Deep Learning Based Fall Detection System
Talha Aslam, Faezah Harun, Aizat Faiz Ramli, Kushsairy Kadir, Muhammad Noor Nordin · 2023
A number of organizations, both medical and non-medical, have adopted assistive technology, such as medical wireless wristbands and fall detection devices. Most research on fall detection using deep learning techniques are based on image data sets, which are not readily available. So, researchers have focused on making a deep learning-based classifier that takes numerical input of sensors data and perform effectively, so that medical costs can be reduced. Deep learning-based fall detection is an assistive device to detect fall particularly for elderly people which can significantly reduce the mortality rate from falls by providing a response as part of a medical system. This paper developed a deep learning methodology to increase classifier accuracy using numerical data for fall detection. With the help of Deep Learning-Artificial Neural Network (DL-ANN), accelerometer, gyroscope, and infrared dataset are feed into DL-ANN consisting of convolution layer, pooling layer, and fully connected layer containing hidden layers to improve the classification accuracy. The DL-ANN was implemented on Jupyter Notebook with the help of a Python library for this Artificial Neural Network approach. This research also employs framework such as NumPy, Pandas, Scikit-Learn, Matplotlib, and Tensor Flow libraries. Moreover, the accuracy of the DL-ANN classifier was tested on eight optimizers with the help of different Keras layers in sequential model. The adaptive moment estimation (ADAM) algorithm performed the best in terms of loss while training the model. Deep Learning-Artificial Neural Network method helped to improve the classification accuracy of the wearable fall detection devices by up to 97%. This was done by training the data with a lot of computing power. Also, if falls could be reliably detected using a deep learning-based classifier like ANN, which got a 97% accuracy, 96% precision, 97% recall, and 96% F1-score rate in this paper, it would pave the way for the development of a fall detection solution that would work for lifelong.