Adaptive gait recognition model and automatic velocity-robust feature selection based on a new Constrained Expectation Conditional-Maximization learning algorithm

A robust and compact human motion model is desirable in many security applications from public facilities to personal devices. Shape features are extracted from the perspective of computer vision in most researches. However, most of them are application-dependent. In order to explore more dynamical features of human motion and to make the human model adaptable to the varying environments, a new Stochastic Switched Auto-Regressive Model together with an innovative Constrained Expectation Conditional-Maximization algorithm which utilizes pre-knowledge from feature space analysis is proposed. The proposed model has a circular topology consisted of 2 pairs of correlated states and the constrained ECM algorithm is proposed under the model´s unique structure. The problem is complicated by the fact that, even though the dominant features are dynamic, there are significant static features. Modeling the underlying behavior is challenging especially when the parameter estimation algorithm does not guarantee that the updated model converges to a maximum likelihood estimator. The proposed method can produce a probability distribution over the latent variables with point estimates. The modeling method can be reviewed as approximating maximum likelihood in a non-Bayesian way with adaptability to changing walking velocity.