FiPR: A Fine-grained Human Posture Recognition

Abstract

Various pioneering human posture recognition techniques based on Channel State Information (CSI) of WiFi devices have been proposed. The main issue of existing techniques, however, lies in such recognition methods are extremely sensitive to the impacts of random noise derived from indoor environments. In this paper, we present a fine-grained human posture recognition (FiPR) scheme to overcome this issue by extracting two unique statistics features in CSI profile, including mutual information (MI) and cross correlation (CC). In order to eliminate the influences of noise components on the recognition accuracy, a corresponding Discrete Wavelet Transform (DWT) strategy is introduced to denoise by using signal decomposition. Furthermore, FiPR can recognize four basic human postures by measuring the correlation between a given unknown posture and pre-constructed postures profiles. Compared with existing Doppler-based recognition methods, the recognition accuracy of the proposed FiPR scheme can be improved effectively. We implement FiPR scheme on the commercial WiFi devices and evaluate its overall performance in a typical indoor environment. Experiment results demonstrate that our prototype can estimate human posture recognition with average accuracy of 95%.