ABSTRACT

In this work, methods for the identification of walkers based on
measurements by a pressure-sensitive floor were studied and
developed. A 100 square meter pressure-sensitive floor was covered
with ElectroMechanical (EMFi) material and used as part of a smart
room in a research laboratory. Single footsteps were segmented from
the raw data and featurized to train the classification models. Two
different classification methods were studied: Learning Vector
Quantization (LVQ) and discrete Hidden Markov Models (HMM). The
features used in identification were mainly based on the extreme
points of the signal, such as heel strike and toe-off peak, and their
interrelations. The amplitude spectrum values from the frequency level
presentation were also used as features.

In LVQ, the classification is based on a codebook consisting of a
finite set of codebook vectors assigned to every class. The
initialized codebook is trained with featurized training footsteps to
describe the class boundaries in the feature space. Finally, an
unknown sample is classified to the closest codebook vector. In HMM
classification, a single footstep is presented as a sequence of
temporally variable feature vectors, called observations. These
sequences are encoded as discrete symbols using a trained LVQ
codebook. HMMs are initialized for every class and trained with
sequences of observed symbols. An unknown observation sequence is
classified to the most probabilistic model.

In addition, extended versions of LVQ learning algorithms were studied
to increase the reliability and adaptiveness of the identification
systems. A distinction-sensitive LVQ (DSLVQ), which is able to
automatically detect the most informative features during the training
of the codebook, was applied to footstep identification. Furthermore,
a two-level identification method developed in this project was
proposed. It uses three consecutive footsteps and a reject option to
make a final decision on classification.

LVQ showed a 67% overall success rate in classifying single footsteps
of 11 different walkers, while HMM was able to identify 52%
correctly. When the number of occupants was decreased to five, the
total recognition rates were much more reliable, 91% and 84% for LVQ
and HMM, respectively. DSLVQ was able to select the most relevant
features, which increased the recognition rate up to 70% in
identification tests of 11 persons. The reject-optional two-level
classifier gave very promising results with a recognition rate of 90%
and a rejection rate of 20% in identifying eleven walkers. The results
show that it is possible to recognize a small number of occupants from
the pressure signals of walking steps as part of an intelligent
environment.

Keywords: intelligent environment, gait-based identification,
pressure-sensitive floor, LVQ, HMM, pattern classification