Measuring Sway with Markerless Depth Camera

MSR-TR-2019-22 |

Published by Microsoft

The goal of this study is to present methods to measure properties of posture, specifically sway, using ordinary depth cameras. Current methods either use markers which require accurate placement of reflectors or use virtual skeletons which may fail for people with atypical body structure. We introduce algorithms to extract sway metrics from the sensor which do not require markers or virtual skeleton extraction. Three experiments were conducted to measure the proposed algorithm: (1) Measurements on a mannequin connected to a robotic arm were used to measure sensitivity and repeatability (2) Measurements using 20 healthy subjects were used to compare against measurements by a pressure sensor (3) Measurements using 13 healthy subjects were used to verify that subjects pose is within the sensitivity of the sensor. Results show that frequency based metrics and axes invariant metrics were repeatable and showed little sensitivity to the positioning of the sensor. Axes based metrics (lateral-ventral) showed sensitivity to the viewing angle of the sensor when the angle was > 5deg. When testing on human subjects results indicate that many of the measurements from the depth camera are correlated well with measurements of a pressure sensor (Pearson correlation >= 0.65) and are more accurate that results obtained by using the virtual skeleton. Moreover, it is sufficiently accurate to distinguish between the sway imposed by different postures, such as standing with feet in tandem vs feet parallel (p-value ~ 0.05 – 0.005). We use inter-class correlation coefficient to measure the reliability and found it to be ~ 0.5 for most metrics. Last, we show that although the sensor is sensitive to the viewing angle, an acceptable viewing angle (0.55deg +- 1.89) is achieved when subjects are asked to stand behind a line marked on the floor. These experiments show that methods to measure sway with markerless depth cameras without using virtual skeleton extraction are reliable and outperform methods that use the virtual skeleton.

An implementation of the methods is available at