A self-growing Bayesian network classifier for online learning of human motion patterns

This paper proposes a new self-growing Bayesian network classifier for online learning of human motion patterns (HMPs) in dynamically changing environments. The proposed classifier is designed to represent HMP classes based on a set of historical trajectories labeled by unsupervised clustering. It then assigns HMP class labels to current trajectories. Parameters of the proposed classifier are recalculated based on the augmented dataset of labeled trajectories and all HMP classes are accordingly updated. As such, the proposed classifier allows current trajectories to form new HMP classes when they are sufficiently different from existing HMP classes. The performance of the proposed classifier is evaluated by a set of real-world data. The results show that the proposed classifier effectively learns new HMP classes from current trajectories in an online manner.