Abstract:
Common fall occurrences among the growing elderly population pose significant
issues in the public healthcare realm. Falls frequently result in severe or even
dangerous injuries that are the greatest cause of long hospitalizations and even death
among elderly people. To overcome this issue, it is important to build robust fall
detection systems. The main aim of this study is to develop thigh mounted
smartphone-based solution which can handle tricky falls-like Activities of Daily life
(ADLs) efficiently along with actual falls. The main dataset used in this research is the
FallDroid dataset which consists of tricky ADLs along with real falls. In this research
work, multiple machine learning and deep learning models were also compared on
several publicly available datasets that are recorded by simulating adult falls. The
Random Forest, a machine learning model, and the bi-directional long short-term
memory (BiLSTM), a deep learning model along with two feature selection
techniques (Mutual Information and Eigenvector Centrality) are evaluated for both
the reference dataset and the publicly accessible datasets to provide a robust
solution. After the robust testing, the BiLSTM model tends to perform better overall,
with the highest average accuracy of 95.81% on the reference (FallDroid) dataset and
98.44% on other widely used public datasets.
