Privacy-Aware Human Activity Classification using a Transformer-based Model

Fall detection in a bathroom requires privacy as an important issue. Ultra-wideband sensors have the ability to protect human privacy because their output has only limited information. As a result, interpreting the output is a challenging task. This research implemented a transformer model that learned time-series signals from an ultra-wideband sensor in a bathroom. First, the signals were preprocessed into the two-dimensional range-time format. Second, the range-time data were passed into a convolutional neural network encoder before going into a transformer. Basic movements of humans were used for training. The encoder and the transformer were trained separately. The model achieved a good accuracy on static postures but not good on transitions due to their overlapped similarity with the static postures.