The bounded asymmetric Gaussian mixture model (BAGMM) has proved that it generally performs better than the classical Gaussian mixture model. In this thesis, we investigate the learning of the BAGMM. Indeed, we propose an Expectation-Maximization (EM) algorithm to estimate the model parameters. A model selection criterion for BAGMM using minimum message length (MML) is proposed to determine the optimal number of clusters. The MML is shown to perform better than other model selection criteria.
In this thesis, we additionally propose an unsupervised feature selection framework using BAGMM to determine the structure of high dimensional data without knowing in advance the number of clusters nor the importance of the involved features. The validation for this framework involves several human-related recognition challenges, such as human activity categorization and human gender recognition.
Finally, we integrate the BAGMM into a hidden Markov model (HMM) framework, which uses BAGMM to model the emission probability distribution. The BAGMM-based HMM is evaluated with several real-world applications and compared with other Gaussian mixture-based HMMs.