Visual Fall Detection From Activities of Daily Living for Assistive Living

Health facilities have increased life expectancy, a key factor for the growth of the elderly population. Elderly people are at increased risk of falls, causing physical and psychological damage. Falls occur rarely compared to other activities of daily living. Due to such a class imbalance, supervised techniques are not the solution for fall detection systems. In addition, domain-level features for the fall activity are hard to generalize due to their diversity. In this work, the fall detection problem is formulated as anomaly detection in the time series where deviation from the activities of daily living is computed. On the basis of the deviation score, a fall is detected. We propose TCHA, Temporal Convolutional Hourglass Autoencoder, to learn spatial and temporal features from the videos. Hourglass units in the Temporal Convolutional Encoder help us extract multiscale features by expanding the receptive fields of neurons, reducing the information loss in deep learning methods. The proposed methodology is evaluated on the five data sets, including a compiled data set from publicly available Toyota Smarthome data set and four benchmarked datasets that include the UR-Fall dataset, IMVIA dataset, SDU dataset, and Thermal Fall dataset. Our methodology shows 4.1% superior results to existing state-of-the-art methods for unseen falls.