WheelPoser: Sparse-IMU Based Body Pose Estimation for Wheelchair Users
Yunzhi Li, Vimal Mollyn, Kuang Yuan, Patrick Carrington · 2024 · ASSETS '24: Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663548.3675638
Summary
This paper presents WheelPoser, a real-time body pose estimation system designed specifically for wheelchair users using only four inertial measurement units (IMUs). Existing pose estimation methods are largely developed for standing and ambulatory users, relying on assumptions about lower-body movement and ground contact that do not hold for seated wheelchair users. The authors address this gap by creating a purpose-built dataset and model architecture. The WheelPoser-IMU dataset comprises 167 minutes of paired IMU and optical motion capture data collected from 14 wheelchair users performing a range of activities including desk work, eating, exercise, household tasks, sports, and self-care. Participants included both full-time wheelchair users and individuals simulating wheelchair use to expand the dataset. The system places sensors on the wrists, lower legs, and the back of the wheelchair, which serves as a consistent reference point unique to this population. The model architecture uses a three-stage approach: first, a bidirectional recurrent neural network predicts joint angles from IMU data; second, forward kinematics converts these angles to 3D joint positions; and third, a physics-based optimization step corrects for biomechanical constraints like joint limits. The authors evaluated several model variants and found that the biRNN approach with kinematics integration and physics optimization produced the best results.
Key findings
WheelPoser achieved a mean joint angle error of 14.30 degrees and a mean joint position error of 6.74 cm, representing a threefold improvement over prior sparse-IMU pose estimation methods when applied to wheelchair users. The physics-based optimization stage reduced position error by approximately 1 cm and significantly improved physical plausibility of the predicted poses. The system generalizes across different wheelchair types and user populations, though performance varied across activity categories — structured activities like desk work yielded lower errors than dynamic movements like sports. A key methodological finding was that mounting one IMU on the wheelchair itself, rather than on the body, provided a stable reference frame that substantially improved accuracy. The authors also demonstrated that their purpose-built seated dataset was essential, as models trained on standing-user data performed poorly for wheelchair users. Real-time performance was validated with the system running inference at interactive rates suitable for practical applications.
Relevance
WheelPoser addresses a significant gap in assistive technology research by making body tracking accessible to wheelchair users, a population largely excluded from existing pose estimation systems. The practical applications are substantial: clinicians could monitor patients for pressure injury risk by tracking weight shifts and repositioning frequency, rehabilitation professionals could assess wheelchair skills remotely, and wheelchair users could participate more fully in VR experiences and motion-controlled gaming. The work also highlights a broader accessibility principle — that systems designed around ambulatory assumptions create barriers for wheelchair users that require deliberate, purpose-built solutions rather than adaptations. The open-source release of both the dataset and model supports future research and development in this space. For accessibility practitioners, this research underscores the importance of including wheelchair users in motion-tracking and embodied interaction design from the outset.
Tags: wheelchair users · pose estimation · inertial measurement units · motion capture · body tracking · health monitoring · assistive technology · machine learning