Artificial Intelligence Algorithm And Embedded System Design For Continuous Motion Estimation

State-of-the-art robots can perform intensive workloads in harsh environments,however,they are still incapable to achieve autonomous decision-making in a complex environment.Especially in medical and military applications,efficient human-machine cooperation can enhance the safety and reliability levels of mission-critical tasks.Consequently,the design of a more natural and efficient human-machine interaction system has become the research hotspot.Surface Electromyography(sEMG)contains abundant information on motion intention,which has the characteristics of easy collection and ahead of the motion.sEMG provides an excellent interface for human-machine interaction.However,modern sEMG induced motion estimation methods are mostly designed for discrete classification,which does not meet the requirements of refined interaction in advanced scenarios.Furthermore,current methods suffer in flexibility,stability and latency during the extraction of motion information from sEMG.To address such issues,this thesis proposes a series of continuous motion estimation methods based on sEMG,which aims to design a fine-grained,natural and flexible human-machine interaction system.The main contributions of this thesis are as follows:1.Propose a Time-Domain Deep Learning Feature set(TDDLF)and present a comparison with existing feature sets.TDDLF addresses the problem that current deep learning methods for continuous motion estimation have a poor ability to extract original EMG features.This work leads to a paper in “IEEE Robotics and Automation Letters”(JCR Q1).2.Propose a Long Exposure Convolutional Memory Network(LE-ConvMN)to improve the accuracy of continuous motion estimation.LE-Conv MN effectively extracts the temporal and spatial distribution of sEMG signals.The experimental result proves a significantly higher accuracy than existing methods.This work has been published in “Journal of Neural Engineering”(JCR Q1).3.Propose a Multi-Attention Feature Fusion Network(MAFN)based on Transformer to further improve the number of compatible actions and the inference speed while ensuring the accuracy of the algorithm.Measurements on a mobile device show an inference latency of only 82 ms with 28 actions.This work has been submitted to IEEE Transactions on Human-Machine Systems.4.Design of a low-power,low-cost,real-time inertial data glove for human-computer interaction.The data glove is used for hand gesture acquisition during the research of sEMG induced continuous motion estimation.