Gupta and Dallas have proposed feature selection and activity recognition system using a single tri-axial accelerometer in 2014 [1]. The physical activities recognized in this study are walking, running, jumping, sit-to-stand/stand-to-sit, stand-to-kneel-tostand and being stationary (sitting and standing at one place). The data are sampled at 126Hz during the experiments. The acceleration data are segmented into windows of 6s each with a 50% overlap between two consecutive windows. All features were evaluated for the longitudinal axis (Y-vertical), resultant acceleration(X-Y-Z), and acceleration in horizontal plane (X-Z). Window of 6s sample of transitional event (i.e. sit-to-stand) might contain some data corresponding to walking of being stationary that followed or preceded the transitional event. Thus, features such as energy, entropy, mean and variance of transitional event samples were observed to closer to those of walking or being stationary. In order to capture the transition, new features were developed which would break the 6s window of samples into smaller windows and analyzed it. The features are mean trend and windowed mean difference, variance trend and windowed variance difference, DFA coefficient, X-Z energy uncorrelated and maximum difference acceleration.
The set of features are Mean [2], Variance [2], Detrended Fluctuation Analysis coeff., X-Z Energy Uncorrelated and Maximum Difference Acceleration [3].
[1] Gupta, P., and T. Dallas. ‘Feature Selection and Activity Recognition System Using a Single Tri-Axial Accelerometer’. IEEE Transactions on Biomedical Engineering Early Access Online (2014). doi:10.1109/TBME.2014.2307069.
[2] Bao, Ling, and Stephen S. Intille. ‘Activity Recognition from User-Annotated Acceleration Data’. In Pervasive Computing, edited by Alois Ferscha and Friedemann Mattern, 1–17. Lecture Notes in Computer Science 3001. Springer Berlin Heidelberg, 2004. http://link.springer.com/chapter/10.1007/978-3-540-24646-6_1.
[3] Gupta, Piyush, Gabriel Ramirez, Donald Y. C. Lie, Tim Dallas, Ron E. Banister, and Andrew Dentino. ‘MEMS-Based Sensing and Algorithm Development for Fall Detection and Gait Analysis’. 7593:75930U–75930U–8, 2010. doi:10.1117/12.841963.