Intuitive motion classification from EMG for the 3-D arm motions coordinated by multiple DoFs

Surface Electromyography (EMG) has been considered as a viable human-machine interface in the context of human-centered robotics. In order to interpret human muscle activities into motion intentions, various pattern classification methods was proposed for human motion/gesture classification, which provided binary command for myoelectric control. To obtain complex motions coordinated by multiple DoFs, single DoF was usually sequentially classified and activated, which is not intuitive and efficient comparing with the natural motor strategy of the human. In this work, we investigated the motion classification methods from EMG for intuitive and simultaneous activation of multiple DoFs during 3-D arm motions. In the experiments, all motions were performed naturally rather than under the condition of maximum muscle contractions or other kinematic constraints. The combination of two EMG time-domain features after principal component analysis (PCA) processing is considered as the suitable choice considering both the classification accuracy and feasibility for robot control. For the motion classification method, least-square support vector machine (LS-SVM) represents higher classification accuracy for five arm motion classification across eight subjects with respect to other four methods which were popularly used in the previous works. The proposed method is hopefully applied in a EMG-driven simultaneous and proportional kinematics estimation systems for decoding model selection according to the motion intention.