Human Hand Movement Recognition based on HMM with Hyperparameters Optimized by Maximum Mutual Information

Performing dexterous and versatile movements is essential for multi-finger manipulators for human-robot collaboration, and designing effective control methods for the robotic manipulator is challenging. To recognize human hand movements, we used surface electromyography (sEMG) for sensing myoelectric activity due to its portability and low-cost compared to cameras, and proposed a hidden Markov model (HMM) based method to characterize the transition of action primitives. For building HMMs for hand movements, the hyperparameters, including features, the window length and the number of states, are optimized by the maximum mutual information (MMI) criterion. The optimal features - marginal Discrete Wavelet Transform (mDWT) and mean value - are extracted from multichannel signals acquired from 12 electrodes. Our proposed method is validated by recognizing 40 hand movements from activities of daily living (ADL) in the second NinaPro database. Using MMI as the optimization criterion for hyperparameters, we have improved the average recognition accuracy over 40 subjects in the database from 92.02% to 97.32%.