Falls are unexpected occurrences during daily activities that lead to significant difficulties in life. Not only do falls frequently happen among the elderly, but they also account for a high percentage of accidents in industries and hazardous occupations. For people who working in high-risk jobs such as firefighters or even working in industrial site, the rate of accident realated with falls are more than 20%.
To address these incidents caused by falls, research related to fall detection is actively being conducted. This study aimed to develop and evaluate various models that classify whether a fall has occurred when given measured data using a device attached to body which is equipped with an accelerometer and a gyroscope to measure people's movements. The goal was to ascertain which model performed best through performance evaluations.
To gather data, an experiment was conducted where devices were attached to the waists of participants who repeatedly performed actions defined as falls and non-falls.
Collated data tested by 5 models which were Support Vector Machine (SVM), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), and CNN-LSTM model. Performance evaluations were carried out while maintaining some of the hyperparameters the same for the deep learning models, to make the appropriate model selection. Among the implemented models, the LSTM model showed the most superior performance on the configured dataset.