VQ-HMM classifier for human activity recognition based on R-GBD sensor

Abstract

This study presents a framework for understanding the human activities in home by using 3-D skeleton joints captured by a Kinect sensor. The system is developed for the visual system of home robot to enhance the humane as well as the abundant for robot application. The proposed system treats the human activities as a time series of representative 3D poses data. Since the skeleton joints are encoded into pose vocabularies by Vector Quantization, an activity can be described as a series of poses. Discrete HMMs are trained to classify sequential poses into activity type. Experiments are performed on online test with the average accuracy 95.64% obtained. The experimental results have demonstrated the effectiveness and efficiency of the proposed system in real time application.