Comparative Study on Gesture Recognition Using Multiple Kernel Learning via Multi-mode Information Fusion

Abstract

An approach is introduced to the design of a multi-mode information fusion classifier to combine an inertial measurement unit with multichannel surface electromyography (sEMG) sensors to implement gesture control for a mobile robot movement, which can exceed the required control performance of vison-based methods in terms of portability, robustness, intuitiveness, and availability. A comparison of 4 groups of feature extraction project and multiple kernel relevance vector machine (MKRVM) based on multiple kernel expansion via kernel alignment were used to get better recognition performance. It was found that feature extraction method which combined time-domain analysis and time-frequency domain analysis can obtain better performance. Then, after comparing experiments, it was proved MKRVM based on multiple kernel expansion via kernel alignment not only achieved a higher recognition rate, its generalization ability was also significantly better than the traditional MKRVM. Genetic algorithms (GA) are used to optimize the best parameters for each kernel of the MKRVM algorithm. In the online robot control experiment, the gesture online identification system can accurately identify the operator gestures in real time and accurately control the youbot robot to move and do simple assembly operations.