Few-shot Fall Detection using Shallow Siamese Network

The threat of falling down is significantly higher for the geriatric population and can lead to serious injuries including death. In the past, classical Machine Learning/Deep Learning-based methods have been successfully investigated for the detection of falls. However, most of these methods require a lot of data in order to be successfully trained for accurate detection. In this work, we propose a shallow architecture using 1 × 1 filters for use in a few-shot Siamese network. The proposed architecture was used in a Siamese network-based fall detection system. The proposed detection system is shown to effectively learn feature representations for the detection of falls when trained with few signals acquired from wearables containing inertial motion unit (SisFall dataset). The proposed system achieved a performance of 93% ± 7% and 72.5% ± 10% in 15 and 1-shot scenarios respectively. Performance comparisons with Siamese convolutional autoencoders and transfer learning¬based approaches demonstrated the superiority of the proposed few shot fall detection system.