A study on feature extraction from human body motion by using cepstrum analysis

In this paper, we propose a method that classifies human body motions into four states, running and walking on flat ground, going up and going down the stairs. In our approach, we first use the 3-axis accelerometer of a mobile device to collect timeseries data. Then, we focus on the periodicity of body motion and apply cepstrum analysis for classification. In this research, cepstrum analysis is a typical analysis method used in audio signal processing, but also used for feature extraction from human body motion. And, we apply continuous Hidden Markov Model (HMM) to automatically clustering the four types of data from the mobile device. As a result, just only 45% data were classified into the four classes correctly by using the raw data, and 60% by using its power spectrum. However, in the case of using cepstrum analysis, the accuracy reached 73%. This indicates that our approach is effective in feature extraction from human body motion. © 2013 ICIC International.