Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

CNN-Based Detection and Classification of Grasps Relevant for Worker Support Scenarios Using sEMG Signals of Forearm Muscles

: Maufroy, Christophe; Bargmann, Daniel


Institute of Electrical and Electronics Engineers -IEEE-:
IEEE International Conference on Systems, Man, and Cybernetics, SMC 2018. Proceedings : 7-10 October 2018, Miyazaki, Japan
Piscataway, NJ: IEEE, 2018
ISBN: 978-1-5386-6650-0
ISBN: 978-1-5386-6651-7
International Conference on Systems, Man, and Cybernetics (SMC) <2018, Miyazaki>
Fraunhofer IPA ()
Mensch-Maschine-Interaktion; Exoskelett; neuronales Netz; EMG; Griff; Hebehilfe

Intuitive man-machine interaction is one of the main challenges towards a wider use of exoskeletons as assistive systems to reduce fatigue and alleviate the risk of injuries of workers during manual handling or overhead work. In this context reliable human grasp detection and classification is a basic but key component to estimate current state and context. While hand gesture recognition using surface electromyography (sEMG) has been addressed in a number of studies, there has been only little research regarding the detection and classification using sEMG of grasps relevant for worker support scenarios, which is the scope of the present study. Specifically, we investigated detection and classification of one-handed power grasps (relevant for tool manipulation) and two-handed grasps (typical while handling crates and boxes) using the instantaneous sEMG vector of the forearm muscles activity acquired using a consumer-grade wearable sensor bracelet. Detection and classification was performed by convolutional neural networks (CNN) trained using datasets including multiple grasps, positions, sessions and users. The accuracy of grasp detection and classification was evaluated in different scenarios with increasing level of signal variability, including sample cross-validation, inter-session scenario and inter-person scenario. We found that high accuracy of grasp detection can be achieved, even in the inter-person scenario. For one-handed grasps, reasonable accuracy of classification of the object manipulated by the user was possible in the inter-person scenario. For two-handed grasps, reasonable accuracy of grasp type classification was possible for some of the grasps even in the inter-person case. These results, achieved with a simple, low-profile acquisition device, show the potential of proposed approach in the context of worker support scenarios.