This study researched a robust gesture recognition algorithm based on EMG. The proposed algorithm only needs very limited training data (1 or 2 training trials for each gesture). The contribution of the proposed algorithm was mainly three-fold. First, a shrinkage approach was applied to estimate the samples' covariance matrix, which helped to improve the robustness of the algorithm. Second, to evaluate the system performance, classification accuracy and gesture number to be recognized was compromised using information transfer rate (ITR). We found a system which can recognize 10 gestures could achieve similar ITR as the system which can recognize 20 gestures. However, the 10-gesture system was more robust. Third, K-L divergence was used to evaluate the separability of the EMG signals from different gestures. The result of a 5 subject experiment showed that the classification accuracy of 10 gestures using 2 trials as training set can reach 85%.
展开▼
