Temporal pattern recognition for gait analysis applications using an "intelligent carpet" system

This 2015 IEEE SENSORS paper reports an intelligent carpet that classifies different walking manners from temporal floor-sensor signals. About a dozen machine-learning models were compared on short sequences of footsteps.

Presented at the IEEE SENSORS 2015 conference in Busan, South Korea, by Omar Costilla-Reyes, Patricia Scully, and Krikor B. Ozanyan, the paper reports the demonstration of a new floor sensor system for gait analysis in the time domain. The “intelligent carpet” uses plastic optical fibers woven under an ordinary carpet surface; when a footstep deforms a fiber, its light transmission changes, and the signals from all fibers over time form the temporal record of a walk. The paper was part of Dr. Costilla Reyes’s doctoral research at the University of Manchester.

The system’s ability to detect changes in gait was evaluated with pattern-recognition techniques on walking manners that differed in the amplitude, frequency, and type of the temporal signal. Each trial used a few successive footsteps, from which time-series features were computed and passed to about a dozen machine-learning models. A random-forest model, which combines the votes of many decision trees, achieved the strongest reported F-score; the F-score balances how many of the identified cases are correct against how many true cases are found. The authors conclude that pattern recognition on gait activities recorded by the smart carpet is suitable for gait-analysis applications ranging from biometrics to healthcare.

The paper received the Best Student Paper Award in optical sensing applications at the conference, and an expanded version was published in the IEEE Sensors Journal in 2016 as “Temporal pattern recognition in gait activities recorded with a footprint imaging sensor system.” The plastic-optical-fibre sensing layer of the University of Manchester’s iMagiMat smart carpet is described in a separate 2015 paper, and later work analyzed such raw floor-sensor signals directly, as in the 2017 tomography-sensor study.