Simulation-Oriented Parameterization of Virtual Human Head-Orientation Behavior Using Wearable IMU Data

Abstract

Realistic head-orientation behavior contributes to the credibility of virtual human agents in scenario-based simulations. Conventional finite state machine (FSM) agents often rely on manually defined perception rules, producing repetitive head motion patterns that do not reflect action-dependent human behavior. This study derives action-specific head-orientation parameters from head-mounted inertial measurement unit (IMU) recordings collected from 20 participants across six actions and 600 recording sessions. Preprocessed yaw trajectories were segmented into steady-orientation, scanning, and rapid-turn states and summarized using scan amplitude, scan period, directional bias, head-orientation switching interval, reorientation magnitude, and orientation-state proportion. The resulting profiles were applied to an FSM-based virtual human agent. Participant-independent validation used a 20-fold leave-one-subject-out procedure with comparisons against a fixed rule-based agent and a common stochastic agent using one pooled profile for all actions. The action-specific agent outperformed the common stochastic agent across all evaluation metrics. Compared with the rule-based agent, the action-specific agent showed lower switching interval and orientation-state proportion deviations, whereas the rule-based agent showed higher yaw-distribution similarity. Parameter-wise replacement analysis indicated that angular parameters mainly affected yaw-distribution agreement, while switching interval and orientation-state proportion governed their corresponding temporal and compositional measures. The method improves action-dependent temporal and compositional fidelity while retaining explicit FSM control.

1. Introduction

Virtual human agents are used in scenario-based simulation environments for training, decision support analysis, and behavioral evaluation. In these applications, simulation quality depends not only on high-level action selection but also on whether the agent produces plausible perception-related motion over time. Head orientation is one of the most visible cues of perception behavior because it reflects the direction in which a person appears to observe, scan, or respond to the surrounding environment [1,2].

Many virtual human agents are implemented using finite state machines (FSMs), behavior trees, or rule-based control structures [3]. These methods remain useful in simulation because their control logic is explicit, computationally light, and easy to modify. However, perception-related parameters in such models are often assigned manually. As a result, a virtual agent may follow a fixed scan cycle, change direction instantaneously, or repeat an overly regular head motion pattern. Such behavior may be sufficient for simple scenario execution, but it is less suitable when head orientation must vary with the agent’s current action.

Wearable inertial measurement units (IMUs) provide a practical way to record human head motion outside a fixed laboratory setting. A head-mounted IMU can measure yaw, pitch, and roll signals that describe temporal changes in head orientation [4,5]. IMU data have been widely used for human activity recognition and motion analysis, but activity labels alone do not provide the control variables needed for virtual agent simulation. The virtual agent also requires parameters that describe how head direction changes during an action.

Head orientation and eye gaze are related but not equivalent. Large overt changes in viewing direction generally involve head movement, whereas small gaze shifts can occur through eye movement without a corresponding change in head direction. A head-mounted IMU captures the broader directional component of orientation behavior but does not measure visual fixation or covert attention. For FSM-based virtual human simulation, this level of representation is still useful because the yaw direction of a virtual agent can be controlled directly. The practical challenge is to convert measured head-orientation trajectories into explicit parameters that can be inspected, adjusted, and reused in a virtual human model.

In the simulation model, the agent’s field of view is updated according to the head-orientation parameters. At each simulation step, surrounding objects are tested against the current field-of-view range. Objects outside this range, including objects approaching from behind, are treated as outside the agent’s current head-oriented perception field until the head direction changes. This field-of-view test is a simulation assumption based on head orientation rather than a direct estimate of human gaze allocation or visual fixation. The extracted parameters serve as head-oriented perception-control variables, not merely as descriptors of motion shape.

Wearable IMU recordings are used here to derive action-specific head-orientation profiles for FSM-based virtual agents. Preprocessed yaw trajectories are divided into three orientation states, and the resulting parameters are mapped to FSM perception-control variables. The method does not classify human activities from IMU data. It uses measured head motion to parameterize how a virtual agent updates its head direction during an action. Participant-independent leave-one-subject-out (LOSO) validation compares the action-specific agent with a fixed rule-based agent and a common stochastic agent using held-out human recordings.

2. Related Work

Research on virtual human agents has examined behavior representation, motion generation, and interaction control in simulations and virtual environments. Classical virtual human and crowd simulation studies provide the conceptual basis for representing human agents in synthetic environments [1,2]. In practical simulation systems, FSMs, behavior trees, and rule-based control structures are frequently used because they allow designers to define explicit states, transitions, and conditions [3,6,7]. Zhang et al. [7] used behavior trees for autonomous agent control under hybrid constraints. Their work supports the use of structured behavior models, but perception-related parameters still need to be specified for each simulation context.

Data-driven methods have also been used to improve motion quality and behavioral variation in virtual characters. Learning-based methods can generate complex locomotion or control policies from measured or simulated data [8]. Motion capture and virtual reality studies have used measured human behavior to improve motion synchronization and interaction realism in virtual environments [9,10]. These studies mainly address full-body motion reproduction, virtual collaboration, or perceived behavioral realism, rather than the extraction of field-of-view control parameters from wearable head-orientation data.

Wearable IMU-based motion analysis is another closely related research area. IMUs are widely used to measure human movement in unconstrained environments, and orientation signals such as yaw, pitch, and roll are suitable for describing rotational motion of the body or head [4,5,11]. Many wearable sensor studies have used IMU data for activity recognition, motion classification, or movement assessment [12,13,14]. Their outputs are generally intended for recognition or analysis rather than direct control of an FSM-based virtual agent.

Head motion has also been studied as an indicator of attention-related behavior. Prior studies have investigated attention estimation from head motion using wearable sensors, indicating that head movement can provide useful information about attention-related behavior at a coarse behavioral level [15]. Studies on gaze prediction in virtual environments have also examined head-based gaze estimation and the relationship between head and eye movements [16,17,18,19]. In virtual agent and virtual reality research, attention mechanisms, gaze behavior, limited perception fields, and behavioral realism have been studied to improve interaction quality and user experience [10,20]. Head orientation can be used as a coarse representation of overt directional behavior, but it does not directly correspond to visual fixation or eye gaze. Gaze and attention models may also require sensing modalities that are not always available in outdoor or scenario-based simulation data collection.

A direct link between wearable head motion measurements and FSM perception-control variables is still lacking. Virtual agent studies provide interpretable control structures, but their perception parameters are often manually assigned. The method converts measured head-orientation trajectories into action-specific head-orientation parameters for FSM-based virtual agent control.

Table 1. Comparison of representative research directions and the proposed method.

Research Direction	Typical Output	Limitation for This Study
Rule-based virtual agents	States, rules, behavior trees	Perception parameters are manually tuned
IMU-based activity recognition	Activity labels	Labels do not provide control parameters
Motion capture virtual characters	Reproduced motion	Full-body motion focus, limited FSM mapping
Attention/gaze-based agents	Gaze or attention rules	Often not derived from wearable head-orientation data
Proposed method	FSM control parameters	Requires action-specific IMU recordings

summarizes the role of the related research directions in relation to this study.

3. IMU-Derived Head-Orientation Parameterization for Virtual Agents

The proposed method converts head-mounted IMU recordings into action-specific head-orientation parameters for FSM-based virtual human agents. The parameters are assigned to the FSM according to the current action state.

Figure 1 shows how the procedure connects head-mounted IMU recording with virtual agent simulation. Head-mounted IMU data are collected during predefined actions and represented as yaw, pitch, and roll signals. The orientation signals are preprocessed to obtain temporally consistent head-orientation trajectories. The yaw trajectory is divided into explicit orientation states, from which action-specific parameters are extracted. The extracted parameters are mapped to FSM perception-control variables and loaded according to the agent’s current action state. The resulting orientation state and head direction are updated during simulation. The simulated behavior is evaluated against held-out human recordings and compared with the common stochastic and rule-based agents.

3.1. Head-Orientation Representation and Preprocessing

In the head-centered coordinate frame, yaw describes left–right head rotation about the vertical axis, pitch describes upward–downward rotation about the lateral axis, and roll describes lateral head tilt about the longitudinal axis. Because the simulation model controls the horizontal perception direction of the virtual agent, yaw is used as the primary signal for detecting scanning and head-orientation switching. Pitch and roll are retained as complementary orientation signals, but they are not used as the main criteria for identifying the three orientation states.

Yaw angles are unwrapped, duplicated or invalid timestamp entries are removed, and each recording is interpolated to a uniform 10 Hz time grid. The median yaw during the initial 1 s is subtracted as the session-level reference orientation. Samples exceeding the 250 deg/s artifact threshold are removed and linearly interpolated. A second-order Butterworth low-pass filter with a cutoff frequency of 2.0 Hz is then applied using forward–backward zero-phase filtering. The 250 deg/s threshold is used only for transient artifact removal, whereas rapid-turn states are identified separately using the action-specific percentile threshold. Sensitivity to nearby filter cutoffs and artifact thresholds is summarized in Appendix A.4.

Figure 2 shows a representative yaw trajectory before and after zero-phase low-pass filtering. The smoothed trajectory preserves the low-frequency head-orientation pattern associated with scanning and reorientation while reducing small sensor fluctuations. The preprocessed yaw trajectory is analyzed using 1.5 s windows with 50% overlap. The window duration provides a local interval for estimating yaw variance and directional changes while retaining the temporal resolution needed to identify short head-turn events. The overlap reduces sensitivity to the location of an event relative to the window boundary. Yaw variance and absolute yaw angular velocity are computed within each window.

3.2. Orientation-State Identification

The preprocessed yaw trajectory is segmented into three orientation states: steady-orientation, scanning, and rapid-turn. The steady-orientation state represents windows in which the yaw angle remains within a limited angular range and the angular velocity is low. The scanning state represents repeated directional changes over a bounded angular range. The rapid-turn state represents short-duration windows with high angular velocity and a relatively large yaw displacement.

States are identified using explicit threshold-based rules. For each action, the 30th percentile is used to identify the lower-motion portion of the yaw variance and absolute angular velocity distributions, whereas the 80th percentile is used to isolate short high-velocity reorientation events. These values are action-adaptive operational thresholds rather than universal physiological criteria. The use of separate lower and upper percentiles leaves an intermediate range for scanning behavior and avoids applying the same absolute angular velocity threshold to actions with different motion intensities. For participant-independent validation, these thresholds are determined from the 19 calibration participants in each LOSO fold and are then applied to the held-out participant.

A window is labeled as steady-orientation when both yaw variance and absolute yaw angular velocity are below the lower thresholds. A window is labeled as rapid-turn when the peak absolute yaw angular velocity exceeds the upper threshold. A window that does not satisfy either condition is labeled as scanning when the yaw trajectory contains at least two alternating local extrema. The overlapping windows are used for local state estimation, and the resulting window-level state information is aggregated at the recording-session level for parameter extraction.

The 15° switching threshold is used separately to identify significant changes in head direction when calculating the head-orientation switching interval. It was selected to exclude small fluctuations around the current orientation from being counted as directional switches while retaining clear changes in head direction. As with the percentile criteria, this value is an operational threshold for simulation-oriented parameter extraction rather than a physiological boundary. Sensitivity to the percentile settings, window length, switching threshold, and artifact threshold is examined in Appendix A.4.

Figure 3 illustrates an example of the resulting state segmentation. The shaded regions indicate the three orientation states assigned from the yaw angle dynamics. This segmentation converts the measured yaw trajectory into orientation-state information used for parameter extraction and FSM control. The relative duration of each state can then be compared across actions and used to set the orientation-state proportions of the virtual agent.

3.3. Parameter Extraction and FSM Mapping

The segmented yaw trajectory is converted into a compact set of head-orientation parameters.

$$A_e=\underset{t\in {\mathcal{T}}_e}{max }y\left(t\right)-\underset{t\in {\mathcal{T}}_e}{\mathrm{m}\mathrm{i}\mathrm{n} }y\left(t\right),$$

(1)

where $A_e$ denotes the scan amplitude of episode $e$, $y\left(t\right)$ is the filtered and normalized yaw angle at time $t$, and ${\mathcal{T}}_e$ is the set of time samples belonging to episode $e$. The unit of $A_e$ is degrees. For each recording session, the scan amplitude is obtained by averaging over all scanning episodes within that session.

Scan period is estimated from the temporal interval between consecutive same-direction extrema in the yaw trajectory. Directional bias is computed as the mean yaw offset from the session-level reference orientation during scanning. The head-orientation switching interval is defined as the elapsed time between consecutive significant yaw direction changes exceeding the predefined switching threshold. Reorientation magnitude is computed from rapid-turn events whose peak yaw angular velocity exceeds the action-specific upper threshold. Orientation-state proportion is obtained by normalizing the duration of each of the three states within each recording session.

Each parameter corresponds directly to an FSM perception-control variable. Scan amplitude determines the horizontal sweep range of the virtual agent’s head. Scan period controls the oscillation timing during scanning. Directional bias shifts the center of the scanning motion when an action shows a consistent directional tendency. During scanning, the center of the oscillatory yaw trajectory is defined as the current reference direction plus the action-specific directional bias. The bias is applied as an offset and is not accumulated after each scanning episode. The head-orientation switching interval determines the dwell time of the orientation state. Reorientation magnitude controls the angular displacement applied during rapid turns. Orientation-state proportion determines the relative occurrence or duration of the three orientation states.

Table 2. Mapping between extracted head-orientation parameters and FSM perception-control variables.

Parameter	Description	Role in Virtual Human Agent
Scan amplitude	Angular range of yaw during scanning behavior	Controls horizontal head sweep angle
Scan period	Temporal cycle of oscillatory head movement	Determines scanning frequency
Directional bias	Asymmetry of scanning toward left or right	Introduces preference in perception direction
Head-orientation switching interval	Time between consecutive significant head orientation changes	Governs dwell time of orientation states
Reorientation magnitude	Angular change during rapid head turns	Controls angular displacement during rapid turns
Orientation-state proportion	Relative duration of the three orientation states	Determines orientation-state composition

lists the extracted parameters and their corresponding FSM perception-control variables.

The FSM model separates action execution from head-orientation dynamics. The current action state is determined by the simulation scenario or higher-level decision logic. Within each action state, the head-orientation behavior is generated using the corresponding action-specific parameter profile. The same action state can reproduce the measured temporal variability without requiring every aspect of perception behavior to be embedded in the FSM transition rules.

During simulation, the active action state selects the corresponding parameter profile. The parameter profiles are computed before simulation and loaded as FSM control values. No online model inference is required during head-orientation updates. The controller updates the yaw direction at the simulation update rate of 10 Hz. The orientation state is then updated according to the orientation-state proportions and empirical switching interval distribution extracted from the IMU data. When the active state is steady-orientation, the head direction is maintained within a limited angular range. When the active state is scanning, the head direction oscillates around the bias-adjusted scan center according to the scan amplitude and scan period. When the active state is rapid-turn, an abrupt yaw displacement is applied according to the reorientation magnitude.

During each simulation update, the generated head orientation is used to rotate the agent’s modeled perception field. An object is considered geometrically available to the agent when its relative bearing falls within the assigned horizontal and vertical field-of-view limits. This rule represents a simulation-level visibility assumption and does not indicate measured human gaze allocation or visual fixation.

4. Data Collection and Simulation Evaluation

4.1. Head-Mounted IMU Data Collection

Head motion data were acquired using a wearable inertial measurement unit (WT9011DCL, WitMotion Inc., Shenzhen, China). The sensor integrates a triaxial accelerometer, gyroscope, and magnetometer and provides orientation estimates in yaw, pitch, and roll angles through onboard sensor fusion. The IMU was mounted on the top of a cap worn by each participant. The forward direction of the cap was defined as the reference front direction for head-orientation measurement.

Before each recording session, the sensor was zeroed and calibrated while the participant maintained a forward-facing posture. This posture provided the reference condition for sensor calibration. The calibration procedure was repeated for each recording session to reduce inter-session differences in initial heading and to limit the influence of sensor placement on the subsequent analysis. Because yaw estimation includes magnetometer information, the analysis uses session-relative rather than absolute heading. The forward-facing orientation at the beginning of the session is used as the reference direction. The resulting yaw values are interpreted as relative head orientation rather than absolute compass heading.

Twenty volunteers participated in the data collection. The participant group consisted of 16 male participants and 4 female participants, with an age range of 21–26 years. Data were collected for six actions: standing, walking, running, stationary monitoring, sitting, and conversing. For each action, every participant completed five independently recorded sessions, resulting in 100 sessions per action and 600 sessions in total. Running sessions were limited to approximately 3 min. Walking, standing, stationary monitoring, sitting, and conversing sessions lasted approximately 5–7 min depending on the recording condition. The five sessions were separate recording repetitions rather than subdivisions of a continuous recording.

In the standing condition, participants remained in a waiting posture and generally faced forward. In the walking condition, participants walked along a straight path at a natural pace. In the running condition, participants ran along a straight path without externally imposed head movement instructions. In the stationary monitoring condition, participants remained in place while slowly scanning the surrounding area as if observing the environment. In the sitting condition, participants remained seated in a relaxed posture while allowing ordinary head movements. In the conversing condition, participants engaged in free-form conversation with another person.

No explicit visual stimulus, target-following instruction, or externally imposed scanning rhythm was provided during the standing, stationary monitoring, sitting, or conversing conditions. This arrangement allowed head-orientation behavior to be recorded in action contexts relevant to virtual human simulation without imposing a predetermined scanning rhythm or target-following pattern. Data collection was conducted in an open outdoor area where participants could perform walking, running, stationary monitoring, sitting, and conversing under consistent recording conditions.

Sensor data were recorded as comma-separated values (CSV) files, and timestamp information was used to estimate the effective sampling rate. The effective sampling rate was approximately 8–10 Hz depending on the recording condition. The recorded data included orientation angles and raw inertial measurements, but the proposed method used yaw, pitch, and roll signals because the objective was to parameterize head-oriented perception behavior rather than full-body inertial dynamics.

Each independently recorded session was analyzed using overlapping windows for orientation-state identification and parameter extraction. Window-level labels and parameter values were aggregated at the recording-session level and then summarized by action. Overlapping windows were used for signal-level estimation, whereas the recording sessions served as the repeated measurement units. The five session-level estimates were averaged within each participant and action, and the resulting participant-level estimates were summarized across the 20 participants using the mean and standard deviation.

4.2. Simulation Scenario, Comparison Agents, and Participant-Independent Validation

A simulation scenario was constructed to evaluate head-orientation behavior under the predefined six action states. Each simulation condition corresponded to one of these states. Each simulation condition was run for 120 s and repeated 20 times per action. The environment contained six spatially distributed points of interest, and the same scenario configuration was applied to the proposed action-specific agent, the common stochastic agent, and the rule-based agent. The points of interest were not used as direct target-following commands. Instead, they served as spatial objects whose detectability depended on the current field-of-view direction generated by the head-orientation control model.

Action transitions were externally controlled by the scenario logic. This arrangement isolated the effect of head-orientation control from differences in high-level action selection. The evaluation focused on how the agent updated its perception direction within each action state, not on whether the agent selected different actions.

The rule-based agent was implemented as a hand-tuned FSM agent with fixed perception settings. Its head direction follows a constant scan amplitude and a fixed scan period, and rapid-turn events are generated using a predefined event probability. The rule-based agent does not use IMU-derived action-specific parameters. It is used to evaluate differences in agreement with measured head-orientation profiles.

A common stochastic agent was also included to distinguish the effect of action-specific parameterization from the effect of stochastic motion variation. This agent used the same stochastic state generation mechanism as the proposed agent, but a single parameter profile pooled across all six actions was applied to every action. The pooled profile was estimated only from the calibration participants in each LOSO fold. The comparison between the common stochastic and action-specific agents isolates the effect of using action-dependent parameter profiles.

All three agents used the same simulation duration, update rate, and points of interest for each action condition. The rule-based agent represented fixed heuristic control, the common stochastic agent represented stochastic behavior without action-specific adaptation, and the proposed agent used stochastic variation with action-specific parameter profiles. Participant-independent validation used 20 leave-one-subject-out folds constructed from the existing recordings. In each fold, recordings from 19 participants were used to determine the orientation-state identification thresholds and to construct the action-specific and common stochastic parameter profiles. The remaining participant’s five sessions for each action were used only for evaluation. Each participant served as the held-out participant in one fold.

For the parameter-wise replacement analysis, each action-specific parameter was replaced separately with the corresponding parameter from the common stochastic profile while the other parameters remained action-specific. The same LOSO evaluation procedure was used for each replacement.

4.3. Evaluation Metrics

Three metrics were used to evaluate agreement between simulated and measured head-orientation behavior. The metrics were selected to correspond to the main outputs of the proposed parameterization method: yaw angle distribution, head-orientation switching interval, and orientation-state composition.

Yaw-distribution similarity was computed using the normalized histogram intersection between the measured yaw angle distribution and the simulated yaw angle distribution. A higher value indicates closer agreement between the simulated head-orientation range and the measured reference profile.

$$D_{\mathrm{y}\mathrm{a}\mathrm{w}}={\displaystyle \sum _{k=1}^K}\mathrm{m}\mathrm{i}\mathrm{n}\left(h_k^{\left(H\right)},h_k^{\left(S\right)}\right).$$

(2)

where ${\mathrm{D}}_{\mathrm{yaw}}$ denotes the yaw-distribution similarity, $K$ is the number of yaw angle histogram bins, $k$ is the bin index, $h_k^{\left(H\right)}$ is the normalized histogram value of the measured human yaw data in the $k$-th bin, and $h_k^{\left(S\right)}$ is the normalized histogram value of the simulated yaw data in the same bin. The histograms are normalized so that the sum of bin values is one. $D_{\mathrm{yaw}}$ ranges from 0 to 1, where a larger value indicates closer agreement between the measured and simulated yaw angle distributions.

Head-orientation switching interval deviation was computed as a normalized absolute error between the mean switching interval derived from measured IMU data and that generated in simulation. A lower value indicates closer agreement with the measured temporal persistence of head-orientation states.

$$E_{switch}={\displaystyle \frac{\mid {\overline{T}}^{\left(S\right)}-{\overline{T}}^{\left(H\right)}\mid }{{\overline{T}}^{\left(H\right)}}}.$$

(3)

where $E_{\mathrm{switch}}$ denotes the normalized switching interval deviation, ${\overline{T}}^{\left(H\right)}$ is the mean head-orientation switching interval derived from the measured human IMU data, and ${\overline{T}}^{\left(S\right)}$ is the mean switching interval generated by the simulated agent. The switching interval is measured in seconds, whereas $E_{\mathrm{switch}}$ is dimensionless. A smaller value indicates closer agreement with the measured temporal persistence of head-orientation states.

Orientation-state proportion deviation was computed from the relative proportions of steady-orientation, scanning, and rapid-turn states. The metric measures the mean absolute difference between the measured and simulated orientation-state proportions. A lower value indicates that the simulated agent reproduces the measured state composition more closely.

$$E_{state}={\displaystyle \frac{1}{3}}{\displaystyle \sum _{j=1}^3}\mid p_j^{\left(S\right)}-p_j^{\left(H\right)}\mid ,$$

(4)

where $E_{\mathrm{state}}$ denotes the orientation-state proportion deviation, $j$ indexes the three orientation states, $p_j^{\left(H\right)}$ is the measured proportion of the $j$-th state in the human IMU data, and $p_j^{\left(S\right)}$ is the corresponding simulated orientation-state proportion. Each orientation-state proportion is computed by dividing the duration of the state by the total duration of the action segment. A smaller $E_{\mathrm{state}}$ indicates closer agreement between the measured and simulated orientation-state composition.

These metrics evaluate different aspects of head-orientation behavior. Yaw-distribution similarity measures agreement in directional range, switching interval deviation measures temporal persistence, and orientation-state proportion deviation measures the composition of orientation states. Appendix A summarizes the preprocessing, orientation-state identification, and simulation execution settings needed to reproduce the evaluation.

For participant-level statistical analysis, metric values were averaged across the held-out sessions within each action and then across the six actions, yielding one observation per participant and agent ($n=20$). For each metric, the proposed agent was compared separately with the rule-based and common stochastic agents. Normality of the paired differences was assessed using the Shapiro–Wilk test. A paired t-test was used when the normality assumption was satisfied. A Wilcoxon signed-rank test was used otherwise. Statistical significance was assessed at $\alpha =0.05$. Holm correction was applied to the six comparisons. Cohen’s $d_z$ was reported for paired t-tests, and rank-biserial correlation $r_{rb}$ was reported for Wilcoxon signed-rank tests. For yaw-distribution similarity, the paired difference was calculated as the proposed agent value minus the comparator value. For the two deviation metrics, it was calculated as the comparator value minus the proposed agent value because lower values indicate better agreement. The same directional convention was used for the effect sizes. Positive values favor the proposed agent, whereas negative values favor the comparator.

5. IMU-Derived Parameter Profiles and Simulation Evaluation

5.1. Extracted Action-Specific Parameters

The extracted head-orientation parameters varied across the six actions.

Table 3. Action-specific head-orientation parameter profiles extracted from IMU data.

Action	Scan Amplitude (deg)	Scan Period (s)	Directional Bias (deg)	Switching Interval (s)	Reorientation Magnitude (deg)	Steady-Orientation/Scanning/Rapid-Turn
Standing	18.2 ± 4.1	2.4 ± 0.5	0.5 ± 2.5	3.1 ± 0.8	21.5 ± 5.2	0.52/0.34/0.14
Walking	14.8 ± 3.7	1.9 ± 0.4	0.5 ± 1.3	2.2 ± 0.6	24.7 ± 6.0	0.36/0.46/0.18
Running	11.3 ± 3.2	1.6 ± 0.3	−0.6 ± 1.5	1.8 ± 0.5	28.4 ± 6.8	0.28/0.49/0.23
Stationary monitoring	25.7 ± 5.8	2.8 ± 0.6	−2.5 ± 7.0	2.5 ± 0.7	31.2 ± 7.1	0.31/0.50/0.19
Sitting	21.4 ± 4.9	2.6 ± 0.6	2.4 ± 2.7	3.6 ± 0.9	18.8 ± 4.7	0.64/0.24/0.12
Conversing	16.9 ± 4.4	2.1 ± 0.5	−2.7 ± 10.2	3.4 ± 0.8	19.6 ± 5.0	0.61/0.25/0.14

presents the resulting action-specific profiles.

Stationary monitoring has the largest scan amplitude and the longest scan period, whereas running has the smallest values for both parameters. Sitting and conversing have relatively high steady-orientation proportions, while walking and running have larger scanning proportions. Directional bias was generally small relative to scan amplitude, although stationary monitoring and conversing showed greater participant-level variation.

Table 3 summarizes the participant-level parameter estimates across the 20 participants. For LOSO evaluation, the proposed agent used fold-specific action profiles constructed only from the 19 calibration participants in each fold. Within each fold, the action state determined which action-specific profile was used.

Figure 4 shows representative stationary monitoring segments concatenated for visualization, together with the orientation-state markers derived from the original segments. The concatenated trajectory illustrates broad horizontal scanning and repeated rapid-turn events, which are consistent with the large scan amplitude and scanning proportion reported in Table 3.

5.2. Participant-Independent Simulation Comparison

The proposed action-specific agent was compared with the common stochastic and rule-based agents under the same simulation scenario. Performance was evaluated against the held-out recordings using the three agreement metrics.

Figure 5 shows representative relative yaw trajectories from the held-out human data and the three agents under the stationary monitoring condition. The common stochastic agent introduces temporal variation but applies one pooled parameter profile to all actions, whereas the proposed action-specific agent uses profiles estimated separately for each action from the calibration data. The rule-based agent follows fixed scanning parameters.

The trajectories are presented to illustrate differences in temporal structure rather than point-by-point agreement or overall superiority. The rule-based agent follows a regular periodic pattern, whereas the common stochastic and proposed agents introduce temporal variation. Quantitative agreement with the held-out recordings is evaluated separately using the three metrics reported in

Table 4. Participant-independent action-level comparison among the rule-based, common stochastic, and proposed action-specific agents.

Action	Agent	Yaw-Distribution Similarity	Switching Interval Deviation	Orientation-State Proportion Deviation
Standing	Rule-based	0.425	0.341	0.443
	Common stochastic	0.305	0.245	0.114
	Proposed action-specific	0.364	0.245	0.106
Walking	Rule-based	0.263	1.052	0.363
	Common stochastic	0.223	1.456	0.150
	Proposed action-specific	0.198	1.083	0.130
Running	Rule-based	0.176	0.111	0.347
	Common stochastic	0.184	0.529	0.190
	Proposed action-specific	0.127	0.052	0.157
Stationary monitoring	Rule-based	0.514	0.261	0.334
	Common stochastic	0.402	0.321	0.157
	Proposed action-specific	0.390	0.248	0.123
Sitting	Rule-based	0.356	0.421	0.507
	Common stochastic	0.313	0.247	0.140
	Proposed action-specific	0.381	0.204	0.096
Conversing	Rule-based	0.467	0.380	0.503
	Common stochastic	0.372	0.217	0.130
	Proposed action-specific	0.411	0.196	0.085
Participant-level mean ± SD	Rule-based	0.367 ± 0.036	0.428 ± 0.428	0.416 ± 0.041
	Common stochastic	0.300 ± 0.016	0.503 ± 0.649	0.147 ± 0.019
	Proposed action-specific	0.312 ± 0.024	0.338 ± 0.489	0.116 ± 0.022

.

Table 4 reports the participant-independent descriptive results obtained from the 20 LOSO folds. The proposed action-specific agent has a lower orientation-state proportion deviation than both comparison agents in each of the six actions. Its switching interval deviation is lower than that of the common stochastic agent in five of the six actions, equal in standing at the reported precision, and lower than that of the rule-based agent in five of the six actions.

Yaw-distribution similarity depended on the action and the agent being compared. The proposed agent showed higher yaw-distribution similarity than the common stochastic agent for standing, sitting, and conversing, whereas the rule-based agent had the highest participant-level mean. The relatively high yaw-distribution similarity of the rule-based agent does not imply closer temporal trajectory agreement, because this metric compares the marginal distribution of yaw angles rather than their temporal sequence. Fixed periodic scanning can match the yaw distribution under some action conditions even when its temporal persistence and orientation-state composition differ from the held-out recordings.

Table 5. Participant-level pairwise comparisons under leave-one-subject-out validation.

Metric	Comparator	Proposed Mean ± SD	Comparator Mean ± SD	Paired Difference [95% CI]	Holm-Adjusted $\mathit{p}$	Effect Size
Yaw-distribution similarity	Rule-based	0.312 ± 0.024	0.367 ± 0.036	−0.055 [−0.075, −0.035]	<0.001	$r_{rb}=-0.700$
	Common stochastic	0.312 ± 0.024	0.300 ± 0.016	0.012 [0.004, 0.020]	0.006	$d_z=0.690$
Head-orientation switching interval deviation	Rule-based	0.338 ± 0.489	0.428 ± 0.428	0.090 [0.042, 0.137]	0.002	$d_z=0.884$
	Common stochastic	0.338 ± 0.489	0.503 ± 0.649	0.165 [0.083, 0.247]	<0.001	$r_{rb}=0.900$
Orientation-state proportion deviation	Rule-based	0.116 ± 0.022	0.416 ± 0.041	0.300 [0.280, 0.319]	<0.001	$d_z=7.118$
	Common stochastic	0.116 ± 0.022	0.147 ± 0.019	0.030 [0.021, 0.040]	<0.001	$d_z=1.543$

reports the participant-level paired comparisons, including 95% confidence intervals for the paired differences, Holm-adjusted p-values, and effect sizes. The proposed agent had lower yaw-distribution similarity than the rule-based agent but higher similarity than the common stochastic agent. Its switching interval deviation and orientation-state proportion deviation were lower than those of both comparison agents. All six pairwise differences remained significant after Holm correction. These results indicate a metric-dependent trade-off. The rule-based agent better matches the marginal yaw distribution. The action-specific agent more closely reproduces switching persistence and orientation-state composition.

Table 6. Parameter-wise replacement results under leave-one-subject-out validation.

Action-Specific Parameter Replaced by the Corresponding Common Stochastic Profile Parameter	Yaw-Distribution Similarity	Switching Interval Deviation	Orientation-State Proportion Deviation
None: full action-specific agent	0.312	0.338	0.116
Scan amplitude	0.308	0.338	0.116
Scan period	0.311	0.338	0.116
Directional bias	0.309	0.338	0.116
Switching interval	0.315	0.497	0.119
Reorientation magnitude	0.305	0.338	0.116
Orientation-state proportion	0.307	0.339	0.146

reports the parameter-wise replacement results under the same LOSO validation procedure.

The parameter-wise replacement results indicate that scan amplitude, reorientation magnitude, and directional bias mainly affect yaw-distribution similarity, whereas the switching interval and orientation-state proportions primarily affect their corresponding temporal and compositional measures. The effect of directional bias is smaller than that of the scan amplitude and reorientation magnitude.

5.3. Limitations

Eye movements that occur without corresponding head movement are not represented by the model. The field-of-view rule was not validated as a measure of human gaze allocation or visual fixation.

The participant group and actions were selected to support simulation-oriented parameter extraction. The results should be interpreted as action-specific parameter profiles for virtual agent control rather than as population-level behavioral statistics. Additional participants, age groups, and task contexts would be required to generalize the extracted profiles beyond the experimental conditions used in this study.

The evaluation included a fixed rule-based agent and a common stochastic agent. These comparisons separate action-specific parameterization from fixed heuristic control and general stochastic variation, but they do not cover learning-based or hybrid orientation controllers. Future evaluation should examine additional agent control models and more complex multi-agent scenarios.

The simulation scenario also used a simplified environmental context. The points of interest were included as spatial objects whose detectability depended on the current field-of-view direction, but the model did not explicitly consider object saliency, occlusion, task priority, or interaction with other agents. Incorporating environmental context would allow the same parameterization method to support more detailed perception–action interactions.

The evaluation focused on agreement with measured head-orientation profiles and did not examine how update latency affects an observer’s interpretation of the virtual agent. Response timing can influence the perceived fluency and predictability of autonomous behavior [21], even when the generated motion follows an appropriate directional pattern. Future evaluation should examine controller timing together with observer-based assessments of behavioral naturalness.

6. Conclusions

This study derived action-specific head-orientation parameters from head-mounted IMU recordings and mapped them to FSM perception-control variables in a virtual human agent. Participant-independent LOSO validation showed that the action-specific agent produced lower switching interval and orientation-state proportion deviations than both comparison agents. Its yaw-distribution similarity was higher than that of the common stochastic agent but lower than that of the fixed rule-based agent. The parameter-wise replacement analysis showed that scan amplitude, reorientation magnitude, and directional bias mainly affected yaw-distribution similarity, whereas switching interval and orientation-state proportion had greater effects on their corresponding temporal and compositional measures. The main benefit of action-specific parameterization was observed in temporal persistence and orientation-state composition rather than in marginal yaw-distribution agreement.

The resulting parameter profiles provide reusable control values for FSM-based simulation without replaying individual recorded trajectories or requiring online model inference. They allow action-dependent head-orientation behavior to be incorporated while retaining explicit FSM control. Future work will examine environmental context, additional sensing modalities, and multi-agent interactions.