Invited seminar at Nottingham Trent University
In May 2017, Dr. Omar Costilla Reyes gave a seminar at Nottingham Trent University (NTU) as part of the School of Science and Technology’s joint seminar program, at the invitation of Dr. Eiman Kanjo, who hosted the visit. At the time he was a doctoral researcher at the University of Manchester, and he said afterward that he looked forward to working together on intersecting research projects.
Learning from space and time
The seminar title names two ideas. Spatio-temporal means combining where something happens, here across the surface of a floor sensor, with when it happens, over successive moments of a walk. A convolutional neural network is a deep-learning model that slides small filters across a grid of numbers and learns for itself which local patterns matter, instead of relying on hand-designed gait measures such as step length. In the related paper, “Deep neural networks for learning spatio-temporal features from tomography sensors” (2017), the University of Manchester’s iMagiMat floor sensor, which records how plastic optical fibers deform under footsteps, supplied the raw readings. The authors arranged those readings into spatio-temporal matrices and trained a convolutional network to recognize a range of walking activities, with very high accuracy on their dataset. Learning directly from the sensor’s own spatial and temporal structure, rather than from a few summary numbers, is what allowed the network to separate walking activities so reliably.
Why the approach matters
Treating a walking sequence as a structured spatio-temporal signal lets a model compare representations of movement and classify walking patterns, an approach relevant to both healthcare monitoring and biometric recognition. The seminar also drew on the floor-sensor studies collected in his University of Manchester thesis, and the same framing underpinned his footstep-recognition work with deep residual networks, presented later in 2017 at the Arm Research Summit in Cambridge.