A dynamic segmentation based activity discovery through topic modelling

Recent developments in ubiquitous and pervasive technologies have made it easier to capture activities through sensors. The “bag-of-word” topic models have been applied to discover latent topics in corpus of words. In this paper, we propose the Probabilistic Latent Semantic Analysis to discover activity routines. The framework we propose set latent topics as corresponding class labels and use the Expectation Maximization (EM) algorithm for posterior inference. The experimental results we present are based on the Kasteren dataset which validates our framework and shows that it is comparable to existing activity discovery approaches.