Abstract:
Accidental and sudden fall is a significant cause of injury and even death in most
cases, among the old age people. As it has a remarkable impact on the national
health system, so extensive development and proper research for fall detection
systems are very necessary. This thesis provides a comprehensive review of state of-the-art technologies and finds deep learning as the most powerful methodology.
In this study, we proposed an Inception Resnet inspired cnn plus bidirectional
LSTM neural network. We have trained and fine tune this model on state of
the art Korean dataset which was recently published. We have calculated the
magnitude of all the 3 channels of acceleration, gyroscope, and Euler data in
the preprocessing step before feeding it to the network. In this study, we have
considered 1D, 3D, and 6D data and trained the same model accordingly, and
achieved remarkable results. It has been observed that our model performed
better in the case of 8 overlapping step sizes.
