Point Cloud Noise Removal: Gaussian Smoothing Explained + Tips | LRTK

Basics of Point Cloud Processing and the Problem of Noise

Point cloud data acquired by 3D scanners, LiDAR, photogrammetry, and the like record real-world coordinates as a collection of many points. While point clouds contain geometric information about targets, they always include unavoidable noise introduced during acquisition. Consider, for example, surveying terrain with a laser scanner. Because of surface reflectance properties, the sensor’s own measurement errors, and environmental factors (rain, dust, sunlight interference, etc.), points can be recorded at positions that deviate from the actual terrain surface. When scanning structures, laser scattering at steel edges or positional fluctuations in areas with poor GPS reception can introduce inaccurate points—sometimes called “shadow points.”

As acquired, point clouds contain noise points and outliers of various sizes, and if left unprocessed they can adversely affect subsequent analysis and modeling accuracy. For example, trying to estimate a surface from a noisy point cloud can produce locally anomalous bumps and depressions, making accurate volume calculations or shape determinations difficult. In construction site quality control, designs are compared to acquired point clouds to evaluate construction accuracy; if the point cloud contains noise, the errors may be large and lead to false judgments. Therefore, noise removal and smoothing are very important steps in point cloud processing.

Why Noise Removal and Smoothing Are Needed

There are several clear benefits to removing noise and smoothing point cloud data.

• Reduced measurement error: By averaging out sensor measurement variability (random noise), each point’s position is brought closer to the true geometry. This leads to an overall improvement in point cloud accuracy.

• Improved shape discriminability: A denoised point cloud represents planes and surfaces smoothly and clarifies edges and boundaries. This makes it easier to understand terrain undulations or structural shapes both visually and analytically. For instance, slope gradients or road undulations can be evaluated more accurately on smoothed data than on noisy data.

• Reliability in as-built verification: On construction sites, differences between design surfaces and as-built point clouds are used to compute fill/cut volumes and verify as-built dimensions. Excessive noise can cause difference calculations to pick up trivial deviations, leading to unnecessary rework decisions. A smoothed point cloud makes it easier to capture real deviations and improves the reliability of as-built verification.

• Increased processing efficiency: Removing obvious outliers (points that are extremely distant) reduces the load on subsequent processing. Cleaned point clouds are smaller than noisy raw data, shortening processing time on PCs or in the cloud.

For these reasons, it is recommended to apply appropriate filters to remove noise and smooth point clouds after data acquisition. Below, we focus on the Gaussian filter as a method for smoothing point cloud data.

What Is a Gaussian Filter

A Gaussian filter is a filter that smooths data using weights based on the Gaussian (normal) distribution. Originally common in image processing—often called “Gaussian blur”—a Gaussian filter averages surrounding pixel values with weights that follow a Gaussian distribution. The center pixel receives the largest weight and the weight decreases exponentially with distance, so the filter reduces noise while preserving edges to some extent. Compared to a simple uniform average, a Gaussian filter better preserves the original shape while reducing random noise (with high effectiveness against Gaussian-like noise, rather than impulse noise).

Extending this Gaussian weighting concept to 3D point clouds yields Gaussian-filter-based point cloud smoothing. Unlike images, point clouds lack a regular pixel grid, so for each point you first gather neighboring points and compute a Gaussian-weighted average over that set. In a Gaussian filter, closer points receive larger weights and distant points smaller weights when computing the averaged coordinates. This smooths local irregularities and produces a smoother point cloud representation.

The Gaussian function’s standard deviation σ (sigma) is the parameter that determines the filter’s strength: larger σ blurs over a wider area (stronger smoothing). However, if σ is too large, finer details of the original shape (edges and small undulations) are lost, so selecting an appropriate value is important.

Applying a Gaussian Filter to Point Clouds and Neighborhood Search

When applying a Gaussian filter to point cloud data, the definition of the neighborhood region—i.e., which points to average—and the method used to compute it are key. Two representative neighborhood search methods are as follows.

• k-Nearest Neighbors (k-NN): For each target point, search for the k spatially nearest neighboring points and use that set for smoothing. For example, you might set “the 20 nearest points for each point” and average them with weights based on distance. k-NN is easy to implement and guarantees a fixed number of local points, but with nonuniform point density it can include distant points or remain confined to a very small area.

• Radius Neighbors: Use a predetermined radius R—represented as a sphere in 3D or a circle in 2D—and collect points within that range as neighbors. For example, gather points within a radius of 50 cm (19.7 in). Radius neighbors allow smoothing over a fixed physical distance scale regardless of point density, which is an advantage; however, the number of neighbors may be extremely few or many depending on point sparsity, which can produce uneven results.

For the collected neighbors, apply a Gaussian function such as w = exp(-(d^2)/(2σ^2)) to weight by distance d and then compute the average. In practice, spatial indexes like KD-trees are commonly used to efficiently search neighbors for each point. The resulting weighted average is used either to update the original point’s position or to produce a new smoothed point cloud.

An advantage of the Gaussian filter is that it is less influenced by outliers than a simple average. Distant anomalous points (outliers) receive extremely small Gaussian weights and thus contribute negligibly to the average (some implementations also set a distance threshold beyond which neighbors are not considered). Meanwhile, points within the normal range are averaged appropriately, smoothing small jitter.

A caveat is that if the neighborhood contains a geometric edge (a step or corner), the filter can round that edge. For example, smoothing the corner where two walls meet at a right angle can blur the boundary between the two wall surfaces. Therefore, when preserving important edges, use a smaller radius or σ for the filter, or mask out edge areas and skip smoothing there.

Effects of Gaussian Filtering on Civil Engineering Sites and Accuracy Improvement

Consider the practical effects of Gaussian smoothing in civil surveying and structural measurement.

First, in terrain surveying (e.g., as-built management for fill/cut), drone or terrestrial LiDAR scans produce large terrain point clouds but may also include noise points from vegetation, heavy machinery, or people. Applying a Gaussian filter to smooth or remove these points yields a terrain point cloud that represents the intended smooth surface—resulting in a terrain model with fine irregularities removed. This improves fill/cut volume calculations and enables soil quantity estimates that are less sensitive to extraneous noise. In slope gradient measurements, areas that appeared to have abrupt slopes due to noise become stable and close to the true average slope after smoothing. On-site, this clarifies comparisons with design values and allows accurate judgment of corrective work.

Next, in structural measurement (e.g., bridge or tunnel inspections), high-precision scans capture minute surface irregularities and defects, but they also include sensor noise. For example, tunnel wall scans may have millimeter-level noise added across the surface, making an otherwise smooth wall appear rough. Gaussian smoothing removes unwanted surface roughness and highlights the truly noteworthy defects (such as bulging or delamination). For bridge scans, parts of the surface other than rivets or bolts become smoother, making it easier to understand overall distortion or deflection. Gaussian smoothing is particularly effective for assessments of global planarity (for example, “Is this deck overall flat?”). Because it is robust against random outliers, Gaussian smoothing stabilizes flatness verification results and helps capture average surface trends.

Applying a Gaussian filter reduces noise-induced error components and thus improves analysis accuracy. In concrete validation cases, comparisons of RMSE (root mean square error) to reference planes before and after smoothing have shown post-smoothing errors reduced to less than half in some reports. In as-built management workflows that follow specification guidelines, combining position correction using control points with smoothing further improves the final deliverable quality.

Comparison of Gaussian Filter with Other Filtering Methods

There are various methods besides the Gaussian filter for removing noise from point clouds, each with different characteristics. Here we compare the Gaussian filter with two representative alternatives: the median filter and plane fitting (plane approximation).

• Median filter: The median filter reduces the influence of outliers by taking the median of neighboring points. In image processing it is commonly used to remove salt-and-pepper noise; in point clouds, an application might compute the median of the x, y, and z coordinates among neighbors to determine a new point position. The strength of the median filter is that a single extreme outlier does not affect the median, so it is very robust against impulse noise. It is effective even in small neighborhoods and is suitable for removing isolated speckle points in a point cloud. Its downside is that, because it uses medians rather than averages, its smoothing effect on small fluctuations is weaker than that of a Gaussian filter. Also, if the window (neighborhood) is too large, edges can become rounded or points may be moved to positions that do not exist in reality.

• Plane fitting (Moving Least Squares and similar): Plane fitting smoothing fits a virtual surface (plane or curved surface) to each point’s neighborhood and repositions points onto that surface. A representative method is Moving Least Squares (MLS). By finding the mathematically closest local surface to a noisy point neighborhood, MLS removes noise while smoothing along the original geometry. The difference from the Gaussian filter is that smoothing is done by approximating the neighborhood with a geometric surface rather than by averaging. Advantages include very smooth and high-precision results for data containing planes or smooth surfaces—surfaces are more faithfully reproduced, which is beneficial for component surfaces. Downsides are higher computational cost and longer processing time, and when multiple different surfaces (e.g., at edges) exist in the neighborhood, the result can be averaged and shapes blurred. The algorithms are also more complex and harder to implement.

In summary, the Gaussian filter is a balanced method: easy-to-implement weighted averaging that excels at reducing random noise. The median filter is effective for removing extreme outliers, and plane-fitting methods are suitable for high-precision surface recovery. In practice, multiple filters are often combined sequentially. For example, a workflow might remove obvious outliers using Statistical Outlier Removal and then apply a Gaussian filter to smooth fine noise, leveraging the strengths of each filter.

Efficient Filtering for Point Clouds Acquired with LRTK

Recently, [LRTK](https://lrtk.lefixea.com) has emerged as an accessible, high-precision surveying system using a smartphone, attracting attention in civil surveying. LRTK combines a smartphone-connected centimeter-class GNSS receiver (RTK) with the smartphone’s built-in LiDAR scanner to acquire detailed point clouds with absolute coordinates simply by walking the site. Filtering plays an important role in efficiently handling the point clouds obtained in this way.

Smartphone-based point cloud scanning is easier than traditional laser scanners, but it produces a very large number of points at once and includes noise characteristic of smartphone LiDAR. LRTK uploads acquired data to the cloud and automatically performs noise removal and reduction of unnecessary points. For example, duplicate points from multiple scans and points from dynamic objects such as people and vehicles are detected and removed in the cloud to generate a clean, orderly point cloud. Combining Gaussian smoothing further reduces surface variability of terrain and structures, producing a high-quality point cloud. As a result, the point cloud’s data volume is reduced and the accuracy and speed of subsequent analyses (e.g., as-built heatmap displays and volume calculations) are improved.

Filtered LRTK point clouds have fewer false detections due to noise, allowing immediate use of measurement results on site. For example, when displaying an as-built heatmap on a smartphone right after scanning to check differences from the design, denoised data produces clearer color distributions so that excesses/deficits of earthwork or construction mistakes can be recognized at a glance. Traditionally, noise removal was performed back at a desktop PC as a manual step; with LRTK, automated on-site processing enables substantial operational efficiency gains.

Conclusion: A New Workflow with Smartphone RTK × Point Cloud Scanning × Gaussian Smoothing

This article explained Gaussian-filter-based noise removal and smoothing for point cloud data, from basic concepts to practical applications. The Gaussian filter is a highly useful tool in civil surveying and 3D scanning workflows because it reduces inevitable noise in point clouds and improves data accuracy and usability. It contributes to correcting measurement errors and improving as-built verification accuracy, making it valuable to both field engineers and researchers.

A new workflow that combines smartphones, RTK positioning, and 3D scanning—smartphone RTK × point cloud scanning—has been put into practical use in recent years. LRTK is a representative example that allows anyone on site to quickly acquire high-accuracy point clouds and automatically apply filter processing such as Gaussian smoothing. This streamlines the entire process from acquisition to analysis and enables far faster accuracy verification and as-built confirmation than traditional methods.

What used to be time-consuming noise processing in specialized software can now be completed on-site in real time with systems like LRTK. The workflow that combines the reliability of smartphone RTK positioning, the detail of point cloud scanning, and the smoothing effects of Gaussian filtering will strongly support DX (digital transformation) in the civil engineering and construction industries. If you are experiencing issues with point cloud noise or processing time on your sites, consider using Gaussian filters and evaluating modern smartphone point cloud solutions such as LRTK.