Real time gait pattern classification from chest worn accelerometry during a loaded road march.

Accelerometers, whether in smart phones or wearable physiological monitoring systems are becoming widely used to identify movement and activities of free living individuals. Although there has been much work in applying computationally intensive methods to this problem, this paper focuses on developing a real-time gait analysis approach that is intuitive, requires no individual calibration, can be extended to complex gait analysis, and can readily be adopted by ambulatory physiological monitors for use in real time. Chest-mounted tri-axial accelerometry data were collected from sixty-one male U.S. Army Ranger candidates engaged in an 8 or 12 mile loaded (35 Kg packs) timed road march. The pace of the road march was such that volunteers needed to both walk and run. To provide intuitive features we examined the periodic patterns generated from 4s periods of movement from the vertical and longitudinal accelerometer axes. Applying the "eigenfaces" face recognition approach we used Principal Components Analysis to find a single basis vector from 10% of the data (n=6) that could distinguish patterns of walk and run with a classification rate of 95% and 90% (n=55) respectively. Because these movement features are based on a gridded frequency count, the method is applicable for use by body-worn microprocessors.