Human Action Recognition Using Acceleration Information Based On Hidden Markov Model
Shinichi Takeuchi, Satoshi Tamura, Satoru Hayamizu · Institutional Repositories DataBase (IRDB) · 2009
This paper investigates the best feature parameter for human action recognition by using Hidden Markov Model (HMM) with triaxial acceleration sensor information. Our target is to recognize six types of basic actions (walk, stay, sit down, stand up, lie down, get up) in the room of daily life. First, as parameters in time domain, acceleration information in three axes and their derivatives are used as the baseline method. Secondly, Mel-Frequency Cepstral Coefficients are used as feature parameters in frequency domain. As the recognition result, the best recognition rate was obtained when MFCC and its Δ elements of the axis that contained most of gravitational acceleration information.