Accelerometry shows promise in providing an inexpensive but effective means of understanding the human gait. Accurate classification of everyday gait patterns could allow a monitoring system to exhibit greater ‘intelligence’, like, improving the ability to detect and monitor the activities of a person, in a more accurate tracking of the person health parameters in an unsupervised environment. This thesis develops a gait classification algorithm using features derived from novel gait models. The Linear Predictive (LP) model is proposed to model the gait with the basic assumption that the gait is a system that comprises impulse trains input, which correspond to the footstrike. From the Linear Predictive model, the Linear Predictive Cepstral Coefficients (LPCC) were proposed as features, as they have better class separation compared to using LP coefficients directly. The filterbank energies were proposed to approximate the LPCC with a smaller number of parameters. Spectral Centroid Amplitude and Spectral Centroid Frequency features were also developed to integrate frequency information with the proposed filterbank features. The second gait model that is proposed in this thesis is a harmonic model. The premise behind this model is that the accelerometer signal’s spectral contains a fundamental frequency found to be the walking stride rate with multiple harmonics that fit a harmonic model. The first four harmonics were found to be a good approximation of the acceleration movement of the hip. The magnitudes of the harmonics were used for classification features and proved to be better features than the linear predictive filterbank features. Another advantage of the harmonic features is that they are independent of walking speed, which therefore enables the extraction of features that are robust to the variations of speed. Novel linear delta zero crossing counts regression features were used as complimentary features to the proposed static features from the model parameters. The assumption that the zero crossing counts provide a good characterisation of the amount of artefacts caused by the muscle movements was the motivation for using these features. Finally, the non-linear Empirical Mode Decomposition (EMD) method was investigated to extract nonlinear features for gait pattern classification. The Bayesian adapted GMM classification system was used as a back-end system to further improve the classification accuracy of the system by removing the subject variation of the system. In addition, score level fusion was proposed to integrate the linear regression delta zero crossing features into the static features.