A Power Wheelchair Controlled using Hand Gestures, a Single sEMG Sensor, and Guided Under-determined Source Signal Separation Luis A. Rivera, Student Member, IEEE, and Guilherme N. DeSouza, Senior Member, IEEE Department of Electrical and Computer Engineering, University of Missouri Abstract—Surface Electromyographic signals (sEMG) find applications in many areas such as rehabilitation, prosthesis and human-machine interaction. Systems reliant on these muscle- generated electrical signals require some form of machine learning algorithm for recognition of specific patterns of muscle activity. Those systems vary in terms of the signal detection methods, the feature selection and the classification algorithm used, however, in all those cases, the use of multiple sen- sors is a constant requirement. In this paper, we present a power wheelchair control system that relies on a single sEMG sensor and a new technique for signature recognition called Guided Under-determined Source Signal Separation (GUSSS). Compared to other approaches in the literature, the proposed technique achieves comparable results even when using a simple distance classifier and a very small number of features. I. I NTRODUCTION The ability to recognize Motor Unit Action Potential Trains (MUAPT) using electromyographic signals collected at the surface of the skin (sEMG) have been used in many applications, including rehabilitation, prosthesis, computer interfacing, exoskeleton robotics, etc. [1], [2], [3], [4], [5], [6]. When it comes to assistive technology, more specifically for power wheelchair control, sEMG signals have often been used as on/off switches. In those cases, menu driven ap- proaches [7], finite state machines [8], and a combination of multiple muscles and sensors [9] are common techniques em- ployed to expand these simple on/off patterns of activation. In general, sEMG-based systems require more sophisticated pattern recognition techniques and they vary widely in terms of the classification approach employed, the feature selection criteria, and the number of sensors used [10], [8], [9]. In terms of the classification algorithm, the most common methods used to classify muscle activity are Artificial Neural Networks (ANN) [4], [11], [5], Fuzzy Logic and Fuzzy Control systems [4], [12]. For example, in [4] an ANN was compared to a Fuzzy Inference System (FIS) for classifica- tion and control of a hand prosthesis. In this work, the authors concluded that for their application the best performance was using the FIS classifier which achieved 83% accuracy. In another work [5], several techniques for classification were employed in order to identify hand gestures using sEMG signals extracted from the forearm of human subjects. The authors compared the performance of ANN, Random Forest (RF), 1-Nearest-Neighbor (1NN), Support Vector Ma- chine (SVM), Decision Tree (DT) and Decision Tree with Boosting (DT/B) as possible classification techniques. They reported the ANN as the approach with best performance among those methods. In terms of feature selection, the features can be ex- tracted from time or time-frequency domains [4], [11], [3]. These features typically include: number of Zero Crossings (ZC), Mean Absolute Value (MAV), Slope Sign Changes (SSC), coefficients of Auto-regressive models (AR) [4], [11]; Absolute Maximum/Minimum, Maximum minus Minimum, Median Value (Med), Variance, Waveform Length (WL) [3]; coefficients of the Short Time Fourier Transform (STFT) [3]; Wavelets Transform (WT) [3], [2], etc. Given the wide range of features and their large dimension- ality, many systems also employ dimensionality reduction techniques. In those cases, Class Separability (CS), Principal Component Analysis (PCA), Analysis of Variance (ANOVA) or Multivariate ANOVA (MANOVA) are the techniques frequently used. In [4], for example, the authors developed a feature selection employing CS and PCA for dimensionality reduction. In that system, as well as in [5] where ANOVA was the technique of choice, the main concern was, as usual, to reduce dimensionality without affecting classification. Finally, in terms of number of sensors used, as far as we know all systems developed to date have relied on multiple sEMG signals and a large number of features. For example, in [4], the authors reported using two differential sEMG electrodes, multiple features, and PCA to reduce dimensionality of those features. In [5], the system relied on even more sensors – 5 to be more specific – and an ANN as the classification algorithm. As it can be inferred from the literature, one constant in most systems is the use of a large number of sensors and the use of sophisticated classification algorithms to help coping with a major disadvantage of surface EMG – i.e. the occurrence of cross-talk from adjacent muscles [1]. Our goal in this work is to present a much simpler and yet effective technique using a single EMG sensor, freeing other muscles to be used in other interfaces or to add modalities of operation to the interface. In this paper, we propose a system for operating a wheelchair that recognizes muscle movements derived from hand gestures. In our framework, we propose a new technique to separate the “cross-talked” MUAPTs signals from a sin- gle sEMG sensor called “Guided Under-determined Source The Fourth IEEE RAS/EMBS International Conference on Biomedical Robotics and Biomechatronics Roma, Italy. June 24-27, 2012 978-1-4577-1198-5/12/$26.00 ©2012 IEEE 1535