included in the average and blur the shape. Conversely, if the neighborhood is too narrow, sufficient averaging may not occur and noise will remain. It is important to select an appropriate neighborhood range according to the point density and object size of the point cloud.

• Adjusting the weight function (σ value): As mentioned, the Gaussian filter’s weights are determined by the standard deviation $\sigma$. A larger $\sigma$ gives considerable weight to more distant points, resulting in strong smoothing (blurring effect). A smaller $\sigma$ mainly relies on very nearby points and yields weak smoothing. Generally, set $\sigma$ to match the magnitude of the point cloud’s measurement error. For example, if range error is around ±5 mm, choose a smoothing level that roughly removes noise within that range. It is advisable to test the filter and identify a $\sigma$ that sufficiently reduces noise without compromising shape.

• Local processing and iterative application: The Gaussian filter is fundamentally computed within each point’s local neighborhood as a local operation, so the computation is easily parallelized and does not affect the overall macro shape of the point cloud. If needed, the filter can be applied multiple times to progressively attenuate noise. However, applying it too often can round off true edges (corners), so the number of iterations and their strength should be adjusted according to the objective.

• Care for shape continuity: Smoothing will make surfaces reproduced from the point cloud more continuous and smoother. Bumpy slopes or tunnel inner surfaces will have fine irregularities averaged out after Gaussian filtering, improving surface continuity. However, very fine discontinuities such as cracks may also be obscured, so when the goal is damage detection you must carefully consider where and how strongly to apply the filter. On the other hand, if the aim is to capture overall shape or obtain dimensional measurements, a slightly rounded edge in exchange for a noise-free continuous surface can improve accuracy. It is essential to balance shape continuity and local detail by adjusting the degree of filtering case by case.

Before applying a Gaussian filter, it is efficient to remove obvious outliers in advance. Statistically isolated points (clearly distant points with no neighbors) should be removed using statistical outlier removal filters (statistical filters) or radius-based removal, which are implemented in many point cloud processing tools. Performing such outlier removal prior to Gaussian smoothing prevents extreme anomalies from skewing the averages and yields more stable noise reduction.

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

Point cloud noise removal and smoothing by Gaussian filters have demonstrated effectiveness across inspection data for various structures. Here we present application examples and effects for typical civil engineering structures—bridges, slopes, tunnels—and for buildings in the architectural field.

Applications to Bridges

Bridge inspections (road and railway bridges) involve many members such as girders, piers, and suspenders in complex combinations. Bridge point clouds obtained by laser scanning or photogrammetry can contain noise from long-range measurements or reflections on metal surfaces. Applying a Gaussian filter suppresses such unwanted points through averaging and reproduces member surfaces smoothly. For example, in measuring deflection of steel girders, noisy point clouds do not yield accurate results, but after filtering the cross-sectional profile becomes clear and millimeter-level displacement measurements become feasible. Also, slope and pavement surface irregularity measurements can obtain accurate longitudinal and transverse profiles from smoothed point clouds, improving detection accuracy for steps and distortions.

While high-resolution images are often used for detecting micro-cracks in bridge inspections, Gaussian-filtered point clouds provide data suitable for overall shape understanding and deformation measurement.

Applications to Slopes

For slope and retaining wall monitoring, laser scanners and UAV laser surveys are used to capture signs of landslides or surface changes. Acquired slope point clouds can show scatter due to vegetation or reduced accuracy from oblique measurement angles. Applying a Gaussian filter removes these noisy points and produces a smooth terrain model of the slope. For example, small surface irregularities (like erosion grooves caused by rain) can be hard to distinguish from noise, but after smoothing the overall terrain undulation becomes clearer and subtle bulging or subsidence that is difficult to notice manually can be quantitatively evaluated.

Also, when comparing point clouds from multiple epochs to analyze slope deformation differences, applying comparable filtering to each epoch increases the reliability of the differential analysis. By removing unnecessary points while capturing terrain trends, the Gaussian filter is a useful tool for slope management.

Applications to Tunnels

3D scanning inside tunnels can investigate inner wall deformation, cracks, and spalling of lining concrete. Tunnel point clouds tend to have relatively little noise because they are enclosed spaces, but small errors still occur due to laser incidence angles or lighting conditions. Gaussian filtering smooths tunnel inner surface point clouds, bringing the tunnel cross-section closer to an ideal circle (or horseshoe shape). This enables high-precision calculation of differences from the design cross-section to quantify sectional contraction (deformation). Especially when monitoring long-term displacement of lining concrete, radii measured from filtered point clouds show less fluctuation and improved reliability. The smoother tunnel wall also facilitates registration when mapping photographs of cracks or spalling.

Conversely, to prevent smoothing from erasing actual concavities such as spalling, filter strength must be controlled. Overall, in tunnel inspections Gaussian filters are expected to improve precise cross-sectional characterization and accuracy of deformation analysis.

Applications to Buildings

In architecture, 3D scanning of building facades and historic structures is becoming widespread. Noise removal is useful for evaluating wall smoothness, column plumbness, and floor flatness in building point clouds. For example, when scanning a concrete facade, fine surface roughness and point omissions from sensor misses can occur; Gaussian filtering can compensate for these and yield a smoother point cloud close to the actual wall surface. As a result, accurate verticality and bulging measurements can be taken from an undistorted wall model, aiding diagnosis such as detecting tile delamination or structural tilt.

In indoor laser surveys, surfaces such as floors, walls, and ceilings can remain rough due to furniture even after removal. Using filtered point clouds enables high-precision detection of floor deflection and uniformity checks of ceiling heights. For historic buildings, where distinguishing intentional ornamentation from age-related distortion is necessary, reducing noise helps to discern true shape differences. Thus, Gaussian filters in building point clouds contribute to improved reliability of dimensional measurements and clarity in shape understanding.

Effects on Accuracy Verification, Error Reduction, and Mesh Generation

Point clouds treated with a Gaussian filter show clear accuracy benefits compared to untreated data. To verify measurement accuracy, experiments applying filters to point clouds of known shapes are common. For example, scanning a flat reference plate and statistically analyzing each point’s deviation from the reference plane may show a standard deviation of several millimeters before filtering, which can drop to less than half after filtering. This is one example that quantitatively demonstrates the error reduction effect.

Also, when comparing point clouds to compute differences (for displacement measurement or deformation detection), filtered data reduce the influence of extraneous points in the difference results, making changes easier to evaluate.

However, excessive filtering can cause adverse effects. Large deformations that are truly present (for example, open structural cracks) may be averaged out by overly strong smoothing. Therefore, after applying the filter it is desirable to compare with the original data or known measurements at key locations to ensure important features have not been lost. Fortunately, since Gaussian filters can be finely tuned by parameters, it is possible to balance required accuracy and detail preservation while reducing errors.

Gaussian filters also have major effects when generating meshes from point clouds. When converting point clouds into triangular polygons, a noisy input will generate countless tiny triangles and a rough model. With filtered point clouds, surfaces are smoother, enabling high-accuracy meshes with fewer polygons. Advanced algorithms like Poisson surface reconstruction also perform more stably and produce cleaner, hole-free meshes when the input point cloud has low noise.

Especially when integrating multiple scan datasets before meshing, slight misalignments between scans can be alleviated by smoothing, enabling creation of a seamless integrated model. Combined with downsampling, filter processing can reduce point cloud volume while preserving shape accuracy. Properly thinning the point cloud after noise removal not only speeds up processing but, by omitting unnecessary fine details during mesh generation, can ultimately improve model accuracy.

As described, noise reduction via Gaussian filtering enhances measurement accuracy and downstream processing quality of point cloud data, producing reliable analysis results. In infrastructure management that leverages point clouds, using Gaussian filtering alongside accuracy verification and error evaluation processes enables quantitative, high-value inspection data utilization.

Outlook: Cloud Utilization and Integration with AR/BIM

In recent years, there has been growing activity in using high-resolution point cloud data in the cloud and integrating it with AR (augmented reality) and BIM. Point clouds cleaned by Gaussian filtering are highly compatible with these advanced technologies and have the potential to further accelerate DX in maintenance. The following points outline future prospects.

• Data sharing and analysis via cloud services: Large point cloud datasets can be easily shared among stakeholders by uploading them to cloud platforms. If scans are synchronized to the cloud immediately after capture, point clouds can be reviewed and analyzed from the office. Cleaned point clouds render smoothly in cloud viewers, enabling accurate online measurements of distances, areas, and volumes. By leveraging cloud computing power, advanced analyses such as AI-based automatic damage detection and temporal difference extraction can be executed. Integrating field scanning with the cloud enables a faster and more efficient inspection data utilization cycle.

• Fusion with AR technologies: Overlaying point clouds or derived 3D models onto real-world scenes via AR allows intuitive on-site understanding and instruction. For example, displaying a high-precision 3D model produced after filtering on a tablet overlaid on the actual structure lets inspectors compare degradation locations and deformation amounts directly. Marking detected crack locations in AR to share with repair teams or, in the future, using smart glasses to overlay point cloud-derived information (such as highlighting areas needing repair) in the inspector’s field of view are conceivable. Low-noise point cloud models have higher alignment accuracy with the real world, maximizing AR support effectiveness.

• Integration with BIM: BIM usage in maintenance is progressing, and cases of comparing acquired point clouds with existing BIM models to detect changes or generating BIM models from point clouds are increasing. Filtered point clouds, with clear shapes unconfused by noise, are useful for BIM integration. For example, scanning a constructed structure and comparing it to the design BIM to verify as-built conditions is easier when the point cloud is not noisy, since setting error thresholds is less problematic with a smoothed point cloud. When building digital twins of existing infrastructure assets, lower-noise point clouds enable more precise modeling and enhance subsequent simulations and monitoring. In the future, integrating point cloud + BIM + sensor data into a real-time digital twin environment could enable continuous health monitoring of structures.

Conclusion: Use with LRTK for Simple 3D Surveying and High-Precision Scanning

This article explained the significance of 3D scanning for structural inspections and the effectiveness of Gaussian filter processing for point cloud noise removal and precision improvement. Removing noise from high-density point clouds to obtain accurate models and measurements directly improves the reliability of infrastructure maintenance. Gaussian filtering is a powerful means confirmed useful across a wide range of structures from bridges and tunnels to buildings.

Recently, technologies like LRTK have appeared as solutions that enable anyone to easily perform high-precision 3D scans. LRTK combines high-precision positioning (RTK-GNSS) with 3D scanning, enabling simple 3D surveying using smartphones or small devices. Using dedicated hardware and apps and following an operation procedure that can be learned with short training, anyone can walk a site and easily acquire point clouds with absolute coordinates. Because the acquired point clouds already have high-precision positional information from the outset, LRTK significantly reduces post-processing alignment work.

By using LRTK together with Gaussian filtering, inspection sites can perform high-precision scans quickly and immediately analyze and share the results. For example, one could scan a bridge pier on site, apply filtering to the point cloud to compute displacement, and send the results to the office via the cloud for immediate expert judgment—enabling remote collaboration.

Going forward, digital solutions that combine 3D scanning, point cloud processing, and cloud/AR technologies will become indispensable in infrastructure inspection and maintenance. By actively adopting advanced point cloud processing techniques such as Gaussian filtering and innovative measurement tools like LRTK, we can achieve safe, efficient, and highly accurate maintenance management.