Free Weight Exercises Recognition Based on Dynamic Time Warping of Acceleration Data - Intelligent Computing for Sustainable Energy and Environment - page 182

a high recognition rate. The method of multi-template average is adopted in this
paper. The process of template building is listed below[15]:

(1)Calculate average length of each type and choose the series nearest to the
average length as the initial reference template.
(2) The other training series are time aligned by the DTW process such that their
lengths will be equal to the chosen initial template.
(3) The final reference template will be created by averaging the time-aligned
series.

4 Experimental Results

4.1 Experiment Setup

The experiments are conducted by three male and a female subjects, whose heights
are from 160 to 185cm. Every subject did each exercise in three sets using dumbbell
of different weight. They were told to try their best to 15 repetitions in each set.
Finally the acceleration data of 1610 repetitions were collected with a sampling rate
of 100Hz. The data were filtered by the 5th order Butterworth low-pass filter with a
cut-off frequency of 5 Hz.

4.2 User-Dependent Results

The user-dependent protocol checks the algorithmic robustness for individual users
with different weights. The reference templates were built using one of three sets for
each subject and the other two sets were used to test. The confusion matrices are
shown as Table 1. The algorithm proposed in this paper is named as the improved
DTW (IDTW) in order to distinguish it from the standard DTW (STDW). The same
dataset is tested by Artificial Neural Network (ANN), Support Vector Machine
(SVM) and SDTW in order to compare their performance.

Table 1. Confusion matrices of IDTW by user-dependent protocol

BC TC BP FL BR LR OP DL SCR
BC 120 0 0 0 0 0 0 0 0
TC 0 115 0 0 0 0 0 0 0
BP 0 0 117 0 0 0 1 0 0
FL 0 0 0 120 0 0 0 0 0
BR 0 0 0 0 108 0 0 9 0
LR 0 0 0 0 0 120 0 0 0
OP 0 0 6 0 0 0 114 0 0
DL 0 0 0 0 0 0 0 120 0
SCR 0 0 0 0 0 0 0 0 120