Evaluation of Sensing and Processing Parameters for Human Action Recognition

Abstract

Accurate recognition of human actions is essential to many health-care, entertainment, and human-computer interface applications. However, the achievable accuracy depends on a variety of parameters for the various stages of recognition, including sensing, feature extraction, and classification. In this paper, we quantitatively evaluate the classification accuracy for varying sensing rates, sensor data filtering techniques, segmentation approaches, and classification algorithms for several different types of action. The results of these evaluations provide guidelines and strategies for the design of future action recognition systems.