Pattern recognition based on HD-sEMG spatial features extraction for an efficient proportional control of a robotic arm

• Published in: Biomedical signal processing and control 2019-08, Vol.53, p.101550, Article 101550
• Main Authors: Nougarou, F., Campeau-Lecours, A., Massicotte, D., Boukadoum, M., Gosselin, C., Gosselin, B.
• Format: Article
• Language: English
• Subjects: Assistive robotics; HD-sEMG; Linear regression; Pattern recognition; Sequentiel and propotionnal control; Spatial features; Spatial filtering
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2019.04.027

•Pattern recognition based on a HD-sEMG sensor and spatial features to characterize the muscular distribution.•Higher accuracy classification and more robust to the muscular contraction variations than time-domain features.•Real-time efficient control of an assistive robotic arm by simultaneously providing both the direction and the velocity. To enable an efficient alternative control of an assistive robotic arm using electromyographic (EMG) signals, the control method must simultaneously provide both the direction and the velocity. However, the contraction variations of the forearm muscles, used to proportionally control the device’s velocity using a regression method, can disturb the accuracy of the classification used to estimate its direction at the same time. In this paper, the original set of spatial features takes advantage of the 2D structure of an 8 × 8 high-density surface EMG (HD-sEMG) sensor to perform a high accuracy classification while improving the robustness to the contraction variations. Based on the HD-sEMG sensor, different muscular activity images are extracted by applying different spatial filters. In order to characterize their distribution specific to each movement, instead of the EMG signals’ amplitudes, these muscular images are divided in sub-images upon which the proposed spatial features, such as the centers of the gravity coordinates and the percentages of influence, are computed. These features permits to achieve average accuracies of 97% and 96.7% to detect respectively 16 forearm movements performed by a healthy subject with prior experience with the control approach and 10 movements by ten inexperienced healthy subjects. Compared with the time-domain features, the proposed method exhibits significant higher accuracies in presence of muscular contraction variations, requires less training data and is more robust against the time of use. Furthermore, two fine real-time tasks illustrate the potential of the proposed approach to efficiently control a robotic arm.