Action recognition to estimate Activities of Daily Living (ADL) of elderly people

This work proposes a method, and preliminary experimental results to detect and recognize a set of Activities of Daily Living, carried out by elderly people in a residential context, by analyzing video of actions, recorded using an RGB Camera. The proposed solution is based on the creation of neural network models, in particular Convolutional Neural Networks (CNN), which are trained on data extracted and preprocessed from the 'Moments in Time' dataset, a resource released by MIT-IBM Watson AI Lab that includes a collection of a million labeled 3-second videos from hundreds of categories. The performances of the models obtained following two different approaches are also described, the one using Auto Machine Learning, which was also necessary in order to have an idea of the achievable performances and the transfer learning approach. One of the main drivers of our research activity was also to explore and challenge the Auto Machine Learning approach in the context of ADL and evaluate its initial accuracy baseline concerning Transfer Learning approaches.