Abstract
At current stage, the majority of the human activity recognition (HAR) technologies are based on supervised learning, where there are labeled data to train an expert system. In this paper, we proposed a framework based on the unsupervised learning to autonomously discover, learn and recognize atomic activities, i.e., the actions. The input to the HAR framework is a sample pool of unlabeled observations of an unknown number of actions. An incremental action discovery algorithm based on K-means is used to discover new actions. For each new action discovered, a learning algorithm is used to model it through an automated training and cross-validation cycle. The algorithm uses Mixture of Gaussians Hidden Markov Model (HMM) to model the actions, and the algorithm autonomously determines the appropriate number of Gaussian components and states. The framework deals with the dynamic and noisy nature of the data. We evaluated the proposed framework on a third party dataset of daily activities and the results show its performance is in-par with that achieved using a supervised learning algorithm to recognize the activities from the same dataset.