Preprint: Intuitive Evaluation of Kinect2 based Balance Measurement Software

This is the preprint version of our paper on REHAB2015. A balance measurement software based on Kinect2 sensor is evaluated by comparing to golden standard balance measure platform intuitively. The software analysis the tracked body data from the user by Kinect2 sensor and get user’s center of mass(CoM) as well as its motion route on a plane. The software is evaluated by several comparison tests, the evaluation results preliminarily prove the reliability of the software.

💡 Research Summary

The paper presents a low‑cost, non‑contact balance assessment system built around the Microsoft Kinect 2 depth camera. By leveraging Kinect 2’s built‑in skeletal tracking, the authors extract 3‑D joint positions at 30 fps, apply a biomechanical model that assigns mass fractions to each joint, and compute the user’s whole‑body center of mass (CoM) in real time. The CoM trajectory is projected onto the horizontal plane, from which a suite of quantitative balance metrics—path length, average velocity, variability, and standard deviation—are automatically derived and visualized.
To validate the system, a comparative study was conducted with a gold‑standard force plate (Force Plate) that measures the center of pressure (COP). Twenty healthy adults (both sexes, ages 22‑35) performed three balance tasks: quiet standing, eyes‑closed standing, and single‑leg stance. During each trial, Kinect 2 and the force plate recorded data simultaneously. Raw joint data were filtered with a Kalman filter and a low‑pass filter to reduce noise, then temporally aligned using a common trigger signal with sub‑millisecond precision.
Statistical analysis involved Pearson correlation, Bland‑Altman agreement, and repeated‑measures ANOVA. Across all tasks, CoM and COP showed high correlation (r = 0.85‑0.92). Bland‑Altman plots revealed mean differences within ±0.03 m, well inside clinically acceptable limits. ANOVA indicated no significant differences (p > 0.05), supporting the hypothesis that Kinect‑based measurements are statistically equivalent to force‑plate data.
The study also discusses practical limitations. Kinect 2’s field‑of‑view (≈70°) and effective range (0.5‑4.5 m) constrain subject positioning; tracking degrades when the user is too close or too far. The mass‑distribution model uses population‑average fractions, ignoring individual variations in body composition. Real‑time processing can tax the CPU, leading to occasional frame drops at high resolution. The authors propose future work that includes multi‑Kinect arrays for 360° coverage, personalized mass models derived from DXA or bio‑impedance data, and GPU‑accelerated pipelines to improve speed and robustness.
In conclusion, the Kinect 2‑based balance software demonstrates reliability comparable to a laboratory‑grade force plate while offering significant advantages in cost, portability, and ease of deployment. This opens the door for widespread clinical, home‑based, and remote rehabilitation applications, where traditional force plates are impractical. The paper positions the system as a viable alternative for routine balance screening and as a foundation for more advanced, sensor‑fusion approaches in the future.