What Can Machine Learning Infer about the Way You Walk?

This poster at the EuroScience Open Forum 2016 in Manchester applied machine learning to walking activities recorded with a smart-carpet floor sensor. The activities included cognitively demanding tasks.

Dr. Omar Costilla Reyes presented the poster “What can machine learning and a smart carpet infer about the way you walk?” in the Marie Curie Alumni Association (MCAA) poster session at the EuroScience Open Forum (ESOF) 2016, a European interdisciplinary science meeting held in Manchester on July 24–27, 2016. He took part as a Marie Curie young researcher from the University of Manchester. The poster, by Omar Costilla-Reyes, Patricia Scully, and Krikor Ozanyan, was published in the MCAA’s ESOF 2016 poster proceedings under the title “What Can Machine Learning Infer about the Way You Walk?” During the forum he presented at the poster stand and ran a public Open Labs session on machine learning and gait at the university.

The University of Manchester’s smart carpet is a floor-sensor prototype that records how people’s footsteps press on it. It uses plastic optical fibers: footstep pressure slightly deforms the fibers, which changes the amount of light they transmit, and those changes are turned into a record of each footstep in space and time. The poster calls this system iMagiMat. Because the carpet produces a large volume of raw spatio-temporal data, the poster proposes deep neural networks, which learn their own features from raw signals instead of relying on hand-designed ones, for gait characterization. The applications it points to range from security verification systems to identifying markers of neurodegenerative diseases such as Alzheimer’s disease.

The poster reports a spatio-temporal analysis of 13 gait activities performed on the carpet, including cognitively demanding tasks. It concludes that the activities have distinct patterns of frequency, amplitude, and signal type that deep neural network models can identify. This shows that a floor sensor captures enough of the structure of walking, including how it changes when a person is also thinking, for deep learning to tell the activities apart, which supports the carpet as a tool for gait characterization. The temporal analysis behind the poster is in the authors’ 2016 IEEE Sensors Journal paper, “Temporal pattern recognition in gait activities recorded with a footprint imaging sensor system,” and the work developed into Dr. Costilla Reyes’s PhD thesis on gait analysis for healthcare and security. Both are described elsewhere on this site.