Gait recognition via deep learning of the center-of-pressure trajectory: A proof-of-concept study for biometric applications
Download (626.46 kB) This item is shared privately
journal contribution
modified on 2019-01-16, 21:20 Each
human being has a distinctive style of walking. Gait recognition has therefore
been proposed for identification purpose. Using end-to-end learning, I
investigated whether the center-of-pressure trajectory on the ground was
sufficiently unique to identify a person with a high certainty. 36 adults
walked for 30min on a treadmill instrumented with a force platform that
recorded the position of the center of pressure at 50Hz. The raw 2D signals
were sliced into segments of two gait cycles. A balanced set of 20,250 segments
from 30 subjects was used to configure and train convolutional neural networks
(CNN). The best CNN classified a separated set containing 2,250 segments with
99.9% overall accuracy. A second set of 4,500 segments from the 6 remaining
subjects was then used for transfer learning. Several small subsamples of this
second set were randomly selected and used to fine-tune the top layers of the CNN.
Training with 12 segments (2 per subject) was enough to reach 100% accuracy. The
results suggest that everyone produces a unique trajectory of underfoot
pressures. CNNs can learn the distinctive features of these trajectories. Using
transfer learning, few strides could be enough to learn and identify new gaits.