Files
Download Full Text (686 KB)
Document Type
Poster
Publication Date
Fall 10-1-2026
Abstract
Gait analysis is a useful tool for early detection, diagnosis of neurological and orthopedic conditions, and rehabilitation support. The gold standard for assessment of gait requires laboratory motion capture systems. The system requires calibrated cameras, reflective markers, specialist software, and professionals to operate. For these reasons, gait analysis is expensive and inaccessible in some clinics, schools, and home settings where continuous or early monitoring would have a significant impact. KinetiScan addresses this gap with just a single ordinary camera. This browser-based system uses transformer architecture to perform 2D-to-3D mapping. When provided with distance and height entries, KinetiScan outputs the 3D positions of 20 main body joints and clinical gait metrics. The model was fed deliberately noisy distance measurements while still being supervised against the true values to prevent it from echoing the input. The system’s components were validated separately. We used ground-truth-derived rays for the transformer-based model and laboratory motion capture data (30 healthy adult participants, 474 walking trials) for the kinematics engine. The pose model achieved 46.8 mm MPJPE and 29.0 mm PA-MPJPE. Root-depth RMSE was 249.1 mm with per-frame ground-truth distance and 262.7 mm with four ground-truth-derived camera-group means. The kinematics engine achieved 0.89-1.00 correlation and 1.2–6.6° range-of-motion difference. End-to-end validation remains untested, and the system is not yet ready for clinical use.
Department
Biomechanics Lab, LC RM 1053, Department of Kinesiology, DePauw University, Greencastle, IN
Project Mentor
Reuben Addison
Recommended Citation
Tran, Thu '29; Agyemang, Stephen '28; and Addison, Rueben, "KinetiScan: Evaluating a Browser-Based Monocular System for 3D Gait Analysis" (2026). Annual Student Research Poster Session. 255.
https://scholarship.depauw.edu/srfposters/255
Funding and Acknowledgements
Supported by Faculty Development funds, DePauw University. Work conducted in the Biomechanics Lab (LC 1053), Department of Kinesiology. We thank the authors of the public H3WB, BioCV, Schreiber, overground (Riglet et al.), and multi-sensor (Santos et al.) datasets for making this validation possible.