Modelling Duration in Activity Recognition using Generative and Discriminative Models

Human activity recognition allows many applications in areas such as intelligent environments and health monitoring. Typically probabilistic models such as the hidden Markov model or conditional random fields are used to map the observed sensor data onto the hidden activity states. A weakness of these models, however, is their inaccurate modelling of state durations. Hidden semi-Markov models and semi-Markov conditional random fields model duration explicitly. Our experiments show that modelling duration accurately can lead to a significant increase in recognition performance, in real world activity recognition.