Fall Detection Using Self-Supervised Pre-Training Model

Abstract

Several researchers have developed fall detection using wearable sensors due to their flexibility and nature of privacy. Most of those developed methods are supervised deep learning methods. However, data annotation is expensive because we use camera video recording and playback of each participant's recorded video to label the data. This paper presents how to use unlabeled data to pre-train FCN and ResNet models, and use labeled data to fine-tune those pre-trained weights. We collected unlabeled and labeled data and applied self-supervised learning to detect falls. The experiment in this study suggested that the best performance can be achieved by using pre-trained weights of unlabeled data from the accelerometer and gyroscope sensors. Furthermore, oversampling and modified loss functions are used to handle the dataset's imbalance classes. With the ResNet pre-trained weights and training using the labeled data, the experiments achieved an F1Score of 0.98.