Factored State-Abstract Hidden Markov Models for Activity Recognition Using Pervasive Multi-modal Sensors

Current probabilistic models for activity recognition do not incorporate much sensory input data due to the problem of state space explosion. In this paper, we propose a model for activity recognition, called the Factored State-Abtract Hidden Markov Model (FS-AHMM) to allow us to integrate many sensors for improving recognition performance. The proposed FS-AHMM is an extension of the Abstract Hidden Markov Model which applies the concept of factored state representations to compactly represent the state transitions. The parameters of the FS-AHMM are estimated using the EM algorithm from the data acquired through multiple multi-modal sensors and cameras. The model is evaluated and compared with other exisiting models on real-world data. The results show that the proposed model outperforms other models and that the integrated sensor information helps in recognizing activity more accurately.