A technical summit in Cambridge
The 2017 Arm Research Summit, titled “Computing for the Connected World,” was held September 11–13 at Robinson College in Cambridge, United Kingdom. Arm described it as a multi-track technical conference with talks, demonstrations, workshops, and posters on topics including computer architecture, security, and machine learning, following an inaugural summit in 2016. Dr. Omar Costilla Reyes announced in July that he would take part, and he presented “Analysis of Spatio-Temporal Representations for Robust Footstep Recognition with Deep Residual Neural Networks” at the summit in September 2017.
Recognizing people by their footsteps
Footstep recognition is a biometric method: it asks whether the pressure pattern of a person’s footstep, recorded by a sensor in the floor, is distinctive enough to tell one person from another. The work treats each footstep as a spatio-temporal signal, meaning a pattern of pressure across the floor surface that also changes over successive moments. Deep residual networks are neural networks whose blocks add a shortcut from input to output, which makes very deep models easier to train. The architecture figure on this page shows two such networks working in parallel, one reading a single image of a footstep and the other reading a short sequence of frames, with their class scores fused at the end.
Results and related work
The paper that followed, published in IEEE Transactions on Pattern Analysis and Machine Intelligence in 2018 with his co-authors, evaluated a large database of footstep signals from many people and reported a low equal error rate for one tested configuration, meaning that wrongly accepted impostors and wrongly rejected genuine users were both rare. Gait is attractive as a biometric because a footstep is a natural, contact-based signal that a person produces simply by walking, and the paper shows that deep residual networks can turn such signals into reliable identification. The presentation sat within his doctoral research on floor-sensor gait analysis for security and healthcare. Related results appear in “Spatial footstep recognition by convolutional neural networks for biometric applications” (2016) and in later seminars at the University of Cambridge and the Max Planck Institute for Intelligent Systems.