Multi-feature fusion method for hand motion recognition based on multiple kernel learning

Abstract

In order to improve the performance of hand multi-movement pattern recognition in myoelectric control system, a multi-feature fusion method is proposed. The effect of identification method based on single surface electromyography (sEMG) feature is limited. To obtain more comprehensive information about multi-movement pattern, we investigate a multi-feature fusion method based on multiple kernel learning framework. General multi-kernel learning (Generalized-MKL) algorithm is adopted to integrate different kinds of features, which takes full advantage of valuable features for movement classification. Each kind of sEMG feature is endowed with a base kernel respectively. The proposed method adaptively combines multiple features together with optimal combination weights and finds a mixed kernel to classify. In the experiment, we classify multiple different hand movements for fifteen participants, such as wrist, gross hand and finger movements. The experimental results indicate that the feature set formed by the proposed fusion method has the lowest classification error value across all the multi-feature sets. The classification error value of the proposed feature set is 2.86%-4.24% lower than the other commonly used feature combinations. It proves the effectiveness of the proposed multi-feature fusion method. The proposed method is quite promising for hand multi-movement classification with low classification error value.