Improving AR Integration Accuracy with High-Precision Point Cloud Acquisition by LRTK and Gaussian Smoothing

Background: Challenges and Opportunities in AR–Point Cloud Integration

Augmented reality (AR) technology is attracting attention in construction and surveying. AR, which overlays virtual 3D models and information on a site simply by pointing a smartphone or tablet, is a powerful tool for intuitively sharing plans and CIM models. However, a major challenge for practical AR use is the accuracy of alignment with the real world. For example, when projecting a building design model onto the actual site, a position offset of just a few tens of centimeters can make the model appear to float or sink into the ground. Even slight jitter in the display can hinder precise verification and work.

To address this alignment accuracy issue, it is important to measure the real world with high precision and use those measurements as the AR reference. This is where 3D point clouds become valuable. A point cloud is three-dimensional data composed of countless points captured by laser scanners or LiDAR that records the real-world shape in detail. Using a point cloud as the basis for AR makes it possible to detect differences between the virtual model and the real shape down to the millimeter level. Historically, however, acquiring point clouds required expensive equipment and expert skills, and the resulting data often contained noise and errors. These factors meant that combining point clouds with AR sometimes still resulted in jitter and position offsets, preventing the expected accuracy improvements.

A recent solution to these challenges is the combination of high-precision point cloud acquisition using LRTK and the application of Gaussian smoothing (Gaussian filter) to that data. LRTK is a system that enables anyone to easily acquire high-precision point clouds with absolute coordinates using RTK-GNSS technology, and applying Gaussian kernel–based smoothing to these point clouds can dramatically improve AR integration accuracy. This article explains the mechanisms and effects in detail.

What are RTK-GNSS-referenced high-precision point clouds acquired with LRTK?

LRTK combines a smartphone’s (for example, an iPhone’s) built-in LiDAR sensor or camera with high-precision positioning from real-time kinematic GNSS (RTK-GNSS) to form a point cloud measurement system. By attaching a compact dedicated receiver weighing about 165 g to the smartphone and using a proprietary app, centimeter-level positioning and 3D scanning that once required specialized equipment become accessible to anyone. With RTK-GNSS corrections, location errors that normally amount to about 5–10 m with standard smartphone GPS can be reduced to approximately ±1–2 cm horizontally and around ±3 cm vertically. Each point in the acquired point cloud is assigned an absolute coordinate composed of latitude, longitude, and elevation, so the point cloud itself becomes a high-precision 3D map aligned with the site coordinate system (public coordinates or an arbitrary local coordinate system).

LRTK can acquire such high-precision point clouds with surprisingly simple procedures. For example, even on wide roads or slopes, you can scan point clouds covering tens to hundreds of meters in a short time (in some cases 1–2 minutes) by walking while holding a smartphone. Because every captured point includes global position information, post-processing for alignment is unnecessary. It is like obtaining a “digital replica” of the entire site from actual measurements. These LRTK-acquired high-precision point clouds are highly effective as the foundation for AR integration. If the 3D design models (BIM/CIM, etc.) are created in the same coordinate system, simply overlaying the acquired point cloud with the design model will result in an exact positional match, removing the need for cumbersome alignment tasks. Moreover, ensured vertical accuracy reduces problems where models sink into or float above the ground.

Spatial noise in point clouds and issues for AR use

Even though LRTK-acquired point clouds are high precision, as with any sensor-derived measurement, they are not free from noise and error. Spatial noise in a point cloud refers to point variations such as unwanted bumps on surfaces that should be flat or slight spread of points that should be concentrated. Causes include LiDAR ranging errors, environmental factors (varying reflectivity or occlusion), and slight device motion. For example, a point cloud scanned from a perfectly flat concrete floor can still show height variations of several centimeters, creating a cloud-like spread. Using a noisy point cloud directly in AR can cause problems.

First, when detecting or recognizing planes and objects from a point cloud to anchor virtual content, noise can cause false detections or instability. AR apps detect planes in the environment to place virtual objects, but if the floor point cloud is uneven, it won’t be recognized as a precise plane and virtual objects may wobble.

Second, if the point cloud itself is used as the AR background reference, point dispersion from noise appears as AR overlay jitter. Imagine a user slightly moving a device and seeing the virtual model shift relative to the point cloud. Localized errors in the point cloud cause the relative positions between the virtual model and the camera view to change irregularly with viewpoint changes.

Third, when comparing acquired point clouds with design models to evaluate consistency, noise adversely affects the assessment. Slight offsets caused by measurement noise may appear even in areas that should perfectly match, making it hard to determine what is construction error and what is measurement noise.

As described above, noise and point dispersion in point clouds lower the achievable accuracy when integrating with AR. Even with high-precision LRTK point clouds, millimeter- to centimeter-scale variations cannot be ignored. In the real world, a few centimeters translate into visibly noticeable shifts or jitter in AR, so filtering of point cloud data is necessary. Among methods, smoothing—averaging each point’s position with its neighbors—is a fundamental and effective technique for noise reduction. The next section explains the principles of the representative smoothing method: Gaussian smoothing.

Basics of Gaussian smoothing: spatial weighting and the kernel mechanism

Gaussian smoothing corrects each point’s values (such as positional coordinates) by smoothing them using a weighted average of neighboring points on the point cloud. The key element is the Gaussian kernel weighting. The Gaussian kernel is a weight function based on the familiar normal distribution (the “bell curve”) that gives larger weights to closer points and smaller weights to more distant points. Specifically, a neighbor at distance d from the target point receives a weight proportional to exp(-d^2/(2σ^2)) (σ is the standard deviation of the Gaussian distribution). Applying such continuous, distance-dependent weighting enables spatially smooth smoothing.

A good mental image is placing each point on an elastic sheet and letting neighboring points pull on each other to adjust positions. Extremely isolated outliers are hardly pulled by neighbors and thus their conspicuous noise is reduced. Dense clusters of points mutually correct positions, smoothing surfaces and curves. The advantage of Gaussian smoothing over a simple mean filter is that it adjusts the degree of smoothness according to distance: close points are largely preserved while distant points have less influence, helping to maintain edges (boundary shapes) while reducing noise. Very sharp corners or fine details may become slightly rounded, but for AR integration the benefit of noise reduction usually outweighs the slight shape change.

Implementation and parameter tuning: effects of standard deviation and neighbor selection

Applying Gaussian smoothing to point cloud data in practice requires several parameters and implementation considerations. The main points are the Gaussian kernel standard deviation σ (smoothing strength) and the method for selecting neighboring points used in smoothing.

The standard deviation σ directly controls the smoothing range and strength. Smaller σ values make the Gaussian weights decay quickly, so only very close points are averaged. This removes fine noise while preserving small bumps and edges. Larger σ values give non-negligible weights to more distant points, averaging over a broader area; this smooths larger irregularities but may blur fine features such as sharp edges. The appropriate σ depends on the noise level of the point cloud and the scale of the target shapes. For typical LiDAR point cloud fine-noise removal, σ is often set on the order of millimeters to centimeters and then adjusted while observing results. The key is to balance removing noise while preserving important shape features, avoiding extremes.

Selecting neighbors determines which points around a target are considered “neighbors.” Typical approaches include taking points within a fixed radius or the nearest k points. When using a radius, set the radius to be on the order of σ or larger to include enough neighbors; too small a radius yields insufficient neighbors and unstable smoothing, while too large a radius can incorporate unrelated points from different objects and distort shapes. For k-nearest neighbors, k should be sufficiently large (for example, 20 or 50) to achieve stable averaging. However, varying point density across the cloud can cause fixed-k selection to over- or under-smooth in different areas, so combining radius and k constraints can be effective.

Another practical consideration is computational efficiency. Neighbor searches and weight calculations for every point can be expensive for large point clouds. Using spatial indices like KD-trees for neighbor search enables efficient computation. In LRTK systems, cloud processing can automatically apply noise reduction and smoothing to uploaded point clouds so users obtain consistent effects without handling parameters. If processing locally, it is prudent to start with relatively weak smoothing and gradually increase strength as needed.

Post-smoothing point cloud models and examples of reduced jitter in AR overlays

So what effect can Gaussian smoothing have on AR displays when applied to point clouds? Consider a paved road point cloud example. The raw LRTK point cloud showed several centimeters of vertical variation even on areas that visually appeared flat. When used directly to display virtual design lines (e.g., center lines or target finished heights) in AR, slight vertical jitter of the virtual lines was observed as the device moved. The virtual lines were dragged by micro-unevenness in the road point cloud, causing visible instability with viewpoint changes.

Applying a Gaussian smoothing filter to this road point cloud averaged out fine irregularities. Specifically, smoothing used a kernel with σ around 5 cm over a radius of several tens of centimeters. The smoothed point cloud showed the road surface as an almost ideal gentle plane, with the original centimeter-scale bumps much less noticeable. Using this smoothed point cloud as the AR reference, the previously jittery virtual line became firmly stable and remained steady even when the device moved—users reported that “the virtual line looked as if it had really been painted on the ground.”

This case shows that Gaussian smoothing removes random error components in point cloud data, reducing jitter in AR overlays. With LRTK’s globally referenced point clouds, gross positional error is already minimal, so residual fine-scale noise is the main challenge. Averaging out that noise stabilizes the apparent point cloud across consecutive frames, and consequently the appearance of virtual objects also becomes stable. In field AR use, stability when users walk or move their gaze is important for safety and trustworthiness. Gaussian smoothing significantly contributes to such practical quality improvements.

Application to consistency assessment between point clouds and CIM/3D design models

Gaussian smoothing is useful not only for stabilizing AR displays but also for evaluating the consistency between point cloud data and design models (BIM/CIM). On construction sites, scanned point clouds of completed structures or terrain are compared with original design models to verify compliance with plans. For smooth evaluation, point cloud data should be free of extraneous noise so that only genuine differences with the design model are highlighted.

Raw unsmoothed point clouds often contain several-centimeter variations, as mentioned earlier. When overlaying a design model and inspecting differences, noise can cause colored discrepancy regions (for example, nonzero errors on a distance map) to appear where the model actually matches, making it hard to detect true construction deviations. By Gaussian smoothing the point cloud, random variations are greatly reduced and clear trends emerge on a difference map. For instance, if a region’s point cloud is uniformly 3 cm higher than the model, it becomes easier to conclude that the construction is actually 3 cm too high. Conversely, sub-centimeter differences masked by ±3 cm noise before smoothing may effectively disappear—acceptable to ignore—thereby removing noise-driven uncertainty from the assessment.

Gaussian kernels also assist in alignment (pose) estimation between point clouds and models. Algorithms such as ICP (Iterative Closest Point), commonly used to minimize distances between corresponding points when registering point clouds or point clouds to models, can suffer from incorrect correspondences or unstable convergence when the point cloud is noisy. Smoothed point clouds reduce the influence of outliers and make it easier to align based on larger-scale shapes. In other words, smoothing local irregularities with a Gaussian kernel facilitates global alignment.

Thus, Gaussian smoothing aids the entire process of comparing and reconciling point clouds with CIM/3D models, enabling more reliable verification. When deciding, for example, “this area has been overfilled compared to the drawings” or “this slope is gentler than planned,” smoothing lowers the risk of misjudgment due to noise.

Effects of introducing Gaussian smoothing: stable AR and high-precision matching

As discussed, combining high-precision point clouds with Gaussian smoothing leads to a qualitative leap in AR–point cloud integration. The introduction effects can be summarized as follows:

• Stabilized AR overlays: Noise-driven jitter from point clouds is suppressed, so virtual models overlaid in AR appear firmly fixed to the real world. Reliance on markers or manual alignment is reduced, allowing users to utilize AR displays without stress.

• High-precision model matching: Smoothed point clouds with global coordinates enable high-accuracy overlap between site point cloud data and design models in the same coordinate system. This allows differences between model and reality to be understood to millimeter- to centimeter-level precision, supporting quality control and as-built inspection with confidence.

• Improved consistency of scan data: Gaussian smoothing reduces frame-to-frame inconsistencies and positional fluctuations in sequentially acquired point cloud frames. Large datasets scanned over time can be treated as a single smooth model, easing post-processing and analysis.

• Improved field work efficiency: With point clouds that require no alignment and stable AR, on-site tasks such as verifying measurements or issuing instructions become much more efficient. For example, continuously referencing a design model overlaid in AR during construction reduces surveying and inspection workload and enables early detection of errors.

In these ways, Gaussian-smoothed point clouds significantly enhance the reliability and accuracy of AR use, strengthening the bridge between digital and physical realms.

The future of on-site AR using LRTK × Gaussian smoothing

The combination of LRTK high-precision point cloud acquisition and Gaussian smoothing is likely to drive AR adoption on site to the next level. Future developments may include:

• Real-time feedback: LRTK can provide cm-level positioning in real time. Combining this with fast point cloud smoothing could enable the construction of smoothed point cloud models on the fly while holding up a device, with immediate reflection in AR. Integrating surveying and verification in real time would allow instant checks and feedback throughout construction processes.

• Integration with AR glasses: While smartphones and tablets are currently the main AR platforms, widespread adoption of AR glasses could enable hands-free viewing of LRTK position data and smoothed point clouds. Workers wearing glasses could see BIM models consistently overlaid without drift, digitally supporting tasks that previously relied on experience and visual inspection.

• Cloud sharing and remote support: Sharing LRTK-acquired point clouds and AR data via the cloud would allow office-based managers and designers to inspect and verify alignment on remote point cloud models and provide instructions. Smoothed point clouds also reduce data size, lowering communication loads and making remote AR support more practical.

• New quality-control standards: As high-precision point clouds and AR become common, standards for construction management and inspection may evolve. For example, position marking previously done with strings and chalk could be replaced by AR-projected guidelines, and as-built inspections could be judged via AR-based color-coded difference maps. Stable smoothed point clouds would be an indispensable component of such digital-first methods.

Thus, LRTK positioning and Gaussian smoothing expand the possibilities for AR use on site. AR with ensured accuracy and stability can move beyond mere visualization to become core infrastructure for fieldwork.

First steps toward practical use: start high-precision point cloud acquisition with LRTK

We have described the benefits of combining advanced point cloud processing and AR technologies. To realize these benefits in projects, obtaining high-precision point cloud data is essential. With LRTK, absolute-coordinate point cloud measurement that previously required professional surveying instruments can be started easily with just a smartphone. Preparation and operation are simple, so even technicians without special training can perform on-site 3D scans in a short time. Acquired point cloud data can be automatically uploaded to the cloud and used on-site for measurements of distance and area, model overlays, and more. If necessary, Gaussian smoothing can be applied to further reduce noise and improve accuracy.

The combination of high-precision point clouds and AR is expected to become increasingly important in practice. As a technology for bridging the digital and real worlds, it has potential applications across construction management, infrastructure maintenance, urban planning, and disaster prevention. As a first step, consider starting with simple surveying using LRTK. By precisely visualizing site conditions and intelligently smoothing and using that data, you can achieve stable AR experiences and improved accuracy control.