Comparison Between Features Extracted in the Time and Frequency Domain with the Triangulation Method in the Recognition of Activities of Human Movements

In this research, classical features of the literature in the time and frequency domains were compared with the features based on the triangulation technique, which was applied to the input signals of human movements divided into categories (walking on a treadmill, running on a treadmill, walking in a circle, erasing a blackboard, going up and down stairs and hand tremors). Signals are captured with the internal accelerometer of a smartphone, using embedded software, which reads 3 orthogonal axes. The comparison was performed only at the highest percentage value in the rate, and the analysis of statistical significance was not performed. The triangulation method uses Euclidean concepts and basic statistics to extract features based on straight lines, angles, areas, perimeter, derivative and triangle counter in the composition of patterns. The results are promising, in view of the high accuracy rates in the final classification of the categories of movements object of study, with values similar to those obtained in the classification of patterns composed by the classic features used in the literature, with the MLP (Multilayer Perceptron) classifiers and KNN (K-nearest neighbors) for k = 5. The highest average hit rates obtained for the experiment with 500 points were 99% with the literature methods, 97.9% for the triangulation method and in the experiment with 1000 points windows, the highest average hit rates were achieved with the triangulation method 99.3% and 98.7% for the methods found in the literature, both with the KNN classifier for k = 5.