On Recognizing Abnormal Human Behaviours by Data Stream Mining with Misclassified Recalls

Abstract

Human activity recognition (HAR) has been a popular research topic, because of its importance in security and healthcare contributing to aging societies. One of the emerging applications of HAR is to monitor needy people such as elders, patients of disabled, or undergoing physical rehabilitation, using sensing technology. In this paper, an improved version of Very Fast Decision Tree (VFDT) is proposed which makes use of misclassified results for post-learning. Specifically, a new technique namely Misclassified Recall (MR) which is a post-processing step for relearning a new concept, is formulated. In HAR, most misclassified instances are those belonging to ambiguous movements. For examples, squatting involves actions in between standing and sitting, falling straight down is a sequence of standing, possibly body tiling or curling, bending legs, squatting and crashing down on the floor; and there may be totally new (unseen) actions beyond the training instances when it comes to classifying "abnormal" human behaviours. Think about the extreme postures of how a person collapses and free falling from height. Experiments using wearable sensing data for multiclass HAR is used, to test the efficacy of the new methodology VFDT+MR, in comparison to a classical data stream mining algorithm VFDT alone.