Selective Factorized Coupled Hidden Markov Model - Exploiting Collective Sensor Information for Human Activity Recognition

The availability of wearable and ambient sensors allows more information to be captured for human activity recognition. However, noises and signal disconnections are common in complicated environment. Resolving noise and variability from different inputs, as well as accounting for the human-object interactions is challenging under complex settings. To address the challenge, I present a novel Hidden Markov Model variant that includes both coupled and factorized states for estimation and learning problems.
In this thesis, I provide the detailed formulation of selective factorized coupled hidden markov model (SFCHMM), including its model definition, forward-backward procedure for conditional observation probabilities, optimal state path decoding and parameter estimation. In addition to the algorithmic discussion, I also test the model by simulating on synthetic CHMM processes and applying to a real world sensor-rich benchmark dataset that recorded human daily activities. The performance analysis based on the experiments demonstrates that this model is capable of consolidating the fuzzy information from a collective pool of sensors and improving human activity recognition in interactive context, which is highly applicable to real world settings such as surveillance, smart home and multimedia games.