A Deep Q-Network based hand gesture recognition system for control of robotic platforms.

Hand gesture recognition (HGR) based on electromyography signals (EMGs) and inertial measurement unit signals (IMUs) has been investigated for human-machine applications in the last few years. The information obtained from the HGR systems has the potential to be helpful to control machines such as video games, vehicles, and even robots. Therefore, the key idea of the HGR system is to identify the moment in which a hand gesture was performed and it's class. Several human-machine state-of-the-art approaches use supervised machine learning (ML) techniques for the HGR system. However, the use of reinforcement learning (RL) approaches to build HGR systems for human-machine interfaces is still an open problem. This work presents a reinforcement learning (RL) approach to classify EMG-IMU signals obtained using a Myo Armband sensor. For this, we create an agent based on the Deep Q-learning algorithm (DQN) to learn a policy from online experiences to classify EMG-IMU signals. The HGR proposed system accuracy reaches up to [Formula: see text] and [Formula: see text] for classification and recognition respectively, with an average inference time per window observation of 20 ms. and we also demonstrate that our method outperforms other approaches in the literature. Then, we test the HGR system to control two different robotic platforms. The first is a three-degrees-of-freedom (DOF) tandem helicopter test bench, and the second is a virtual six-degree-of-freedom (DOF) UR5 robot. We employ the designed hand gesture recognition (HGR) system and the inertial measurement unit (IMU) integrated into the Myo sensor to command and control the motion of both platforms. The movement of the helicopter test bench and the UR5 robot is controlled under a PID controller scheme. Experimental results show the effectiveness of using the proposed HGR system based on DQN for controlling both platforms with a fast and accurate response.

Recognition of Hand Gestures Based on EMG Signals with Deep and Double-Deep Q-Networks.

In recent years, hand gesture recognition (HGR) technologies that use electromyography (EMG) signals have been of considerable interest in developing human-machine interfaces. Most state-of-the-art HGR approaches are based mainly on supervised machine learning (ML). However, the use of reinforcement learning (RL) techniques to classify EMGs is still a new and open research topic. Methods based on RL have some advantages such as promising classification performance and online learning from the user's experience. In this work, we propose a user-specific HGR system based on an RL-based agent that learns to characterize EMG signals from five different hand gestures using Deep Q-network (DQN) and Double-Deep Q-Network (Double-DQN) algorithms. Both methods use a feed-forward artificial neural network (ANN) for the representation of the agent policy. We also performed additional tests by adding a long-short-term memory (LSTM) layer to the ANN to analyze and compare its performance. We performed experiments using training, validation, and test sets from our public dataset, EMG-EPN-612. The final accuracy results demonstrate that the best model was DQN without LSTM, obtaining classification and recognition accuracies of up to 90.37%±10.7% and 82.52%±10.9%, respectively. The results obtained in this work demonstrate that RL methods such as DQN and Double-DQN can obtain promising results for classification and recognition problems based on EMG signals.

Hand Gesture Recognition Using EMG-IMU Signals and Deep Q-Networks.

Hand gesture recognition systems (HGR) based on electromyography signals (EMGs) and inertial measurement unit signals (IMUs) have been studied for different applications in recent years. Most commonly, cutting-edge HGR methods are based on supervised machine learning methods. However, the potential benefits of reinforcement learning (RL) techniques have shown that these techniques could be a viable option for classifying EMGs. Methods based on RL have several advantages such as promising classification performance and online learning from experience. In this work, we developed an HGR system made up of the following stages: pre-processing, feature extraction, classification, and post-processing. For the classification stage, we built an RL-based agent capable of learning to classify and recognize eleven hand gestures-five static and six dynamic-using a deep Q-network (DQN) algorithm based on EMG and IMU information. The proposed system uses a feed-forward artificial neural network (ANN) for the representation of the agent policy. We carried out the same experiments with two different types of sensors to compare their performance, which are the Myo armband sensor and the G-force sensor. We performed experiments using training, validation, and test set distributions, and the results were evaluated for user-specific HGR models. The final accuracy results demonstrated that the best model was able to reach up to 97.50%±1.13% and 88.15%±2.84% for the classification and recognition, respectively, with regard to static gestures, and 98.95%±0.62% and 90.47%±4.57% for the classification and recognition, respectively, with regard to dynamic gestures with the Myo armband sensor. The results obtained in this work demonstrated that RL methods such as the DQN are capable of learning a policy from online experience to classify and recognize static and dynamic gestures using EMG and IMU signals.

An Energy-Based Method for Orientation Correction of EMG Bracelet Sensors in Hand Gesture Recognition Systems.

Hand gesture recognition (HGR) systems using electromyography (EMG) bracelet-type sensors are currently largely used over other HGR technologies. However, bracelets are susceptible to electrode rotation, causing a decrease in HGR performance. In this work, HGR systems with an algorithm for orientation correction are proposed. The proposed orientation correction method is based on the computation of the maximum energy channel using a synchronization gesture. Then, the channels of the EMG are rearranged in a new sequence which starts with the maximum energy channel. This new sequence of channels is used for both training and testing. After the EMG channels are rearranged, this signal passes through the following stages: pre-processing, feature extraction, classification, and post-processing. We implemented user-specific and user-general HGR models based on a common architecture which is robust to rotations of the EMG bracelet. Four experiments were performed, taking into account two different metrics which are the classification and recognition accuracy for both models implemented in this work, where each model was evaluated with and without rotation of the bracelet. The classification accuracy measures how well a model predicted which gesture is contained somewhere in a given EMG, whereas recognition accuracy measures how well a model predicted when it occurred, how long it lasted, and which gesture is contained in a given EMG. The results of the experiments (without and with orientation correction) executed show an increase in performance from 44.5% to 81.2% for classification and from 43.3% to 81.3% for recognition in user-general models, while in user-specific models, the results show an increase in performance from 39.8% to 94.9% for classification and from 38.8% to 94.2% for recognition. The results obtained in this work evidence that the proposed method for orientation correction makes the performance of an HGR robust to rotations of the EMG bracelet.

Sign Language Recognition Using Wearable Electronics: Implementing k-Nearest Neighbors with Dynamic Time Warping and Convolutional Neural Network Algorithms.

We propose a sign language recognition system based on wearable electronics and two different classification algorithms. The wearable electronics were made of a sensory glove and inertial measurement units to gather fingers, wrist, and arm/forearm movements. The classifiers were k-Nearest Neighbors with Dynamic Time Warping (that is a non-parametric method) and Convolutional Neural Networks (that is a parametric method). Ten sign-words were considered from the Italian Sign Language: cose, grazie, maestra, together with words with international meaning such as google, internet, jogging, pizza, television, twitter, and ciao. The signs were repeated one-hundred times each by seven people, five male and two females, aged 29-54 y ± 10.34 (SD). The adopted classifiers performed with an accuracy of 96.6% ± 3.4 (SD) for the k-Nearest Neighbors plus the Dynamic Time Warping and of 98.0% ± 2.0 (SD) for the Convolutional Neural Networks. Our system was made of wearable electronics among the most complete ones, and the classifiers top performed in comparison with other relevant works reported in the literature.

Real-Time Hand Gesture Recognition Using Surface Electromyography and Machine Learning: A Systematic Literature Review.

Today, daily life is composed of many computing systems, therefore interacting with them in a natural way makes the communication process more comfortable. Human-Computer Interaction (HCI) has been developed to overcome the communication barriers between humans and computers. One form of HCI is Hand Gesture Recognition (HGR), which predicts the class and the instant of execution of a given movement of the hand. One possible input for these models is surface electromyography (EMG), which records the electrical activity of skeletal muscles. EMG signals contain information about the intention of movement generated by the human brain. This systematic literature review analyses the state-of-the-art of real-time hand gesture recognition models using EMG data and machine learning. We selected and assessed 65 primary studies following the Kitchenham methodology. Based on a common structure of machine learning-based systems, we analyzed the structure of the proposed models and standardized concepts in regard to the types of models, data acquisition, segmentation, preprocessing, feature extraction, classification, postprocessing, real-time processing, types of gestures, and evaluation metrics. Finally, we also identified trends and gaps that could open new directions of work for future research in the area of gesture recognition using EMG.

Optimal clustering under uncertainty.

Classical clustering algorithms typically either lack an underlying probability framework to make them predictive or focus on parameter estimation rather than defining and minimizing a notion of error. Recent work addresses these issues by developing a probabilistic framework based on the theory of random labeled point processes and characterizing a Bayes clusterer that minimizes the number of misclustered points. The Bayes clusterer is analogous to the Bayes classifier. Whereas determining a Bayes classifier requires full knowledge of the feature-label distribution, deriving a Bayes clusterer requires full knowledge of the point process. When uncertain of the point process, one would like to find a robust clusterer that is optimal over the uncertainty, just as one may find optimal robust classifiers with uncertain feature-label distributions. Herein, we derive an optimal robust clusterer by first finding an effective random point process that incorporates all randomness within its own probabilistic structure and from which a Bayes clusterer can be derived that provides an optimal robust clusterer relative to the uncertainty. This is analogous to the use of effective class-conditional distributions in robust classification. After evaluating the performance of robust clusterers in synthetic mixtures of Gaussians models, we apply the framework to granular imaging, where we make use of the asymptotic granulometric moment theory for granular images to relate robust clustering theory to the application.