Abstract:
Falling is one of the major health concerns faced by elderly people globally. Accurate and
timely prediction or detection of falls could result in a significant reduction in injuries
and associated costs. In this study, we have presented a deep learning model inspired by
InceptionResNet which can detect a fall with an inference time of 52 milliseconds in the
pre-fall phase. We conducted 3-fold testing by extracting temporal features of a falling
signal from three different phases: pre-impact, post-impact, and the fall-cycle phase.
The system is capable of predicting and detecting 15 types of fall activities alongside
21 different activities of daily living (ADL). The classifier has achieved outstanding
results in detecting falls in all phases, with an average sensitivity and F1 score of 98%.
Achieving a sensitivity of 98% implies that the model has effectively minimized the
occurrence of false negatives, which is vital for an accurate and reliable fall detection
system. Furthermore, the proposed model has been evaluated using three different
types of input features with varying dimensions, ranging from 1D to 6D. These features
include the "1D magnitude of Accelerations", "3D Accelerations", and "6D Accelerations
& Angular Velocities", which were used to examine the impact of input size on the
model’s complexity.
