タイトルなし

\# This Is What Causes AR Drift! How to Eliminate Errors On-Site

Table of Contents

• What is the phenomenon of AR drift?

• Main causes of AR drift

• How to eliminate AR drift on-site

• Achieve high-precision AR with simple surveying using LRTK

• FAQ

What is AR positional drift?

Have you ever used AR (augmented reality) and noticed that the displayed 3D models or drawings appear shifted from their real-world positions or become misaligned over time? For example, on a construction site, if you display pipe installation positions in AR on a smartphone but they subtly differ from the actual locations, or the virtual model gradually drifts as you walk around, that would be extremely troublesome. Such AR positional drift may be tolerable as a minor error for entertainment purposes, but it becomes a major problem when used in real-world settings like construction or surveying. A deviation of several centimeters can lead to misplacement of structures, and if you cannot trust the positions confirmed in AR, the technology will be useless on site. To prevent incidents like "it's not where I expected" or "the position was wrong after following the display," it is important to correctly understand the causes of AR drift and take measures to eliminate errors in the field.

Main Causes of AR Misalignment

There are several technical factors behind AR displays becoming misaligned with reality. The main causes can be summarized as follows.

• Insufficient location accuracy (GPS errors): When acquiring a device's position outdoors, the typical smartphone built-in GPS can have errors of several meters (several ft). Therefore, in AR that requires absolute alignment, the initial placement can become greatly misaligned with reality. Especially when displaying models in a large outdoor space based on map coordinates, being off by several meters (several ft) is unacceptable. Also, in urban areas with tall buildings or in mountainous regions, GPS satellite signals are easily reflected or blocked, further degrading positioning accuracy and causing models to not be displayed where expected.

• Device orientation and pose errors: In AR, the direction and tilt of the device (smartphone or tablet) are also important. If the compass (electromagnetic sensor) is off so that the north reference is rotated by several degrees, the model's orientation relative to reality will likewise be incorrect. Magnetic sensors are easily affected near metal or electronic devices, making directional errors likely. Also, small biases in a smartphone's gyro and accelerometer, or errors in detecting the ground's level, can cause the model's tilt or vertical position to be off by several cm to more than ten cm (a few in to several in). Such pose angle errors are often overlooked but cannot be ignored for precise alignment.

• AR tracking drift: A smartphone's AR functionality (spatial recognition using the camera and sensors) fundamentally relies on relative position tracking. When the device moves, it estimates motion from movement of feature points in the camera image and from IMU sensors, but those estimates include slight errors. As these errors accumulate with movement, a phenomenon occurs where the position of virtual objects that was initially correct appears to gradually drift over time. This so-called drift is especially noticeable when walking around over tens of meters or more, or in flat areas with few distinctive features. If, in AR glasses or smartphone AR apps, a model that initially aligned perfectly eventually appears to float or shift out of place, this limitation in tracking accuracy is likely the cause.

• Anchor and marker misconfiguration: In AR, to fix a virtual model in real space, anchors (reference points) and markers (image markers or QR codes, etc.) are sometimes used. Improper use of these can cause positional drift. For example, when a printed marker is attached to the on-site reference position to place a model, if the marker's placement is not accurate or the marker moves during the process, the model will shift along with it. It is a common failure that when a marker is displaced by strong wind, the CG also shifts. Also, if an anchor is aligned by sight from a single location, it may look correct there but the misalignment becomes noticeable when you move to a different viewpoint. This is a case where depth scale or angle was distorted because alignment was done from only a single viewpoint. If the reference alignment procedure or the accuracy of marker placement is insufficient, the misalignment will continue to cause problems later.

• Design data and on-site coordinates mismatch: In the use of AR in construction, the coordinate systems of the underlying drawings or 3D models can sometimes differ from the site coordinates. For example, if the design drawings are created in a proprietary local coordinate system or a different geodetic datum, they will not match the actual site positions as-is, and the AR display can be offset by several meters (several ft). There are cases where the whole model ended up shifted 2 m (6.6 ft) southeast because it was placed without performing a coordinate transformation… In this way, incorrect handling of the source data's coordinates can also be a factor causing AR positioning to be out of alignment.

As described above, multiple causes can combine to produce the problem of "AR misalignment" on-site. So how can we suppress such misalignment and accurately overlay AR displays at the intended positions in the field? Next, we will explain the countermeasures.

How to eliminate AR misalignment on site

To minimize AR misalignment on site and, ideally, bring the error close to zero, it is necessary to take countermeasures corresponding to the root causes. Below are several effective methods.

• Alignment method tips (markers and multiple-point alignment): When placing a model, avoid relying on visual alignment from a single point and align using multiple reference points. For example, matching two or more known points at the model’s edges to their corresponding real-world positions can correct scale and rotation offsets. Also, if you can install AR markers on site, secure them firmly before use. Marker-based alignment is very high precision, but be aware it involves physical installation and management costs. Outdoors, fixing markers may require installing structures and obtaining permits, so assess cost versus benefit when deciding whether to use them. If installation is difficult, using distinctive existing site features (building corners, patterns, etc.) as alternative markers and manually aligning the model to them is also effective.

• Use of AR cloud (spatial scanning): There is also a method that aligns positioning using surrounding environmental data without placing markers. You pre-scan the site with a 3D scan, and by matching that map data with camera footage in the cloud, the system can estimate the device's position with high accuracy. The advantage is that you can perform alignment without actually placing physical markers on site, but it requires the effort of conducting an accurate scan beforehand. Also, if the site's condition (object placement or scenery) changes after scanning, accuracy can suffer, so careful operation and experience are necessary to ensure reliable use. Recently, services such as geospatial AR have emerged that can determine location from the outset using 3D maps of locations around the world. However, such AR cloud systems have limited coverage areas and may not be usable indoors, so they are not a panacea.

• Pre-adjustment of data coordinates: It is important to unify the coordinate system of the data shown in AR, such as drawings and 3D models, with the on-site survey coordinate system in advance. If the design data has absolute coordinates such as latitude/longitude or plane rectangular coordinates, it can be automatically placed simply by matching it with the smartphone’s position information. On the other hand, for drawings created in a local coordinate system, you need to perform a coordinate transformation—for example by taking correspondences of two or more points using known points on site—before importing them into AR. Skipping this extra step can lead to discovering large misalignments later and having to readjust. If you are unsure whether the coordinates match, it is reassuring to perform a simple on-site survey to confirm the actual coordinates of the reference points.

• Sensor calibration: Pre-calibrating sensors can also be effective for reducing device-side error sources. For example, an electronic compass can be calibrated by swinging the device in a Figure-8 (figure-eight) motion, and leveling (tilt) accuracy improves if you place the device on a reliable flat surface and reset it. It is also important to look around and let the environment be sufficiently scanned immediately after launching the AR app. Especially indoors or in locations with few distinctive features, move the device consciously so the camera recognizes walls, floors, etc., creating a situation in which the AR can more easily understand the environment. Frequently resetting and correcting sensor errors can help prevent the accumulation of drift to some extent.

• Periodic anchor reset: When using AR while walking for long periods or across wide areas, tracking errors will inevitably accumulate. In such cases, one option is to reset the AR session and reposition the model. Alternatively, re-create anchors at key points and readjust them so they align correctly at those locations. Some AR glasses and smartphone AR apps allow you to manually correct alignment mid-session. On large sites, dividing the area into sections and repositioning the model for each—split the area and update the AR display sequentially—can be effective so you don’t have to walk large distances at once. If you notice it starting to drift, don’t force it; resetting and trying again as needed will ultimately maintain accuracy.

• Use of high-precision positioning technologies: As a fundamental solution, there is a method that leverages high-precision GNSS positioning. In recent years, RTK (Real Time Kinematic) technology has made centimeter-level (cm) positioning possible even on smartphones. By using an RTK-compatible external antenna, the typical smartphone GPS error of about 5 m (16.4 ft) can be reduced to a few cm (a few in). This makes it possible to reduce the initial placement offset of a model to almost zero. If the absolute position is aligned, the effort required for subsequent fine-tuning is greatly reduced. In Japan, the Michibiki (Quasi-Zenith Satellite System) provides sub-centimeter-level (cm) positioning services (such as CLAS), and if equipment that can use these services is deployed on-site, AR overlays based on map coordinates can be aligned without offset.

By combining the above measures, AR misalignment can be greatly reduced. However, until now high-precision positioning equipment has relied on specialized surveying instruments or expensive GNSS receivers, making it difficult to put to immediate use on typical job sites. That is where a new approach appears: simplified surveying using LRTK, which will be introduced next.

Achieving high-precision AR through simple surveying using LRTK

Finally, as a practical solution to bring AR errors on site as close to zero as possible, we introduce simple surveying using LRTK. LRTK is the latest high-precision positioning solution that uses smartphones. By attaching a dedicated small antenna to a smartphone to perform positioning, it is designed so that anyone can easily benefit from RTK-GNSS (real-time kinematic positioning). What is revolutionary is that centimeter-level positioning, which until now could only be handled by skilled surveying technicians, can be achieved with one smartphone per person.

By using simplified surveying with LRTK, for example on a construction site, you can instantly obtain the coordinates of reference points with several-cm level accuracy (half-inch accuracy) and accurately set AR model anchors at those locations. Because it can record data at a level of positional accuracy that ordinary GPS cannot achieve, it can be applied not only to AR visualization but also to as-built management and volume measurement; here, however, we focus particularly on its effect on AR alignment. If high-precision coordinates are obtained, the model displayed in AR will precisely overlap the location in the design drawings. For example, when checking pile-driving positions, simply displaying an AR pile model at the position measured with LRTK corresponding to the design coordinates can achieve a state in which there is almost no discrepancy between the real object and the virtual one.

Additionally, LRTK is designed with ease of use on-site in mind, and the operation is simple — for example, positioning can be completed just by pressing a button on the smartphone screen. The acquired point data can be immediately saved to and shared via the cloud, allowing the entire team to leverage AR while referring to the latest reference point information. In other words, LRTK's strength is that it enables high-precision alignment even without specialized surveying skills.

By introducing simple surveying with LRTK in this way, even sites that have long suffered from “AR drifting…” can align virtual models to the real world to a level that can be described as virtually zero error. With a solid foundation of high-precision positional information, AR drift becomes nothing to fear. AR whose positional reliability is guaranteed will no longer be an entertaining demo but will serve as a practical business tool. Note that LRTK using GNSS cannot, in principle, be used in indoor or underground spaces. When using AR in environments where satellite signals cannot reach, such as inside buildings, it is necessary to combine conventional alignment methods using markers or visual references. By using LRTK for large outdoor sites and markers or plane detection indoors, you can broaden the range of AR applications across all kinds of scenes.

FAQ

Q: How much does the position drift in smartphone AR? A: With a typical smartphone-only AR display, errors on the order of tens of centimeters (several to a few dozen in) to several meters (a few to several dozen ft) can occur. For example, when relying on GPS outdoors, it is not uncommon for the position on a map to be off by several meters (a few ft). Even indoors, if the AR’s self-localization is unstable, positional drift of tens of centimeters (several to a few dozen in) can happen. These issues do not prevent rough checks, but they leave doubts for tasks that require millimeter-level accuracy (about 0.04 in). Unless you use high-precision equipment or methods, assume smartphone AR accuracy is roughly at that level.

Q: Why do AR objects drift out of place when you walk around? A: This is due to AR tracking error (drift). The position tracking of smartphones and AR glasses computes relative movement from camera images and sensors, but this calculation contains small errors. Even tiny misalignments that are initially negligible accumulate as you move long distances or as time passes, causing the positions of virtual objects to gradually diverge from reality. Drift tends to be larger when moving across wide outdoor areas or in monotonous environments where sufficient feature points cannot be obtained. This phenomenon cannot be entirely avoided with current technology, but it can be mitigated by resetting anchors midway or correcting with high-precision positioning data.

Q: What should I do to reduce AR positional drift? A: First, take countermeasures according to the cause. For example, initial alignment offsets can be largely resolved by aligning the coordinate system of your drawing data to the site in advance. If markers can be used in the environment, use them proactively; if that’s difficult, be creative by aligning to site landmarks at multiple points. Also perform sensor calibration, and when starting AR look around so the system can sufficiently recognize the surroundings—this is basic. Above all, the decisive factor is raising positioning accuracy. If possible, use RTK-capable equipment to fix positions to the centimeter level (half-inch level) and place the model based on that—that is ideal. That way AR can start with almost no offset from the beginning, and subsequent fine-tuning will be minimized.

Q: What is RTK? Do you need dedicated equipment? A: RTK stands for "Real Time Kinematic" and is a technique that corrects errors in satellite positioning (GNSS) to achieve high accuracy. By calculating in real time the relative relationship with a reference station, positioning errors that are normally several meters can be reduced to a few centimeters. Traditionally, dedicated expensive equipment such as a base station and a rover set was required. However, today there are services that provide correction information from Japan's Quasi-Zenith Satellite System and compact devices that support it (for example smartphone antennas like LRTK), making RTK positioning easy to use.

Q: Can non-specialist survey technicians use LRTK? A: Yes, LRTK was developed with an emphasis on being easy to use even for non-surveying personnel on site. Its operation is simple: attach the antenna to your smartphone and press a button in the dedicated app to complete positioning. Even without complex settings or surveying knowledge, you can obtain accurate position information by following the on-screen instructions. Measurement results are automatically saved to the cloud and are easy to share, making the system convenient for individuals as well as entire teams. It does not require the familiarization training associated with traditional surveying instruments, and can truly be described as a tool aimed at "surveying anyone can do". This enables on-site staff themselves to achieve high-precision AR alignment.