Activity recognition using unsupervised learning
Anca Alexan, Alexandru Iulian Alexan, Stefan Oniga · 2022 IEEE International Conference on Automation, Quality and Testing, Robotics (AQTR) · 2022
In developing a high-performance generic algorithm capable of managing multiple environments, the main challenges are the large and very different data sets and the problem of different types of residential spaces. The human activity recognition domain in residential environments has registered an exponential growth lately. In developing a high-performance generic algorithm capable of managing multiple environments, the main challenges are the large and very different data sets and the problem of different types of residential spaces. The training process of supervised algorithms is also limited since most of the available datasets do not have the activities data labeled. This article describes a classification method for dataset data. The used dataset is Kyoto generated by CASAS. The raw data is cleaned in the pre-processing phase, and the new features are extracted. These extracted features form a new dataset of input data fed to a cluster classification neural network. The processed dataset is labeled in order to simplify the validation of the obtained results. The dataset will contain 120 activities comprised of five main actions: Make a phone call in the dining room, Wash hands, Eat, and Clean. Since we are using an unsupervised classification algorithm, we can also use unlabeled data-sets as well.