Classificação de manobras de skate através de acelerometria e redes neurais artificiais

Abstract

Skateboarding is one of the most popular cultures in Brazil, with more than 8.5 million skateboarders. Today Skateboarding is more recognize as a true sports discipline, also given by its debut on the Olympic Games in Tokyo 2020. The present study aims to develop methods detection and classification of Skateboarding flat ground tricks, using a Inertial Module Unit (IMU) allied whit Artificial Neural Networks (ANN). Using state of the art knowledge on movement detection in Skateboarding, 181 artificial flat ground tricks divided between 5 different classes were generated, equivalent to 543 signals (X, Y and Z) of acceleration. A classification heuristic based on cross-correlation coefficients was applied to the discriminate between classes, and using the Neural Network Toolbox a Multilayer Feed Forward Network was architect with three layers and a scaled conjugate gradient back propagation algorithm. The results showed that with the use of ANNs trained specifically for each axis measured by the IMU, being the Z-axis the best differentiator, we can reach error percentages inferior to 0.05 %, with a computational efficiency that makes real time applications possible. Machine learning is a useful tool in pattern recognition of complex movements, as long that the classifiers are properly architected and the task is expressed with clarity this technique is a promising application in Exergames and motion detection.