Abstract:
Vision based Human Activity Recognition or simply HAR is a widely researched area that
is helpful in understanding human behaviour in images and videos. HAR is an important part of
various research problems such as detecting and preventing crimes with the help of automated
video surveillance, robot movement without human intervention and to provide telecare for
elderly. In this research, an algorithm for activity recognition using 2D pose information extracted
from human skeleton is implemented. The approach is based on angles between the joints and
displacement of joints between frames. Two publically available datasets are used for training and
testing purpose. For activity recognition, five well known techniques of supervised machine
learning are implemented separately including K nearest neighbours, SVM, Linear Discriminant,
Naïve Bayes and Back propagation neural network. Using these techniques, four action classes
Sit, Stand, Fall and Walk, are recognized in videos. Results for all the classifiers are compared to
find the best performing technique for the proposed methodology. All classifiers performed well
with the best performing classifier achieving an overall accuracy of 98%. The results show that
proposed methodology gives compatible accuracy with the state of the art in this field.
