An Active Perception System of Upper Limb Rehabilitation Robot Based son Human Body Action Feedback

An active rehabilitation training method based on the feedback of human action command is proposed. Firstly, the 4 channels of upper limb electromyogram (EMG) signals are decomposed by wavelet packet. The log features of wavelet coefficients are extracted and input into neural network for identification. The recognition accuracy of 8 common movements of the upper limb is 95.8%. Subsequently, the action of identification result is displayed to the patient with a virtual animation, and the patient nods head to accept the identification action or shakes the head to reject it. Kinect is used to get the patient movement video, and the detection range is narrowed through face recognition. Combining the characteristics of light flow with the prior experience, the action recognition results of the voting method, the Sigmoid method and the tanh method are compared and analyzed. When adding jamming artificially, the Sigmoid method still gets a good result. And then the parameters of the Sigmoid method are corrected, which leads to the accuracy of 87.25% and the no-hazard misjudgement of 12.75%. In the active rehabilitation experiment, the proposed active rehabilitation intention perception method based on human action command feedback shows fine effect, which indicates it is suitable for active rehabilitation training.