Diseño y comparación de modelos de reconocimiento de gestos de la mano combinando aprendizaje supervisado y aprendizaje por refuerzo : Diseño de un modelo de reconocimiento de 5 gestos de la mano que funcione en tiempo real usando aprendizaje supervisado DQN (Deep Q-learning).

Abstract

Hand gesture recognition (HGR) is an important application of human-computer interaction (HCI) using electromyographic (EMG) signals and deep learning techniques. In the present work, two HGR models for general users are developed using the Deep Q-Network (DQN) reinforcement learning algorithm and the Double Deep Q-Network (DDQN) algorithm, both based on the EMG-EPN-612 dataset. The developed models are a fine tuning of a pre-trained model with supervised learning and based on a convolutional network (CNN). For such refinement, a reward function capable of rewarding constancy in model predictions through gradual rewards was designed. A post-processing of the model was also performed based on the fashion trend measure to deal with erroneous intermediate labels. For testing, experiments were performed on a publicly accessible dataset containing 612 users. Subsequently, a measurement and comparison of the recognition and classification accuracy between 6 gestures, including "no gesture", was performed. The results showed that the fine-tuned models with DQN and DDQN present similar values in terms of the stated metrics. The DQN model presents a 3.61% improvement and the DDQN a 3.45% improvement in recognition accuracy with respect to the base model. On the other hand, classification accuracy does not show any significant improvement. Finally, we conclude that DQN presents an overall better performance than DDQN in classification and hand gesture recognition tasks.