Detection of stable contacts for human motion analysis

Human motion analysis is one of the important topics in visual surveillance applications,the ultimate goal of which is to achieve automated scene understanding. This paper proposes a novel "stable contact "concept for temporal abstraction of image sequences, and presents a Hidden Markov Model (HMM)based framework to recognize continuous human activities. With the extended star-skeleton representation, stable contacts are formed by stationary extreme points, and image sequences are segmented temporally into adjacent but disjoint primitive intervals. We define a set of primitive motion units (PMU 's)over primitive intervals based on stable contacts and trajectories. Thus frame sequences are abstracted as PMU sequences. Discrete HMM 's are trained on manually segmented sequences to classify segmented testing PMU sequences into predefined activities. The continuous recognition on non-segmented PMU sequences is achieved by searching over the time axis, for the best fit between durations of PMU sequences and types of activities. The experiments on various sequences of (mixed) human activities, including walking, running and climbing (fences or rocks), are presented to show the effectiveness of the proposed concept and framework.