Evaluation of marker-based medical image registration using augmented reality head-mounted display

1 Introduction

The integration of augmented reality head-mounted displays (AR HMDs) into surgical applications has advanced significantly, enabling use cases such as preoperative surgical planning (Wang et al., 2019), intraoperative image-guided surgery (Cao et al., 2020), and surgical training (Cao and Cerfolio, 2019). By superimposing medical imaging data as virtual content, AR HMDs provide enhanced visualization capabilities. In particular, image-guided intraoperative applications utilize segmented three-dimensional (3D) models derived from patient imaging data, such as computed tomography (CT) or magnetic resonance imaging, to overlay virtual models onto real-world patient anatomies. This capability improves surgical precision by offering visual guidance in procedures such as neurosurgery (Meola et al., 2017; Tabrizi and Mahvash, 2015; Kunz et al., 2021), orthopedic surgery (Jud et al., 2020; Andress et al., 2018), and other open or minimally invasive procedures (Farronato et al., 2019; Fida et al., 2018).

Image registration in AR often relies on computer vision algorithms to detect image-based or infrared (IR) reflective markers attached to patients or surgical instruments (Andrews et al., 2020). Using the built-in front-facing camera of an AR HMD, marker-tracking systems such as ARTag (Fiala, 2005), ARToolkit (Kato et al., 2000), and AprilTag (Olson, 2011) enable virtual content placement at the detected marker locations in real environments. Efforts to enhance detection accuracy have included innovations in marker configurations, such as varying sizes, patterns, and multi-marker systems (Pérez-Pachón et al., 2021). In contrast, IR reflective marker tracking typically involves external cameras, such as motion capture systems (Eom et al., 2025), or onboard environmental cameras (Kunz et al., 2020; Martin-Gomez et al., 2023; Gsaxner et al., 2021a). These systems, however, introduce computational demands and latency during rendering.

Achieving precise alignment of virtual anatomical models with patient anatomy remains a critical challenge in surgical applications, where sub-millimeter to millimeter accuracy may be necessary (Chen et al., 2017). Factors such as lighting conditions, computational load on the AR HMD, and the motion or number of tracked objects significantly influence registration accuracy. Operating rooms, with their complex environments, exacerbate these challenges. While IR reflective marker-based registration has demonstrated high accuracy—approximately 1–2 mm (Costa et al., 2024; Eom et al., 2025; Lee et al., 2020) — in aligning virtual models to stationary objects, image-based marker systems exhibit less precision, with jitter and misalignment errors ranging from 3 to 7 mm (Qian et al., 2018; Gibby et al., 2021; Eom et al., 2022). The performance disparity is more pronounced for moving objects, such as surgical tools, where motion variability and marker occlusion further complicate tracking. Marker occlusion is a challenge for all optical tracking systems, including both image-based marker systems and IR reflective marker systems (Martin-Gomez et al., 2023; Gsaxner et al., 2021a; Eom et al., 2025).

Despite technological advancements, a standardized methodology for quantitatively evaluating AR registration accuracy is lacking (Malhotra et al., 2023; Ma et al., 2023). Previous studies have used manual tools, such as digital calipers, to measure alignment errors from the perspective of AR HMD users (Eom et al., 2022; Van Gestel et al., 2024; Cui et al., 2023), or assessed image registration indirectly via task performance metrics (Doughty and Ghugre, 2022; Qian et al., 2019; Li et al., 2022; Mamone et al., 2021). However, these methods are susceptible to human error and are limited in evaluating error metrics beyond basic displacement errors. Many of the optical see-through AR HMDs, such as Microsoft HoloLens and Magic Leap, allow AR image captures using onboard front-facing cameras. However, such captures often fail to accurately represent user perception due to angular misalignments and depth discrepancies.

To address this limitation, high-resolution scientific cameras have been used to directly capture AR images through HMD displays. This approach of using scientific cameras to capture the virtual images has been adopted in standards (International Committee for Display Metrology, 2023; International Electrotechnical Commission, 2020) for assessing optical and video see-through HMD performance across metrics such as visual acuity, contrast sensitivity (Livingston, 2006), radiometric and photometric properties (Penczek et al., 2017), and spatiotemporal performance (Zhao et al., 2022). In medical applications, AR display configurations may be optimized for perceptual image quality, focusing on gamma and luminance calibration for optical overlays (Zhao et al., 2023). These advancements highlight the potential for objective assessments of AR registration accuracy through direct capture methods.

This study introduces a framework for evaluating marker-based AR registration on the HoloLens 2 optical see-through AR HMD, using quantitative error metrics to assess virtual content alignment accuracy and stability. The approach builds on the established use of high-resolution scientific cameras to directly capture images through HMD displays, a technique previously applied to HMD display characterization for spatiotemporal image quality assessment (Zhao et al., 2022) and the quantification of spatial resolution (Johnson et al., 2024). This methodology employs a five megapixel CMOS camera with a global shutter (Teledyne FLIR BFS-U3-51s5c) equipped with a 12 mm fixed focal length lens with a 39° horizontal field of view (FoV) to capture AR images at a constant frame rate, enabling the accurate assessment of registered virtual content in comparison to ground truth (GT) data. We used advanced quantitative metrics, including dynamic time warping (DTW), cross-correlation, and jitter, to assess image-based marker and IR reflective marker tracking approaches. Experimental conditions incorporated varying motion, rendering configurations, and marker placements to quantify the impact of factors such as HMD-to-marker distance, motion speed, and computational load on AR registration performance. The objective of this work is to develop and assess the evaluation methodology itself; while experiments are conducted on the HoloLens 2, the framework is designed to be generalizable to other AR HMD platforms, with adaptation of the tracking implementation to the target device’s hardware and software as needed.

The novelty of the proposed methodology lies in the systematic integration of these components: direct AR display capture following standardized HMD measurement procedures, precision-controlled GT motion that is fully independent of the tracking systems under evaluation, and newly proposed time-series metrics (DTW-based spatial RMSE and cross-correlation latency) that separately quantify spatial fidelity and temporal lag of registered virtual content. Because the virtual points and physical GT are captured simultaneously in the same camera frame and all measurements are relative to the co-located GT, the methodology inherently eliminates temporal synchronization errors and is invariant to systematic coordinate alignment offsets. To our best knowledge, no prior work has combined these elements to provide objective, per-frame characterization of AR registration accuracy and stability under controlled motion conditions. This methodology is intended as a reproducible methodological foundation for the medical AR community, where standardized evaluation tools are critically lacking and urgently needed as AR-guided surgical systems move toward clinical deployment.

Section 2 details the development of the testbench, which integrates an optical see-through AR HMD, precision motion control automation, a phantom model, and image-based and IR reflective marker tracking devices. Section 3 presents experimental results comparing image-based and IR reflective marker performance under stationary, translational, and rotational motion conditions. Finally, Section 4 discusses the insights, limitations of this study, and future directions, followed by conclusions in Section 5.

2 Testbench for AR registration

The objective of this evaluation is to quantify the spatial accuracy and temporal latency of virtual point projections rendered on the AR HMD display, relative to a physical fiber optic light source with a cosine corrector to provide Lambertian emission that was used as GT, as observed from the perspective of the user’s right eye via an external calibrated camera. Specifically, we measure the displacement between: (1) the centroid positions of virtual points rendered by the AR HMD to indicate the tracked marker location, and (2) the centroid of a co-located physical fiber optic light, captured simultaneously in the same frame by an external FLIR camera positioned at the right-eye location. Throughout this paper, we use the following terminology consistently. An image-based marker refers to a planar 2D marker detected by the HMD’s front-facing camera using a computer vision algorithm (Vuforia).

To achieve this, we developed a testbench that integrates: (1) an optical see-through AR HMD (HoloLens 2) running a custom AR application with image-based marker tracking and IR reflective marker tracking, (2) a precision motion controller (Zaber Technologies Inc, 2026) for automated, repeatable movement of both the HMD and the target marker, (3) a calibrated high-resolution FLIR camera (Teledyne FLIR BFS-U3-51S5C) positioned at the right-eye location to capture AR display images at 60 FPS, and (4) a physical light source co-located with the target marker center to provide a per-frame GT reference for centroid position. The overall architecture is illustrated in Figure 1.

All experiments in this study are conducted on the HoloLens 2 platform. Findings related to temporal warping, rendering performance, and marker tracking behavior are therefore specific to this device and its software implementation (Unity 2022.3.19f1 with Vuforia Engine v10.25 for image-based marker tracking, and the OptiTrack Motive software for IR reflective marker tracking). The testbench design is intended to be generalizable to other AR HMD platforms; porting and validation on other devices are discussed as future work in Section 4.3.

Two types of image-based markers are evaluated using Vuforia tracking: an ARToolkit binary-pattern marker (ART), which contains a low density of detectable feature points (Figure 3a), and a high-feature marker (HFM), which contains a high density of feature points and is used to evaluate the effect of feature point density on tracking performance (Figure 3b). An IR reflective marker refers to a retroreflective sphere detected by the external OptiTrack motion capture system (OT). A reference marker is a large stationary marker used exclusively for one-time coordinate system alignment between the HoloLens 2 and OptiTrack world coordinate systems; it is not used in per-frame target tracking. A target marker is the moving object being tracked during experiments, which carries both an image-based marker and IR reflective markers. The GT is the centroid of a physical fiber optic light mounted at the center of the target marker (Figure 3b), whose position is extracted from FLIR camera frames by the same centroid detection algorithm used for the virtual points.

2.1 Overall architecture

The bench setup including the Zaber system, the AR HMD, the marker tracking system, and the phantom model is shown in Figure 2. For the IR reflective marker tracking system, we installed eight Flex 13 OptiTrack cameras, each with a 1.8-megapixel resolution and capable of capturing at 120 frames per second (FPS). These cameras were evenly distributed on the ceiling to provide a 360-degree FoV tracking. For the image-based marker tracking, a printed marker was attached to the target marker. The target marker was designed with one degree of freedom for translational movement along the x-axis, while the AR HMD had three degrees of freedom for translational movement along the x- and z-axes and one rotational movement along the y-axis. Both the target marker and the AR HMD were securely mounted on two separate Zaber stages, positioned to face each other, as shown in Figure 2. The data flow among the three systems; (1) the AR HMD with the marker tracking systems, (2) the FLIR camera, and (3) the Zaber motion control system, is illustrated in Figure 1 and described in detail in the following subsections.

2.2 Marker tracking

To overlay virtual content onto the real world, an AR system must accurately track the position and orientation of physical objects. In this study, we evaluate two categories of marker tracking systems integrated into the HoloLens 2 AR application: (1) image-based marker tracking using the HoloLens 2 front-facing camera and Vuforia software, and (2) IR reflective marker tracking using an external OptiTrack motion capture system. For each system, the tracked position of the target marker center is rendered as a virtual point in the AR display. The displacement between this virtual point and the co-located physical fiber optic light (GT), as captured by the external FLIR camera, forms the primary measurement in all experiments.

We note that the optical marker tracking system (OptiTrack) serves a dual role in this study: (a) as a tracking system under evaluation using IR reflective markers (denoted OT in the results), and (b) as a component of the one-time coordinate system alignment procedure described in Section 2.2.2. These two roles are distinct: the per-frame target tracking (role a) is performed entirely by the OptiTrack system and is independent of Vuforia. The coordinate alignment (role b) uses a stationary reference marker that is tracked by both Vuforia and OptiTrack simultaneously, as described below.

2.2.1 Image-based marker tracking

We employ Vuforia marker tracking (Vuforia Engine v10.25), a commercial software compatible with the Unity and HoloLens platforms. HoloLens 2 is equipped with the following onboard sensor suite: four environment-tracking grayscale cameras used for inside-out spatial mapping and head pose estimation, an 8 MP RGB photo/video camera, and a 1 MP time-of-flight (ToF) depth camera for far-field depth sensing. Vuforia uses feature-based tracking via the built-in front-facing camera on the AR HMD to match feature points detected on registered image-based markers. This method performs particularly well with marker patterns that have a high density of feature points, such as edges and corners. In contrast to binary-pattern markers used in other tracking systems, Vuforia recommends using image-type patterns for optimal performance. To evaluate tracking performance, we selected two markers: 1) a binary pattern marker from the ARToolkit marker database (Kato et al., 2000) (Figure 3a), which features a low number of detectable feature points, and 2) an FDA logo, as image-based marker, which contains a high number of feature points (Figure 3b), to compare the tracking performance. These two markers were selected specifically to compare the effect of feature point density on HMD system tracking performance.

2.2.2 IR reflective marker tracking

We use the OptiTrack motion capture system, which provides high-precision tracking of objects equipped with IR reflective markers. To integrate IR reflective markers into our testbench, we designed a custom 3D-printed mount that allows for flexible configurations of marker attachments, as shown in Figures 3a,b. Four IR reflective markers are attached to the 3D-printed mount, with their configuration registered in the OptiTrack system. The system detects the center point of the IR reflective marker cluster, which aligns with the center of the image-based marker. To ensure optimal marker tracking performance, we minimized reflections by using non-reflective objects and covering all reflective surfaces in the scene with non-reflective black tape. Since the world coordinate systems of the AR HMD and the OptiTrack system differ, we performed a coordinate transformation to align the two systems. A reference marker (Figure 2) was designed, combining a 12 cm by 12 cm square image-based marker with four IR reflective markers, to allow tracking of a shared center point. The transformation of world coordinates is shown in Equation 1, where T F H is obtained by detecting the reference marker on HoloLens 2, T O F is the transformation between the image-based marker and the IR reflective marker, and T M O is the difference between the reference marker and the target marker tracked by the OptiTrack system.

T M H = T F H T O F T M O . (1)

We note that T F H is obtained by detecting the reference marker using Vuforia, which introduces a fixed, time-independent systematic offset into the coordinate transformation used to express OptiTrack-tracked positions in the HoloLens 2 world coordinate system. This offset applies equally to both OT and image-based marker tracking results. To bound the magnitude of this offset, the alignment was performed once per experimental session under static conditions using a large reference marker (12 cm × 12 cm) at D1 (800 mm), where Vuforia detection is most stable. Because the DTW-based RMSE and cross-correlation metrics evaluate relative motion about the mean positions rather than requiring exact absolute spatial localization, this static offset does not affect the quantitative results or the validity of the evaluation methodology.

2.3 Data collection

The AR HMD can capture AR images via its built-in front camera located above the display. However, these images do not fully replicate the user’s perspective through the HMD. To capture AR images that more closely resemble what the user perceives, we positioned a high-resolution FLIR camera at the location of the user’s right eye (Figure 3c). This camera collects AR images as raw data (Figure 3d) at 60 FPS, matching the HMD refresh rate, showing the GT light source and the virtual points rendered at the registered marker locations.

The center points determined by the image-based and IR reflective marker tracking systems are represented as white-colored virtual points in the AR scene, each with a 6 mm diameter (Figure 3d). This size approximates the physical light used as the GT for alignment, ensuring consistency in visual comparison. The virtual points and the physical light are aligned vertically in the AR environment, enabling a straightforward analysis of tracking performance. To compute the center of the virtual points from captured images, the raw AR images are binarized (Figure 3e), and a threshold-based point detection algorithm is applied. Only regions containing clusters of multiple white pixels are retained to minimize detection errors caused by single-pixel noise artifacts.

Because both the HMD and the OptiTrack systems track the same physical target marker center, their corresponding virtual points would otherwise appear at the same pixel location in the FLIR camera image, making independent centroid detection impossible. To prevent this overlap, a fixed vertical offset is introduced in the AR application’s world coordinate system: the ART virtual point is placed 3 cm below the OT virtual point along the vertical axis of the AR scene. The physical GT fiber optic light is co-located with the OT virtual point (i.e., at zero vertical offset). This results in three vertically separated white points in each FLIR camera frame: the GT light source at the top, the OT virtual point immediately adjacent to the GT, and the ART virtual point 3 cm below. This 3 cm offset is applied in 3D world coordinates and is therefore a fixed, distance-independent separation in the AR scene. The corresponding pixel separation between the ART and OT virtual points in the FLIR camera image varies with HMD-to-marker distance according to standard perspective projection, but in all cases the three points remain vertically distinct and independently detectable by our centroid detection algorithm at all experimental distances (800–1400 mm). Critically, this vertical offset does not affect the horizontal displacement measurements used in all RMSE and latency calculations: all spatial error metrics are computed along the horizontal axis of motion only, and each tracking system’s virtual point centroid is detected and analyzed independently.

In addition to AR images collected from the FLIR camera, we gather supplementary data from both the HoloLens 2 and the Zaber motion control system. The data from HoloLens 2 include 3D coordinates of all virtual contents, the status of marker tracking, and the rendering speed of the virtual scenes. Simultaneously, the Zaber system provides absolute displacement data for each axis of motion. All datasets are synchronized at 60 FPS to maintain consistency with the frame rate of the captured AR images, ensuring precise temporal alignment across all measurements. Prior to each experimental session, the testbench was calibrated as follows. The OptiTrack system was calibrated using the manufacturer’s wand calibration procedure, achieving a mean 3D residual error of approximately 0.229 mm as reported by the Motive software.

2.4 Data analysis

Prior work has proposed various methodologies for evaluating AR image registration including user-based assessments using calipers (Eom et al., 2022; Van Gestel et al., 2024; Cui et al., 2023) or task performance metrics (Doughty and Ghugre, 2022; Qian et al., 2019; Li et al., 2022; Mamone et al., 2021). However, these approaches are susceptible to human error and do not provide quantitative, objective characterizations of spatial trajectory accuracy or temporal latency. In this study, we evaluate three quantitative metrics: (1) jitter, which quantifies frame-to-frame instability of the virtual point position at stationary conditions; (2) DTW-based RMSE, which quantifies spatial trajectory accuracy during motion by comparing the 1D centroid coordinate time series of the physical GT light and the virtual point; and (3) cross-correlation latency, which measures the end-to-end phase shift between the GT and virtual point motions. All metrics operate on 1D time series of centroid pixel coordinates extracted from FLIR camera frames, subsequently converted to millimeters using a pixel-to-mm calibration factor derived from the known marker dimensions and HMD-to-marker distance at each experimental condition.

2.4.1 Metric 1: jitter

To quantify the stability of the AR registration at stationary positions, we evaluate a jitter, defined as the 3D Euclidean distance between the virtual point coordinates reported by the HoloLens 2 in consecutive rendering frames (Teather et al., 2009).where p ( t ) ∈ R 3 is the 3D world coordinate of the virtual point at frame t, as reported by the HoloLens 2 AR application. Jitter is reported as mean J over all frames in each condition, expressed in millimeters. When both the AR HMD and the target marker remain stationary, jitter directly reflects the performance of the marker tracking system.

J t = ‖ p t − p t − 1 ‖ 2 , (2)

2.4.2 Metric 2: DTW-based spatial RMSE

To quantify spatial accuracy during motion, we employ DTW to compare the 1D centroid coordinate time series of the physical GT light, x GT ( t ), and the virtual point, x VIR ( t ), along the axis of motion, where t denotes the frame index and coordinates are expressed in pixels (subsequently converted to millimeters). DTW finds the optimal warping path W = { ( i 1 , j 1 ) , … , ( i K , j K ) } that minimizes the cumulative Euclidean distance between aligned elements of the two sequences while satisfying monotonicity and boundary constraints (Senin, 2008). This warping path accommodates temporal distortions such as rendering lag and speed-dependent phase shifts. DTW has been applied in various domains, including the evaluation of human motions during AR-assisted surgical training (Lahanas et al., 2011), human-robot collaboration (Cao et al., 2019), rehabilitation (Escalona et al., 2020), and motor skill acquisition (Liu et al., 2023). However, its use as an evaluation metric for AR registration remains unexplored.

We compute the warping path using the TSlearn package in Python (Figure 4b). The spatial accuracy metric is the RMSE of this optimal warping path, computed as the root mean squared difference between the GT and virtual point coordinates at each matched pair in the path:where K is the total number of matched coordinate pairs in the warping path. A lower R M S E DTW indicates that the virtual point trajectory more closely follows the GT trajectory, accounting for any temporal offset. An example of DTW path matching between a GT sinusoidal motion and the corresponding virtual point trajectory is shown in Figures 4a,b, illustrating noticeable latency and step-like positional jumps caused by inaccurate tracking detections. The R M S E DTW was computed independently for each period, and the mean and standard deviation (SD) across periods are reported throughout Section 3. The R M S E DTW is initially computed in pixels from the centroid coordinate differences at matched DTW path pairs, then converted to millimeters.

R M S E DTW = 1 K ∑ k = 1 K x GT i k − x VIR j k 2 , (3)

2.4.3 Metric 3: cross correlation latency

While the DTW path matching quantifies the spatial accuracy of marker-based registration relative to the GT, it does not isolate the temporal phase shift (latency) between the GT and virtual point motions. To measure this latency, we compute the mean-subtracted cross-correlation between the GT centroid coordinate time series x GT ( t ) and the virtual point centroid coordinate time series x VIR ( t ):where τ is the lag in frames, N is the total number of frames, and x ̄ GT and x ̄ VIR are the global means of the respective time series computed over all N frames. Mean subtraction removes DC offset differences between the two signals prior to correlation. The latency lag is identified as the first local maximum of C ( τ ) at negative lag, i.e., τ ̂ = max { τ < 0 : τ ∈ P }, where P = { τ : C ( τ ) is a local maximum } is the set of all local peak locations of C ( τ ) (Figures 4c,d). This peak-selection strategy isolates the primary rendering-induced phase shift from other periodic peaks that arise from the sinusoidal nature of the motion. The latency is then computed as L = τ ̂ × 16.6 m s, based on the 60 FPS FLIR camera capture rate. The measured latency L reflects the end-to-end delay in the AR rendering pipeline as observed in the captured display images, including contributions from marker detection, coordinate data transmission over the wireless network, rendering computation, and the HMD display pipeline.

C τ = ∑ t = 1 N x GT t − x ̄ GT x VIR t + τ − x ̄ VIR , (4)

2.5 Graphical user interface

We developed a graphical user interface (GUI) to facilitate the experiment setup. This GUI allows for precise manipulation of speed and acceleration settings on the Zaber system to automate motion. It also enables real-time monitoring of AR images captured by the FLIR camera, which are processed using the OpenCV package for Python. Additionally, the GUI retrieves real-time data from the AR HMD, including the 3D coordinates of virtual points, marker tracking status (e.g., tracking status, frame rate, and timestamps), via a wireless network. The GUI uses multi-threading to initiate the data acquisitions and synchronized data collection from the FLIR camera and Zaber automation systems at a consistent frame rate of 60 FPS. However, the frame rate of the AR HMD data varies depending on the conditions of rendering loads. Therefore, we collect all data with timestamps to align AR HMD data with the other systems. The collected experimental data are stored for post-experiment analysis.

3 Experimental results

Table 1 shows a list of all experimental conditions. We investigated the stability of AR registration under three scenarios: stationary positions (i.e., no movement of the AR HMD or the target marker), translational movements of the AR HMD and the target marker with variations in rendering conditions and motion speeds, and the rotational movements of the target marker. All translational and rotational movements were automated as approximately sinusoidal motions with fixed amplitudes: 50 mm for translational movements and 10° for rotational movements for both the motion of the AR HMD and the target marker. Each experimental condition was repeated with an ARToolkit image-based marker, noted as ART in the results, a high-feature image-based marker with Vuforia tracking, HFM, and an IR reflective marker with OptiTrack tracking, OT.

TABLE 1. Experimental conditions for all trials repeated for each marker type (ART, HFM, OT). D: HMD-to-marker distance; S: translational speed (peak-to-peak amplitude 100 mm, linear acceleration 85 mm/s 2); R: rotational speed of HMD (peak-to-peak amplitude 10°, linear acceleration 50°/s 2).

3.1 Rendering conditions

While HoloLens 2 tries to maintain a consistent rendering speed of 60 FPS during operation, the rendering performance can degrade significantly when displaying a high-resolution 3D anatomical model in the virtual scene. Reduced rendering speed introduces lag in updating virtual content to align with newly tracked positions and orientations, which may disrupt surgeons’ perceptions or distract them from focusing on surgical tasks. For this study, we manually segmented the anatomical model from an anonymous patient’s CT scan, creating layers that include the skull, brain, and brain ventricular system (Figure 5) (Eom et al., 2025).

To evaluate the impact of rendering load on performance, we experimented with various anatomical model configurations, as described in Table 2 with their respective model numbers, M. The full model including the brain ventricle, brain, and skull was tested with both opaque, M1 (Figure 5a), and transparent textures, M2, (Figure 5b), comprising approximately 2.7 million polygons. The simplest model, consisting only of the brain ventricle, M7, contained approximately 390,000 polygons (Figure 5c). Additionally, intermediate variations of the full model with reduced resolutions (M3 to M6) were created. The polygon counts for all models are reported in Figure 6.

TABLE 2. Rendering configurations used in experiments. Resolution indicates the fraction of full polygon count retained after down-sampling. NA: not applicable.

From the experiments, we discovered that the AR HMD maintained a rendering speed of 60 FPS for M6, M7, and M8 conditions. However, as the resolution of the model (i.e., number of polygons in Figure 6) increased, the rendering speed progressively deteriorated down to as low as 11 FPS for the M2 (Figure 6). The rendering frame rate was slightly better for the full resolution model with opaque texture (M1: 13.4 FPS) compared to the transparent texture (M2: 11.7 FPS), however the difference was unperceivable. While 30–60 FPS is considered optimal for HoloLens 2 (Microsoft, 2021), a lower frame rate accompanied by significant lag in updating virtual content can lead to a suboptimal user experience. Models with stacked anatomical layers, such as the skull combined with internal brain systems (Figure 5b), require higher computational resources, further straining the AR HMD’s rendering capabilities.

3.2 Stationary positions

Analysis of the 3D coordinate data from HoloLens 2 revealed that the mean jitter computed using Equation 2 in AR registration increased with greater distances between the AR HMD and the target marker, as shown in Figure 7. Particularly, image-based marker AR registration (i.e., the ARToolkit marker, ART, and the high-feature marker, HFM) exhibited higher jitter compared to IR reflective marker-based registration. Furthermore, rendering the full model (M2) resulted in a greater jitter than conditions without any model rendering (M8). The HMD-to-marker distances evaluated (800–1400 mm) reflect the range of working distances between a surgeon wearing an AR HMD and a patient or surgical instrument in a typical operating room configuration, where the surgeon may stand close to the patient (D1, 800 mm) or at an extended distance during pre-operative planning or instrument preparation (D3, 1400 mm).

For the M8 condition, the jitter for the ART marker gradually increased from D1 to D3, reaching 1.90 mm of jitter per frame at D3. In contrast, the HFM marker tracking maintained consistent jitter levels until D2, but increased to 1.78 mm at D3 (Figure 7a). A similar trend was observed under the M2 condition, but the jitter at D3 was slightly higher for both the ART (4.16 mm) and the HFM (3.19 mm) markers. On the other hand, the IR reflective marker (OT) demonstrated consistent jitter between 0.05 mm and 0.10 mm across all distances and rendering conditions (Figure 7b). The observed discrepancy in the jitter between image-based and IR reflective markers can be attributed to the FoV of the tracking systems. The OptiTrack system, configured to cover a full 360-degrees of view around our testbench, maintained consistent jitter performance regardless of HMD-to-marker distance. This aligns with the calibration results for the OptiTrack system, which reported a 3D projection error of approximately 0.229 mm. In contrast, image-based marker tracking relies on the image captured from the front-facing camera of the AR HMD, resulting in poor accuracy of registration (i.e., high jitter) as the distance increases. The ART marker exhibited a more gradual increase in jitter compared to the HFM marker due to the limitations of Vuforia marker detection on binary marker patterns. However, beyond a critical threshold (i.e., farther than 1400 mm, D3), the jitter levels for both ART and HFM markers became comparable.

Although the HFM marker exhibited lower jitter than the ART marker at D3, the percentage of tracking losses was higher: 13.1% for M8 and 16% for M2, compared to 7.31% (M8) and 5.89% (M2) for the ART marker (Table 3). This discrepancy is due to the higher number of feature points required for tracking the HFM marker, whereas the ART marker relies on simpler binary patterns. However, the lower percentage of tracking losses for the ART marker came at the cost of higher jitter, indicating more frequent inaccurate detections of positions and orientations. These inaccuracies can cause sudden jumps or judder in virtual content, potentially distracting or interrupting surgeons during operations. While surgeons are unlikely to work at distances beyond the threshold (e.g., farther than 1400 mm from registered objects such as patients or surgical instruments), environmental factors like lighting conditions or occlusion of markers by other objects or personnel in the operating room could still impact the stability of AR registration.

TABLE 3. Percentage of frames with lost tracking (%) at stationary positions across HMD-to-marker distances D1–D3, under no-model (M8) and full-resolution transparent model (M2) rendering.

3.3 Translational movement

Using the Zaber motion control, we automated a single-axis sinusoidal motion for each target marker and AR HMD to investigate the accuracy and stability of the AR registration. To quantify spatial accuracy, we calculated the R M S E _ DTW using Equation 3 between the GT centroid coordinate time series and each tracking system’s virtual point centroid coordinate time series, computed per sinusoidal period and converted to millimeters as described in Section 2.4.2. The mean and SD across periods are reported as shown in Figure 8; as each experimental condition was conducted in a single trial, the SD reflects within-trial variability across sinusoidal periods rather than across-trial reproducibility. Additionally, we evaluated the latency of each target marker relative to the GT, corresponding to the delay in the motion of registered virtual points, using cross-correlation analysis as shown in Figure 9. The latency of the HMD was excluded from the results due to the reporting of negligible numbers close to zeros.

3.3.1 Variations in the rendering condition

The R M S E DTW results of the translational movement of the target marker and the AR HMD are shown in Figures 8a,e, correspondingly. The translational movement followed an approximately sinusoidal position trajectory with a fixed peak-to-peak displacement of 100 mm (±50 mm amplitude), executed via repeated back-and-forth motion at a peak speed of 50 mm/s and a trapezoidal velocity profile acceleration of 85 mm/s 2 (S3, Table 1).

For the translational movement of the target marker, the R M S E DTW increased proportionally with the rendering computation demands (i.e., M8, M7, M4, and M2 in order as shown in Figure 8a). Among the tracking systems, the ART exhibited the highest R M S E DTW across all rendering conditions compared to the HFM and OT, 5.34 ± 0.53 mm at M8-S3 to 16.07 ± 1.61 mm at M2-S3. This higher R M S E DTW, coupled with greater variability, was attributed to the ART marker’s lower tracking performance, leading to inaccuracies in the rendered virtual point locations (e.g., sudden positional jumps). These findings are consistent with prior studies reporting higher displacement errors for image-based marker registration compared to IR reflective marker-based registration (Eom et al., 2022; Costa et al., 2024). The HFM marker exhibited substantially lower R M S E DTW, ranging from 0.83 ± 0.07 mm at M8-S3 to 2.72 ± 0.21 mm at M2-S3. Interestingly, the DTW distances for HFM and OT remained similar between M8-S3 (no model rendering, 0.83 ± 0.07 mm and 0.75 ± 0.81 mm, respectively) and M7-S3 (ventricle model rendering, 0.86 ± 0.07 mm and 0.42 ± 0.03 mm, respectively). This indicates that selective model rendering tailored to specific surgical applications can mitigate the impact on AR image registration.

On the other hand, for the translational motion of the AR HMD, the R M S E DTW remained consistent across all rendering conditions. The mean R M S E DTW values for ART, HFM, and OT ranged between 0.49 mm and 0.56 mm (Figure 8e), significantly lower than those reported for the target marker (Figure 8a). This consistency is likely due to the AR HMD’s built-in temporal warping function, which adjusts the rendered image in real-time based on head movements to reduce latency and maintain alignment of virtual objects with the real-world scene (Itoh et al., 2021; Microsoft, 2021; Van Waveren, 2016). This internal compensation ensured that the alignment of tracked virtual points closely followed the GT, even during HMD motion. Similar trends were observed for the translational motion of the HMD under different target motion speeds (Section 3.3.2) and for the rotational motion of the HMD (Section 3.4), as discussed in subsequent sections.

Latency computed using Equation 4 results for the translational motion of the target marker are shown in Figure 9a. Under M8 (no model rendering), all tracking systems maintained latencies below 250 ms for updating tracked virtual points relative to the GT. When rendering the ventricle model (M7), latency for HFM and OT remained stable at approximately 133 ms and 83 ms, respectively, while ART latency increased slightly to 300 ms. The highest latency was observed under M2 (full-resolution model rendering), where ART, HFM, and OT caused 600 ms, 417 ms, and 283 ms, respectively, to update tracked virtual points. High latency in updating registered images is critical in surgical applications, particularly with the movement of surgical instruments. The delays in the AR HMD display (i.e., lag of updating virtual contents in the surgeon’s AR view) can lead to distractions, discomfort, motion sickness, and errors during surgical procedures. These findings underscore the importance of rendering optimization in AR image registration. The R M S E DTW and latency results highlight that rendering conditions significantly influence registration performance, emphasizing the need for efficient rendering of anatomical models that align with the computational capabilities of the AR HMD for surgical applications.

3.3.2 Variations in the speed of the target motion

The R M S E DTW results of the translational movement of the target marker and the AR HMD for M8 and M2 conditions are shown in Figures 8b,c, then Figures 8f,g, correspondingly. The translational displacement was consistent at 100 mm.

Overall, the OT demonstrated stable R M S E DTW mean values regardless of the speed of motion, maintaining approximately 0.52 ± 0.04 mm under no-model rendering (M8) and approximately 2.21 ± 0.15 mm under full-resolution model rendering (M2). This consistency can be attributed to the 120 FPS tracking speed of the OptiTrack camera, which ensured reliable image registration. However, the R M S E DTW for the image-based marker tracking systems (ART and HFM) increased with faster motion. R M S E DTW of the ART rose from 1.30 ± 0.65 mm at M8-S1 (10 mm/s) to 5.19 ± 0.58 mm at M8-S2 (30 mm/s), continuing to increase to 9.27 ± 0.66 mm at M8-S4 (100 mm/s). This indicates that even a slight increase in motion speed (from S1 to S2) negatively impacted the registration accuracy of ART. On the other hand, R M S E DTW of the HFM remained within the 0.26–0.83 mm range from M8-S1 to M8-S3 but increased to 7.42 ± 0.49 mm at M8-S4. With the higher rendering condition (M2), R M S E DTW values were substantially higher for both ART and HFM. For M2-S4, the ART reached 25.25 ± 1.33 mm, while the HFM recorded 7.35 ± 0.37 mm. These findings illustrate the threshold motion speeds for each image-based marker. While the HFM maintained optimal accuracy up to 100 mm/s, the ART’s accuracy began to degrade at 30 mm/s. Marker selection and pattern design therefore play a critical role in optimizing marker-based image registration for surgical applications that involve variable speed of surgical tool movements. Future studies should evaluate more realistic 3D surgical tool movements using open-source pose datasets (Wang et al., 2023; Hein et al., 2025) to assess image registration in clinical settings.

In contrast, the R M S E DTW for the AR HMD was consistently low across all speeds of motion for both M8 and M2 conditions. Notably, R M S E DTW was highest at the slowest motion speed (S1, 10 mm/s). For M8-S1, R M S E DTW values were 0.29 ± 0.06 mm for ART, 0.26 ± 0.06 mm for HFM, and 0.18 ± 0.04 mm for OT (Figure 8f). Similar trends were observed for M2-S1, where ART, HFM, and OT recorded RMSE values of 0.32 ± 0.06 mm, 0.28 ± 0.06 mm, and 0.16 ± 0.03 mm, respectively (Figure 8g). As motion speed increased to S2 (30 mm/s), R M S E DTW increased slightly and remained stable at higher speeds (S3 and S4), ranging from 0.41 mm to 0.60 mm across all marker types and rendering conditions. This stability in R M S E DTW at faster speeds likely reflects the effects of temporal warping on the AR HMD. The built-in time-warping mechanism dynamically adjusts rendered images to align with real-time head movements, reducing misalignment and improving registration accuracy at higher motion speeds. Further studies using real-time data from the AR HMD’s onboard inertial measurement unit (e.g., HoloLens 2 sensor streaming via research mode (Ungureanu et al., 2020)) could provide insights into the temporal-warping mechanism and its role in image registration.

The latency results for the translational movement of the target marker are depicted in Figure 9b (M8) and 9c (M2). Similar to the R M S E DTW trends, latency for image-based marker tracking systems increased with faster motion speeds. The ART exhibited higher latency (183.3 ms at M8-S1 and 333.3 ms at M8-S4) than the HFM (133.3 ms at M8-S1 and 233.3 ms at M8-S4, Figure 9b). Under the M2 condition, latency more than doubled compared to M8 for both ART (500 ms at M2-S1, 733.3 ms at M2-S4) and HFM (416.7 ms at M2-S1, 516.7 ms at M2-S4, Figure 9c). Unlike rendering conditions, controlling the speed and motion of surgical tools is often challenging in surgical applications. However, latency above 200 ms can significantly impact performance in AR-guided surgeries (Khan et al., 2023), leading to distractions, discomfort, and errors during tasks. These results suggest that IR reflective marker-based registration, which exhibited lower latency, is more suitable for surgical applications involving fast-moving instruments than image-based marker registration.

In summary, the findings underscore the importance of optimizing rendering conditions and selecting appropriate marker types for AR image registration. While ART and HFM markers showed variable performance under different conditions, OT consistently delivered high accuracy and low latency, making it a reliable choice for surgical applications.

3.4 Rotational movement

Surgeons often shift their attention between the surgical scene and other visual contexts, such as external monitors, primarily by rotating their heads. To better understand the implications of such rotational movements, we analyzed the performance of an AR HMD under various angular speeds while maintaining a consistent angular displacement (Table 1). The R M S E DTW results for rotational movements are presented in Figures 8d,h. For these experiments, the angular displacement was set to 10°, with an angular acceleration of 50 ° / s 2. The latency of the rotational movement was negligible, as confirmed by cross-correlation analysis, and was therefore excluded from the results. The rotational speeds evaluated (5°–20°/s) reflect the range of head rotation speeds during surgical attention shifts, such as glancing between the surgical field and an external monitor, which are relevant to the practical use of AR HMDs in the operating room.

The R M S E DTW for rotational movements of the HMD was slightly lower compared to translational movements. Under the slowest rotational speed condition (M8-R1), the R M S E DTW values across all marker tracking systems were below 0.52 mm, indicating high similarity between the AR registration and the GT (0.51 ± 0.04 mm for ART; 0.48 ± 0.03 mm for HFM; 0.51 ± 0.04 mm for OT, as shown in Figure 8d). As the angular speed increased, the R M S E DTW values also increased, with M8-R4 results showing 1.34 ± 0.04 mm for ART, 0.99 ± 0.03 mm for HFM, and 0.82 ± 0.03 mm for OT. Similarly, under the M2 rendering condition, the R M S E DTW values followed the same trend, increasing as the angular speed progressed from M2-R1 to M2-R4 (Figure 8h). For instance, at M2-R1, the R M S E DTW values were 0.41 ± 0.03 mm for ART, 0.50 ± 0.04 mm for HFM, and 0.35 ± 0.03 mm for OT. At the highest rotational speed (M2-R4), the R M S E DTW increased to 1.94 ± 0.07 mm for ART, 0.96 ± 0.03 mm for HFM, and 0.84 ± 0.03 mm for OT. These findings suggest that higher angular speeds reduce the effectiveness of temporal warping in rendering adjustments, leading to increased R M S E DTW. This contrasts with the stable R M S E DTW observed during translational HMD movements across all speeds, highlighting distinct behaviors of temporal warping in different types of head motion. For example, a surgeon wearing the AR HMD and moving throughout the operating room to re-align with a registered virtual model may experience different registration accuracy than one remaining stationary and rotating their head to switch attention. These contrasting impacts emphasize the need for further investigation into head movement patterns and their characterization using the onboard inertial measurement unit sensor to optimize AR performance for surgical applications.

4 Discussion

This work introduces a testbench and quantitative approach for evaluating marker-based AR registration on the HoloLens 2 optical see-through HMD. By synchronizing AR images captured through the HMD display with GT motion from precision control and headset-reported tracking data, we obtain objective measurements of spatial error and temporal latency. The combination of jitter, DTW-based RMSE, and cross-correlation–derived latency provides a richer characterization of registration performance than single-point displacement errors alone: jitter isolates frame-to-frame instability at stationary positions, DTW-based RMSE quantifies spatial trajectory similarity while accommodating temporal distortions, and cross-correlation latency directly measures the delay between physical and virtual motion. Together, these metrics enable objective comparison across marker types, rendering configurations, and motion profiles on the HoloLens 2 platform without relying on subjective user assessments.

4.1 Objective quantitative evaluation of AR image registration

Higher motion speed and rendering loads significantly affected the spatial accuracy and latency of image registration, with binary-pattern markers (ART) exhibiting more pronounced deterioration than high-feature markers (HFM) during translational motion of the target marker. In contrast, IR reflective marker–based registration using an external camera system demonstrated more consistent spatial accuracy and lower latency across similar conditions, reflecting its higher temporal resolution and independence from the HMD’s front-facing camera pipeline. Notably, the translation and rotation of the HMD were largely unaffected by motion speed or rendering load, resulting in stable spatial accuracy and near-zero latency across all marker types. This work demonstrates contrasting system behaviors: registration accuracy for mobile objects such as surgical instruments is highly sensitive to system configuration, while HMD motion is effectively compensated. These differences highlight the value of the objective, device-centric approach over user-based metrics. These contrasting behaviors are attributable to HoloLens 2-specific features, in particular, its onboard reprojection pipeline, and may differ on other AR HMD platforms that implement different temporal compensation strategies or none at all. These differences highlight the value of the objective, device-centric approach over user-based metrics, and underscore the need to evaluate each AR HMD platform independently.

4.2 Implications for marker-based AR guidance in surgery

The results carry several practical implications for the design and deployment of marker-based AR guidance in surgery, using the HoloLens 2 platform or similar optical see-through AR HMDs with comparable rendering and temporal warping architectures. IR reflective marker–based registration consistently outperformed image-based marker approaches in spatial accuracy and latency across all motion and rendering conditions, and is frequently used for tracking fast-moving objects such as surgical instruments. Among image-based marker types, high-feature markers (HFM) are generally more robust than binary-pattern markers (ART) at higher speeds and greater distances, though they can be more susceptible to tracking loss under marginal viewing conditions. For applications where image-based markers are required, selecting HFM over ART and limiting the HMD-to-marker distance are effective strategies for preserving registration accuracy.

Rendering load emerged as a significant and modifiable determinant of registration performance. Reducing polygon count or selectively rendering only the anatomical structures relevant to the current surgical task as virtual contents, rather than displaying a full stack of embedded anatomies, can substantially improve frame rate and thereby limit additional spatial error and latency. The observed latencies under full-resolution rendering on the HoloLens 2 (up to 733 ms for image-based markers) far exceed the threshold above which AR-guided surgical performance has been shown to degrade (Khan et al., 2023), underscoring the clinical importance of rendering optimization. In contrast, registration accuracy during HMD motion was largely insensitive to rendering load and motion speed, reflecting the headset’s built-in temporal warping. Taken together, these findings provide quantitative guidance for developers and clinical users in selecting marker types, rendering configurations, and operating conditions that keep spatial and temporal errors within acceptable bounds for specific surgical procedures.

4.3 Limitations

Our experiments are restricted to rigid, marker-based registration, evaluated in a controlled laboratory environment under pre-defined translational and rotational trajectories. While this reflects many current use cases such as neurosurgery and orthopedics, it does not capture non-rigid scenarios such as soft-tissue deformation, brain shift, or organ motion during surgery. Although the recent advances in deep-learning and depth-based markerless AR registration (Gsaxner et al., 2021b; Kim et al., 2023; Yang et al., 2026) could address the registration of non-rigid anatomies, these approaches still require defined GT sources and error metrics for more objective quantitative measurements of registration accuracy. Furthermore, while the controlled environmental settings do not fully represent the complexity of the operating room, the reported errors represent lower bounds under idealized conditions: dynamic lighting, instrument and hand occlusions, specular reflections, and cluttered backgrounds in the operating room are expected to degrade marker detection robustness. Future work will integrate these perturbations, utilize clinically realistic instrument trajectories from surgical pose datasets (Wang et al., 2023; Hein et al., 2025), and evaluate the approach against cadaveric or clinical data.

Furthermore, our measurements are based on a monocular camera positioned to approximate the location of the user’s right eye, following standardized HMD measurement procedures (International Committee for Display Metrology, 2023; International Electrotechnical Commission, 2020). While this approach provides objective, reproducible device-centric measurements, positioning the camera at the right-eye location introduces a parallax-dependent measurement bias relative to the cyclopean or left-eye perspective, dependent on the horizontal offset between the camera and the HMD optical center, the lateral position of the target marker relative to the HMD midline, and the HMD-to-marker distance. In our experiments, the target marker was centered along the HMD midline by design and all motion experiments were conducted at a fixed HMD-to-marker distance of 800 mm (D1), minimizing this bias under our experimental conditions; furthermore, since the virtual points and the physical GT fiber optic light are co-located at the same depth in the AR scene, depth-induced parallax is not a confounding factor in the current experiments. In general use cases where the marker is not centered or the distance varies, this bias could become significant and should be corrected. This approach does not account for inter-user variability in eye position, inter-pupillary distance, or eye relief. The reported spatial errors should therefore be regarded as a lower bound on perceptual misregistration, with user-specific optical and perceptual factors expected to increase the discrepancy between virtual and physical content as perceived by individual surgeons. Future work should complement device-centric stereoscopic measurements of 3D accuracy using a representative sample of surgeons to quantify this additional error under realistic task conditions.

Similarly, the one-time coordinate system alignment between the HoloLens 2 and OptiTrack world coordinate systems relies on Vuforia marker detection for the T F H transformation component, which introduces a fixed systematic offset into the OT evaluation results (see Section 2.2.2). This error was mitigated by the experimental design: the alignment was performed once per session under static conditions using a large reference marker at D1 (800 mm), where Vuforia detection is most stable. Furthermore, because the DTW-based RMSE and cross-correlation metrics evaluate relative motion about the mean positions rather than requiring exact absolute spatial localization, this static offset does not affect the quantitative results or the validity of the evaluation methodology. Future implementations should replace this step with an independent HMD pose determination method such as using the HoloLens 2 onboard IR depth sensor for spatial mapping, or mounting a dedicated IR reflective marker cluster rigidly on the HMD and tracking it with the OptiTrack system, to eliminate Vuforia-dependent error from the coordinate alignment procedure entirely.

In this study, the controlled experimental conditions of a fixed HMD, single-axis constrained motion, and a monocular right-eye camera, are deliberate methodological choices designed to isolate individual sources of registration error so that each can be characterized independently, consistent with standard practice for establishing baseline performance metrics (International Committee for Display Metrology, 2023; International Electrotechnical Commission 2020). The HoloLens 2 spatial tracking pipeline remains fully active throughout all experiments; the fixed mount eliminates only the large-scale SLAM drift that accumulates over free-roaming displacements, while rendering-pipeline and marker-detection errors are still captured. The reported spatial error values therefore represent lower bounds on registration error in clinical use, where additional factors such as SLAM drift, dynamic lighting, marker occlusion, and environmental variability would further increase errors. The translational and rotational motion profiles evaluated here are representative of the controlled, deliberate head and equipment movements that occur during critical phases of procedures such as neurosurgery, orthopedics, stereotaxic, and robotic-assisted surgery. The quantitative results are specific to the HoloLens 2 and its associated software stack (Unity 2022.3.19f1, Vuforia Engine v10.25, OptiTrack Motive), but the evaluation methodology is designed to be applicable to other optical see-through AR HMD platforms.

Finally, depth perception in AR HMD applications is further influenced by individual eye parameters and display technology. HMDs utilizing waveguide-based optical combiners typically display virtual images at optical infinity (Chen et al., 2021), where perceived depth is governed primarily by binocular disparity rather than accommodation cues and may require adaptive calibration for individual users (Han et al., 2025). The HoloLens 2 incorporates additional optical elements to position the virtual image at approximately 2 m focal distance (Microsoft, 2024); however, binocular disparity remains the dominant depth cue over accommodation at this distance. These perceptual factors could lead to a discrepancy between the reported spatial errors and individual user experience and should be considered when comparing registration accuracy or evaluating depth-related registration performance across different HMD platforms in clinical settings.

5 Conclusion

Objective, quantitative evaluation of AR image registration accuracy is critical for surgical applications where sub-millimeter to millimeter precision is required, yet standardized assessment methodologies have remained lacking. To address this gap, we presented a testbench and quantitative approach for evaluating the spatial and temporal accuracy of marker-based medical image registration in optical see-through AR HMDs. By capturing AR images directly through the HMD display with a calibrated high-resolution camera and synchronizing them with ground-truth motion from a precision controller, the approach enables objective, repeatable assessment of registration performance without relying on subjective user measurements. The proposed time-series metrics, including jitter, DTW-based RMSE, and cross-correlation–derived latency, jointly characterize the stability, spatial fidelity, and temporal lag of image-based and IR reflective marker–based registrations, and reveal that registration accuracy for mobile objects such as surgical instruments is more sensitive to marker type, rendering load, motion profile, and viewing distance than registration to relatively static anatomy.

Future work will extend this evaluation methodology in several directions. The monocular camera setup can be expanded to a binocular stereo configuration with two calibrated cameras, enabling 3D spatial accuracy characterization including depth displacement measurement by triangulation and full perceptual registration error assessment incorporating both lateral and depth components. The methodology can also be extended to more complex and random 3D motion profiles using robotic motion control, and controlled environmental perturbations such as varied ambient lighting and occlusions to better approximate operating room conditions. As the evaluation methodology is designed to be platform-agnostic, the same standardized metrics can be applied to additional AR HMD platforms with appropriate adaptation of the tracking implementation, enabling characterization of different optical architectures under consistent evaluation conditions.

Generative AI statement

The author(s) declare that Generative AI was used in the creation of this manuscript. The use of artificial intelligence (AI) in the preparation of this manuscript was limited solely to proofreading and language refinement. No AI tools were used for data analysis, content generation, interpretation of results, or drawing scientific conclusions.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Author disclaimer

This article reflects the views of the authors and does not represent the views or policy of the U.S. Food and Drug Administration, the Department of Health and Human Services, or the U.S. Government. The mention of commercial products, their sources, or their use in connection with material reported herein is not to be construed as either an actual or implied endorsement of such products by the Department of Health and Human Services.

Footnotes

Edited by: Chao Ping Chen, Shanghai Jiao Tong University, China

Reviewed by: Laura Cercenelli, University of Bologna, Italy

Yan Zhang, Shanghai Jiao Tong University, China