System Architecture for VR Yoga Therapy Platform with 6-DoF Whole-Body Avatar Tracking

Abstract

Chronic pain is a leading cause of morbidity among children and adolescents affecting 35% of the global population. Pediatric chronic pain management requires integrative health methods spanning physical and psychological subsystems through various mind-body interventions. Yoga therapy is one such method, known for its ability to improve the quality of life both physically and psychologically in chronic pain conditions. However, maintaining the clinical outcomes of personalized yoga therapy sessions at-home is challenging due to fear of movement, lack of motivation, and boredom. Virtual Reality (VR) has the potential to bridge the gap between the clinic and home by motivating engagement and mitigating pain-related anxiety or fear of movement. We developed a multi-modal algorithmic architecture for fusing real-time 3D human body pose estimation models with custom developed inverse kinematics models of physical movement to render biomechanically informed 6-DoF whole-body avatars capable of embodying an individual’s real-time yoga poses within the VR environment. Experiments conducted among control participants demonstrated superior movement tracking accuracy over existing commercial off-the-shelf avatar tracking solutions, leading to successful embodiment and engagement. These findings demonstrate the feasibility of rendering virtual avatar movements that embody complex physical poses such as those encountered in yoga therapy. The impact of this work moves the field one step closer to an interactive system to facilitate at-home individual or group yoga therapy for children with chronic pain conditions.

I. Introduction

Chronic or persistent pain is pain that carries on for long periods despite medication or treatment. Pediatric chronic pain conditions are the leading cause of morbidity among children, with an estimated 20% to 35% of children suffering from chronic pain worldwide [1, 2, 3]. Without effective management strategies, chronic pain can limit activities, impede life experiences, and contribute to long-term suffering [4, 5, 6, 7]. Children with chronic pain conditions have reported a diminished quality of life, restrictions in school activities, and higher likelihood of using pain medication and seeking medical care [3, 8]. The World Health Organization (WHO) recommends intervention models that include physical, psychological, and social therapy for children with chronic pain for effective pain management [8]. Children denied such interventions for chronic pain are left with little ability to perform basic everyday tasks and limited opportunities to engage in physical and social activities, hindering their quality of life [9, 8, 2, 4].

Recent advances in chronic pain management are being directed towards integrative mind-body rehabilitation methods to manage the three key components of chronic pain – namely the bio (movement, physiological), psycho (depression, anxiety, kinesiophobia, and catastrophizing), and social (engagement, motivation) components [10, 11, 12, 8, 13]. Integrated psychological and social activities like yoga have shown evidence of physical, psychological, and social improvement in children with chronic pain conditions [14, 15]. In fact, studies of the effects of yoga therapy have shown promise for a broad range of chronic pain conditions [13, 14, 15]. The home environment therefore assumes a critical role in achieving pediatric chronic pain rehabilitation outcomes for continuity of care and transition to a more normal life. Yet, the continuation of yoga therapy from the clinic into the home becomes challenging, with limited access to supervision, poor adherence due to short attention span and boredom, and pain-related psychological factors such as kinesiophobia (fear of movement) and catastrophizing (distorting harm) that impede rehabilitation [16, 17, 18].

Virtual Reality (VR) is known for its effectiveness in improving motivation, compliance, and pain tolerance for physical activities and exercises [19, 20]. In fact, there are already several VR platforms studied for the use of chronic pain [21, 22, 23] which offers multi-sensory stimulation for pain distraction combined with virtual embodiment to reduce akinesia. A pilot study in pediatric chronic pain rehabilitation using physical movement in virtual reality, conducted by Griffin et al., created a VR game to alter this perception by encouraging limb movement in VR where the child must move to crush fruit with their hands and feet. The study indicated that individuals felt more distracted from pain when engaged, had greater perceived mobility, and fewer clinician-observed pain behaviors during the VR game [23] compared to conventional exercise, even though the scope of the game covered only a subset of body segments. Another study points out that social interaction in VR can increase distraction and possibly provide social support for those with chronic pain conditions [24]. While these studies demonstrate feasibility in their applications to reduce perceived pain in individuals with chronic pain, they fail to provide avatar embodiment and movement fidelity measures. VR commercial applications for yoga are limited to displaying yoga poses or providing a media for mind-body meditation without person-centric, whole-body quantitative feedback or options for social/clinical interaction. Biopsychosocial aspects of yoga therapy are critical in enabling the user to reduce anxieties related to chronic pain [25, 26], such as akinesia and catastrophizing, by transporting their real body into a virtual world as an avatar [26, 27]. While there exist few examples of accurate solutions for full body tracking in VR, [28, 29] those that do require expensive equipment and complicated calibrations that are not practical for at-home settings. It is evident from the current state-of-the-art technology that there remains a need for improvements in 6-DoF biomechanically accurate avatars, that can be coupled with whole-body VR yoga therapy to strategically address key components of chronic pain management, including relaxation, pain distraction and holistic mind-body benefits through engaging yoga movements. The solution can be developed using a multi-modal system architecture to enable seamless integration of current and future body movement tracking technologies.

In this work, we designed a system architecture for VR platform with a 6-DoF biomechanically accurate whole-body avatar capable of recreating complex poses in VR using a single depth camera and a VR headset. The avatar architecture is driven by multiple machine learning based sensing modalities working in tandem, comprised of a body pose estimation algorithm, an on-board VR hand tracking algorithm, and custom fusion models developed for real-time movement integration and noise removal. We examined the accuracy of our system in a cohort of children and young adults by comparing it with a commercial state-of-the-art body tracking method. We hypothesized that our proposed multi-modal architecture would improve the movement tracking accuracy and degree of smoothness thereby delivering an immersive and fully embodied VR experience. Our results demonstrate the feasibility of developing a VR yoga therapy platform targeted towards children and adolescents with chronic pain conditions, that can track whole-body movements with high fidelity and can provide an immersive and engaging VR environment for individual, clinical, and social engagement.

II. Related work

The popularization of meta-verse technologies has led to advances in the immersion and embodiment of VR avatars. Winkler et al. designed a model to derive animated avatar joint transform positions based on controller and headset articulated movements [30]. The study utilized the tracking data obtained from a standalone Head Mounted Display (HMD) and hand controller devices in a reinforcement learning framework to estimate full body movement in VR. However, the model was assessed on simple activities, such as balance, and walk/jog in place. ControllerPose [31] is another study that investigates full body reconstruction from an inside-out (internal sensors within VR hardware) approach by retrofitting VR hand controllers with webcams that are facing towards the user to derive image based human pose estimation and VR avatar movement. Although the system demonstrated initial feasibility, it was also limited to a small range of human body movements with minimal effort towards improved embodiment.

While these approaches mimic poses within their controlled task specific environments using on-board sensing modalities, a generic method able to capture a larger range of real-world poses is necessary for clinical applications. One such body tracking study that attempts to reproduce complex body movement poses, MixSTE conducted by Zhang et al., uses an offline batch processed spatio-temporal encoding for 3D human pose estimation from single outside-in camera input [32]. MixSTE model was extensively tested against the current state of the art body tracking approaches leading to an improvement of 7.6% in mean per-joint position error (MPJPE) against other comparable methods. Large complex models such as MixSTE produce relatively accurate estimates, however, they are limited by their non-real-time performance which is critical for VR as delayed perception of self-body movements can cause distortion in VR immersion and embodiment [33].

Despite the advances in various VR technologies aimed at treating chronic pain, existing pose estimation algorithms have yet to deliver an immersive and accurate solution to a whole-body VR avatar rendering needed to accommodate the wide variety of yoga poses used in practice. Key design elements to such an approach should include modular integration of various sensing modalities to capture redundancy of joint positions, fusing both on-board and inside-out sensing modalities needed to estimate complex full body joint transforms in real time that enable the development of a VR platform with real-time tracking of 6-DoF biomechanically driven whole-body avatar movements.

III. Design And Implementation

We designed a whole-body immersive multi-modal algorithmic architecture for fusing real-time 3D body joint estimation derived from an RGB-D camera with inverse kinematics models of physical movement. Kalman filtering was implemented to reduce jitter in the predicted joint positions. This resulted in a 6-DoF whole-body avatar capable of embodying individual’s complex real-time yoga poses within a VR environment. The avatar architecture design encompasses four main components: 1) Avatar Rigging – structural makeup of virtual skeleton; 2) Tracking modalities – integration and mapping of a human pose tracking algorithm and hand tracking algorithm; 3) Informed Inverse Kinematics – custom algorithms to fuse joint tracking from different tracking modalities based on biomechanical constraints; and 4) Noise Filtration – a signal processing algorithm to remove measurement noise in movement tracking modalities. A detailed flow-chart of this architecture is presented in Figure 1. All development stages were implemented within Unity 3D Game Engine.

Figure 1.

Avatar Architecture Pipeline: After obtaining measure from a body tracking method, we validate our input using biomechanical constraints and refine the measure with a Kalman filter fusing the input with a generic hand tracking method to obtain a full body representation of an Avatar. Gray blocks indicate avatar rigging, amber indicates tracking modalities, blue indicates inverse kinematics and green indicates noise filtration algorithms.

A. Avatar Rigging

Avatars used in this work were provided via the open-source Mixamo Avatar library from Adobe. These avatars follow the conventional animation guidelines used for non-VR platform video games. The challenge of implementing humanoid skeletons into VR is the need to reconcile the inherent hierarchy of skeleton biomechanics with the head-driven hierarchy inherent to VR headsets, where the virtual body is driven by the user’s perspective via the head-mounted VR device. We mitigated this challenge by adapting the hierarchy of the humanoid skeleton using the head as the primary positional driver. To make the head the main positional driver, the head joint was set to be the root, at the top of the skeleton hierarchy using the HMD tracking shown in Figure 2. Any objects underneath the parent would be affected by positional and rotational changes made to the root. As the root, the head was also affected by the rotation of all objects below it in the hierarchy, introducing the need to constrain the neck’s rotation. To solve this problem, we rotated the avatar’s neck joint based on its local rotation inversely to the head, such that the head node could rotate independently of the rest of the body nodes.

Figure 2.

Illustration of joints and their associated tracking methods with two-bone IK chain. The end joint follows the same position as the target. The target being the input from the sensing modality.

B. Tracking Modalities

Our architecture was based on the fusion of inside-out and outside-in movement tracking modalities irrespective of the specific hardware (camera or HMD) or software applications utilized. For the scope of this study, we developed and evaluated this architecture using an existing inside-out tracking modality – the Meta Quest 2 VR headset and integrated onboard hand tracking solution – and an outside in tracking modality – the Azure Kinect the integrated body tracking algorithm. These modalities were used to determine joint positions concurrently that were then fused into a single overall biomechanical estimate of avatar movement using A custom IIK algorithm and the measurement confidence values determined from the individual modalities.

1). Inside-out Tracking

For the selected hand tracking algorithm, we integrated MEgATrack, Meta Quest 2’s full articulated hand tracking, which leverages the use of four monochrome cameras on board the HMD to estimate key-points on a hand [34]. Based on their study, MEgATrack tracks 21 key-points on the hand with an average fingertip positional error of 1.1cm, an average finger joint angle error of 9.6°, and an average temporal delay of 45 ms. We integrated distal, proximal, and metacarpophalangeal joints from this algorithm to their respective avatar joint position to allow full articulated hand control.

2). Outside-In Tracking

For the selected 3D human body pose estimation we integrated the Azure Kinect body tracking algorithm which uses the RGB-D data frames derived from a single Kinect camera as input to a neural network to derive body key-points. The Azure Kinect full body tracking algorithm captures data with an average delay of 2.6 frames for each 33 ms frame. We integrated 14 keypoints from Azure Kinect’s body tracking solution including forearm (x2) and upper arm (x2), hip (x2), pelvis, spine (x2 : thoracic and lumbar), neck, knee (x2), and ankle (x2) and mapped them to their respective joint rotation transforms along the avatar body. We selected these keypoints as clinically relevant joints based on stakeholder feedback and discarded unnecessary keypoints (ears, nose, toes etc.).

C. Informed Inverse Kinematics Algorithm

High fidelity hand and arm tracking is a key factor in the overall sense of embodiment due to the presence of the hands commonly in the user’s field of view. We used inverse kinematics (IK) to estimate arm pose with respect to three joints shown in Fig. 2 (End, Lower, and Upper). IK provides an advantageous framework to estimate a chain of kinematic joint positions and rotations given its ability to operate from a relatively small set of constraints and inferential information. In our application, we modified the standard two bone IK into a new informed inverse kinematics (IIK) model to account for the non-deterministic nature of the analytical model for the lower joint, which introduced the need for additional inferential information.

We implemented the IIK algorithm by assigning an empirically computed elbow position and orientation to the pole component shown behind the elbow in Figure 2. Since the pole is a component of a standard IK chain that enables the control and direction of bends in the chain, input was applied from the body tracking algorithm to inform the pole’s positional height based on the initial position and rotation of the shoulder and elbow (obtained from Kinect body tracking), and wrist joints (obtained from Meta Quest hand tracking) already integrated within the global frame of the Unity environment. We then derived an equation to estimate the height of the elbow and set the position of the pole as

$$y_p=b_p-\left(\lambda \left(y_e-y_s\right)\right)$$

where y_p is the height of the pole, b_p is the base height of the pole at start, y_e is the height of the Kinect estimate of the elbow, y_s is the height of the Kinect Estimate of the shoulder, and λ is a user defined sensitivity parameter for tuning.

D. Noise Filtration

To mitigate sources of high-frequency noise within our body-tracking algorithm, we implemented an Unscented Kalman filter based on its wide application in other fields to remove noise from continuous time series data, and its capability for predicting future positional states in real-time. For a given frame, the joint tracking raw rotational measure from that respective joint was replaced by a set of projected 3D joint movements from the Unscented Kalman filter leading to smoother rotational movement.

Additional sources of noise caused by sporadic glitches in the body joint tracking outcomes were mitigated by implementing constraints based on the biomechanical joint hierarchy, movement speed, and range of motion in each data frame. One example of such constraints is when the Kinect body tracking model makes an error in identifying the global position across a group of joints and the whole avatar is intermittently snapped into a random incorrect body position. This snapping between positions causes abrupt VR avatar body movements at a rate much higher than volitionally possible. We implemented an empirically driven velocity constraint on biomechanical joint movements that was used to identify and reject such unrealistic body movement glitches.

IV. FEASIBILITY EVALUATION

A. Participant Demographics

Ten children (n=5; 10–15 years) and adolescents (n=5; 21–25 years) participated in the experiment, males and females were equally represented across the cohort. A pre-experiment informational briefing was conducted with each participant explaining the procedures, risks, benefits, and rights according to Institutional Review Board (IRB) regulations. Prior virtual reality familiarity was surveyed by asking participants to score their use of VR based on a 4-point scale: 1-never, 2-once, or twice, 3-multiple times, 4-device owner (refer to Table 1 for additional subject characteristics).

Table 1.

Participant Demographics

ID	Age	Height	Sex	VR Familiarity
1	10	111.8	F	2
2	11	157.5	M	4
3	11	144.7	M	3
4	13	144.3	F	2
5	15	167.6	F	2
6	21	152.3	F	3
7	24	167.6	M	3
8	25	180.3	F	3
9	25	187.9	M	2
10	25	180.3	M	2

B. Experimental Protocol

Two sets of experiments were conducted to evaluate whole-body movement tracking accuracy and VR embodiment of the system platform designed. For the first set of experiments, the participants carried out a series of yoga poses wearing a VR headset to validate the accuracy of the whole-body movement tracking in VR with respect to a reference standard motion capture system (MTX 160, Vicon Systems). A second set of experiments were conducted following the accuracy assessment to evaluate VR immersion and embodiment of our system among the cohort of participants using Likert scale-based surveys.

1). Whole-body Movement Tracking Validation

The accuracy of the whole-body movement tracking in VR was validated with respect to a reference standard multi-camera motion capture system (MTX 160, Vicon Systems) set up with 8 distributed cameras to provide 360° movement coverage. Motion tracking markers were instrumented on the participant’s neck, shoulder (L/R), elbow (L/R), wrist (L/R), spine, knee (L/R), and ankle (L/R) along with additional marker clusters along each peripheral limb for motion tracking robustness as per the guidelines from International Society of Biomechanics (ISB) [35, 36]

A brief calibration routine was performed for each participant prior to movement data collection to customize the avatar height to the participant’s height. All other parameters of the avatar such as its visual appearance and limb-segment ratios were maintained constant across participants. Our custom IIK and signal filtering algorithms were used to translate the body segment positions to recreate a simplified VR avatar in a virtual environment that faithfully captures a participant’s real-world positions.

Ten yoga poses were selected in consultation with the yoga therapy experts at Boston Children’s Hospital (BCH), Boston, MA, USA, and Mass General Brigham (MGB), Boston, MA, USA to identify those used in practice for pediatric rehabilitation. These included Cobra, Triangle, Sandwich, Plank, Downward Dog, Warrior 1, Warrior 2, Supine, Tree, and Lotus. The selected poses were in alignment with evidence-based yoga therapy protocols for pain, recently reviewed by the National Center for Complementary and Integrative Health (NCCIH) [37]. During the poses, concurrent data were collected from the Unity engine rendering the VR avatars movements based on the IIK and signal filtering algorithms designed; the commercial off the shelf Kinect body pose tracking models; and the reference motion capture system for accuracy comparisons.

2). Embodiment Assessment

To assess user embodiment sentiment, a second VR session referred to as a period of Free Movement was conducted along with evaluations of their experience with the avatar during the session. In the free movement period, each participant was given a few minutes to be in VR to experiment with their avatar and move freely in the space as they preferred. The instructions were to take your time to observe the body movements and experiment in any way you feel with your avatar. The participants practiced with the avatar for at least 3 minutes and could leave at any point thereafter. During this time participants tried several self-selected activities including fine hand movements, head movement and jumping among others. Once the free movement period concluded, a VR Sickness Questionnaire (VRSQ) was administered to test for cybersickness using a binary question (Presence/Absence of cybersickness associated with nausea, headaches, or blurry vision [38]. A validated virtual embodiment questionnaire (VEQ) was also administered to investigate ownership and agency [39] using a 7-point Likert scale (1-Strongly Disagree: 7-Strongly Agree). All questionnaires were administered outside the VR platform.

C. Data Processing

Three modalities of movement tracking data were recorded for each trial – 1) reference marker data recoded from the motion capture system; 2) multi-modal avatar movement tracking data obtained from the platform designed in this study; and 3) commercial off the shelf avatar movement tracking using Microsoft Kinect Azure body tracking models. Data recorded for evaluating the movement accuracy comprised of 3D positions across selected whole-body joints (XYZ), and time stamps of each frame from 10 participants who performed 10 poses each with 3 repetitions per pose leading to an overall dataset of 300 recordings of tracked yoga movements for each individual tracking modality. Movement accuracy was compared between our multi-modal avatar and the Kinect avatar with respect to the reference motion capture system based upon the absolute Euclidean distance was used to calculate 3D positional error between each joint in respective data frames.

V. Results

A. Quantitative Comparison of Joint and Pose Tracking

The weighted mean 3D Euclidean joint error comparing the multi-modal avatar and Kinect avatar is demonstrated in Figure 3. We used weighted mean per joint positional error (WMPJPE) to determine average distance outside of the radial joint, weighted to radial size of the avatar joint [32]. Overall, we improved the joint WMPJPE across all joints by 67 ± 4% with respect to the existing Kinect avatar. The WMPJPE from the ground truth of each pose position comparing the multi-modal avatar and the Kinect avatar in Figure 3 show a decrease in average WMPJPE from Kinect by 66 ± 5% consistently across all poses tested. The multi-modal avatar’s average WMPJPE (7.7 ± 6.9 cm) decreased 16cm compared to the Kinect avatar (23cm ± 18.4). A paired t-test (α = 0.05) indicated significantly higher positional fidelity performance (p < 2.2e-16) was achieved by our multi modal avatar compared with the existing Kinect solution.

Figure 3.

(top) WMPJPE over all joints comparing our multi-modal avatar to the raw Kinect avatar, (bottom) WMPJPE averaged across poses comparing our multi-modal avatar to the raw Kinect avatar.

A representative side-by-side comparison of our multi-modal avatar and real-life movements is depicted in Figure 4 showing a participant engaged in each of two poses, a floor (Cobra) and standing pose (Tree). The corresponding multi-modal avatar can be seen in real-time measured for all participant trials an average 79.5 ± 5 FPS (Frames Per Second) across 236,600 frames, matching the user’s movements.

Figure 4.

Avatar Unity renderings of floor (1.A) and standing (2.A) pose and their respective ground truth image for cobra floor pose (1.B) and tree standing pose (2.B).

B. VR Sickness & Immersion Survey

VRSQ responses were analyzed to assess perceived player experiences involving motion sickness from the VR-based intervention. Binomial analysis was performed based on the binary response answers in the VSRQ for headache, blurry vision, and nausea. With the hypothesized value of 0 symptomatic vs non-symptomatic (α = 0.05), no participants experienced a significant degree of blurry vision (p = 0.08), nausea (p = 1), or dizziness (p = 1) following gameplay.

VEQ responses (Figure 5) were analyzed using a Wilcoxon signed rank test (α = 0.05) to assess a participant’s ownership (Q1-Q4) and agency (Q5-Q8) over their VR avatar. Our findings indicate significant improvement in both the ownership and agency categories over a median (neutral) response on the Likert scale (h = 4).

Figure 5.

Virtual Embodiment Questionnaire (VEQ): questions (Left) and responses (Right) categorized as negative, positive, or neutral. Questions 1–4 (Ownership). Questions 4–8 (Agency)

VI. Discussion

Feasibility of a full body avatar for VR yoga therapy was evaluated in this study and found to provide significant improvements in whole body tracking when compared to conventional methods. The beneficial effect provided a realistic and acceptable feeling of embodiment without VR sickness among the sample participants for all ten yoga poses. The findings from this work provide a solid baseline for further development of a VR yoga app for children and young adults with chronic pain to help them adhere to self-directed mind-body treatment regimens when out of a clinical setting.

Experiments carried out in control populations of children and young adults showed our multi-modal avatar architecture significantly improved over an existing off-the-shelf Kinect avatar. From the WMPJPE results, the average improvement across all joints achieved a 67 ± 4%, and across all poses, 66 ± 5%. The closest study on full body VR avatar design for virtual reality using webcam HybridTrak, achieved a mean per joint positional error of 13.8cm when compared to a less accurate markerless motion capture database during relatively constrained everyday tasks [40]. Our framework performs slightly better with a 12.8cm mean per joint positional error with respect to reference standard marker-based motion capture system [41, 42], while maintaining this accuracy improvement across more complex poses than those tested by HybridTrak. Another study, LoBSTr [43] used sparse upper body tracking signals and a wearable laser tracked HTC Vive tracker but obtained a higher average toe-base combined error of 17.9 ± 2.3cm than this study and was limited to 4 participants and only a small subset of activities excluding any complex pose transitions and dynamic movements encountered in yoga therapy. Yet another study, HOOV [44] predicted occluded hand movement when the controller was out of the line of sight to track shoulder, elbow, and wrist movements; achieving an online evaluation error of 16.97 cm across 12 participants which is higher than the error achieved in this study and limited to the use of body-worn sensors. In contrast, Ego-poser [45] achieved average MPJPE of 9.2 ± 1.4 cm over a pre-collected offline (outside VR) dataset, improving avatar tracking beyond the results demonstrated in this study, however, their findings were limited to simple activities of daily living and their evaluation was based only on an offline system outside the VR environment. Our multi-modal VR avatar achieved a consistent 79.5 FPS, well-beyond the 72 FPS needed for interactive VR standards [46].

Positive user sentiment on system feedback and responsiveness from the control group was observed for all users in the study. Results from the VEQ additionally support significant levels of ownership and agency (perceived movement fidelity) when using the avatar. Few users reported slightly negative sentiment when asked about the ownership of the avatar when asked if the body felt like it was their body, this can be attributed to sex of the avatar used during testing or the scale of the limbs. High levels of agency in the virtual avatar were recorded with 100% positive responses for 3 questions indicating they felt like the body was responsive and was being controlled in sync with them. Additionally, we provide results to confirm our system does not cause any symptoms related to cybersickness through our analysis. A single user reported a slight blurry vision after the session related to VR use and no reports of nausea or dizziness following the experiment.

Specific areas of improvement and future work were also uncovered in the areas of user avatar ownership. From the user sentiment questions in the VEQ, Users reported that the avatar was not their body (sex, appearance, limb proportions). Since the primary focus of the work was to test movement and pose fidelity, a single avatar was used for all participants. The importance of user customizability not only sheds light on embodiment not only in appearance, but the proportions of the limbs also can affect how the user interacts with the world. Improvements in avatar customization can improve outcomes both with immersiveness and avatar fidelity.

The contribution of our work creates a foundation on which a complete VR yoga therapy application can be built with features in pose recognition, multi-person interaction, and clinical outcomes. The ability of the avatar to extract rotational and positional data can be used to implement pose recognition of full yoga poses, which can help validate user adherence to exercises as well as teach proper form. In conjunction with higher data fidelity in future iterations, data extraction can provide users and clinicians interested in a patient’s data to observe clinical outcomes such as range of motion and postural balance. Lastly, data extraction from our avatar allows us to network objects with that positional data to be relayed to other users for multi-person sessions or classes.

VII. Conclusion

These findings demonstrate the feasibility of rendering whole-body avatar movements that embody complex physical poses such as those encountered in yoga therapy to facilitate at-home rehabilitation for children or adults with chronic pain conditions that can be enhanced by real-time monitoring of clinical outcomes. The ability for the avatar to mimic complex yoga poses allows a system such as this to implement pose detection and other features with meaningful outcomes which could be extended to other applications reliant on a full body virtual avatar. The significance of this work could aid children with chronic pain conditions providing a technology for pain management and rehabilitation. Future work entails the efficacy evaluation of symptomatic users in an at home setting. Additionally, we plan to expand our sample size evaluate limb segments, and other pose estimation algorithms to further evaluate our model. The impact of this work helps facilitate the creation of a tool to motivate, rehabilitate, and improve the quality of life with children with chronic pain conditions.

Table 2.

Wilcoxon signed rank test on VEQ responses

Metric	Mean	Median	Wilcoxon (V)	Wilcoxon (p)
Ownership	5.281	5.500	42	<0.03
Agency	5.0	5.87	55	<0.01