Diseño y aplicación de un modelo de reconocimiento de 11 gestos de la mano usando señales EMG, IMU, DTW Y KNN.

Abstract

In the present work, a real-time hand gesture recognition model was developed based on signals taken from static and dynamic gestures. The inputs for this model are hand movement signals measured by surface electromyography (EMG) and by an inertial measurement unit (IMU) using sensors embedded in commercial armbands. Both types of signals are obtained from commercial Myo Armband and G-Force Pro sensors. The dataset used for this work contains EMG and IMU signals from 85 users. Those signals were collected during the execution of 12 gestures: 11 gestures: wave in, wave out, fist, open, pinch, up, down, left, right, forward and backward plus the relax gesture. For each gesture, 15 repetitions of 5 seconds were executed, composing a total of 180 samples for each user. The output of the model is the classified label of the gesture that the user executes at each moment and the instant of time where the gesture was performed, i.e., the model provides results in terms of classification and recognition of the gesture performed. This model can recognize 11 gestures. The recognition phase is based on the KNN and DTW algorithms. In this model, the KNN algorithm takes the signal classification function using the DTW algorithm as the similarity distance function between the input signal and the trained signals. The evaluation of the model was carried out in terms of response time, classification accuracy and recognition. The results of this work were a classification accuracy of 80% and a recognition accuracy of 66% of recognition below 0.22 seconds on average.