Método de triangulação aplicado sobre dados capturados com acelerômetro de movimentos humanos, visando a extração de novas características e o reconhecimento de padrões

Abstract

This research presents a triangulation method, which was applied to the 3D electrical signals discretized from normal everyday human movements and abnormal ones such as tremors, captured from the internal accelerometer of a smartphone, aiming at pattern recognition. The main contribution of the proposed method is in the elaboration of an algorithm that uses Euclidean concepts and basic statistics in the composition of new fetures extracted from sequences of triangles, which are based on straight lines (ascent, descent and base), perimeter, angles , angular coefficients of the straights (derived) and triangle count. The new concepts can be applied in the area of biomedical engineering, in view of the amount of studies that use human movements and the accelerometer sensor, as well as the excellent results presented in this study. Comparisons were made between the proposed method and the well-known concepts of FFT (Fast Fourier Transform), which converts a signal in the time domain to the frequency domain, as well as comparisons with statistical features in the time domain. Four sets of patterns that apply the triangulation technique and 3 sets of patterns that apply the classic features of the literature were elaborated, in 4 experiments varying the amount of movements (5 and 9), the amount of features (9, 15 and 36) and the size of the input data windows (100, 200, 500 and 1000 points) in the pattern composition. The classification indices for each supervised class (1 to 9), using two classifiers, Multilayer Perceptron (MLP) and K-Nearest Neighbor (KNN) for k=5, were used in the evaluations and the values of individual and average hits were recorded by movement, among the 9 categories studied. As for the average hit rates, in the third experiment (5 movements and 36 features), the values 99.6% and 99.6% were results achieved with the sets of features applying triangulation (and average) and the sets of classic features from the literature, for windows of 200 and 500 points, both with the MLP classifier, respectively. Favorable result for the triangulation technique due to the smaller input window size (200 points). Using the same 5 movements, in the third experiment and windows of 1000 points, the highest average rate of 100% hit was obtained, with the feature sets that apply triangulation (and sum) and 99.9% using the feature sets from the literature, both with the MLP classifier. In the fourth experiment, using all 9 movements, the average indices of 99.5% were obtained with the features of the literature and 99.5% with the features of triangulation (and average), in this case, for windows of 1000 points and classifier KNN. The results obtained in this research are promising, in view of the high hit rates obtained with the triangulation method, which were equated to the indices obtained with the sets of features in the literature, as well as, when performing the association of all the features (triangulation and literature) in a new processing, the hit rates in the final classification (categories of movements) improved, proving the efficiency of the proposed method.