A correspondence based method for activity recognition in human skeleton motion sequences

In this paper we present an algorithm for efficient activity recognition operating upon human skeleton motion sequences, derived through motion capture systems or by analyzing the output of RGB-D sensors. Our approach is driven from the assumption that, if two such sequences describe similar activities, then, consecutive frames (poses) of one sequence are expected to be similar to consecutive frames of the other. The proposed method adopts a quaternion based distance metric to calculate the similarity between poses and an intuitive method for estimating a similarity score between two skeleton motion sequences, based on the structure of a pose correspondence matrix. Our method achieved 99.5% correct activity recognition, when applied on motion capture data, in a classification task consisting of 18 classes of activities.