融合多源体感信息的人体上肢动作识别与情感推断研究

At present, the studies of human motion recognition based on body sensor networks has made some progress, but the researches on human upper limb motion recognition is relatively few, and the research on human emotion inference based on motion recognition analysis is still in the enlightenment stage. This project develops the research of human upper limb motion recognition and emotional inference based on body sensor networks. Firstly, the traditional feature extraction methods of body sensor signals lack accurate description and are easily distorted. To solve this problem, this project studies how to extract new features of body sensor signals by deep learning method in order to achieve accurate representation of body sensor signals. Secondly, aiming at the complexity and similarity of upper limb motions, a new framework of human upper limb motion recognition based on body sensor networks is designed. Finally, this project explores the mechanism of human upper limb motions and emotional inference, and establishes an emotion recognition method based on human motion and body sensor signals, so as to realize the accurate recognition of human emotions. Through this study, it can provide a certain scientific support for the follow-up research of emotional recognition, and it can also open up a new direction for the popularization and application of body sensor networks technology, and this study has important theoretical and practical significance.

现阶段，基于体感网的人体动作识别研究已取得的了一定的进展，但是对于人体上肢动作识别与分析研究相对较少，而基于动作识别分析的人体情感推断研究更是处于启蒙阶段。本项目主要利用体感网，开展人体上肢动作识别与情感推断方面的研究。首先，针对传统的体感信号特征提取方法存在的缺乏精确性描述、易失真等问题，研究利用深度学习方法提取新的体感信号特征，以实现对体感信号的准确表示。其次，针对上肢动作存在的复杂性以及相似性，设计基于体感网的新的人体上肢动作识别框架方案，以实现对上肢动作的精确识别。最后，探索人体上肢动作与情感推断的相关机制，建立基于人体动作和体感信号的情感识别方法，实现人体情感的准确识别。通过本项目研究，可为后续的情感识别研究提供一定的科学支撑，也能够为体感网技术的推广应用开拓方向，具有重要的理论与应用意义。

人体动作可以表达人体情感，如果智能系统能从人体动作中识别出人的情绪状态，人与机器的交互就会变得更加友好和自然。而目前基于人体动作识别分析的情感推断研究非常少。本项目主要开展人体动作识别与情感推断方面的研究。主要研究内容包括：(1) 在特征提取方面，基于人体3D骨架，设计了一种轻量级卷积神经网络的特征提取方法，同时针对人体惯性数据，提出了一种一种基于注意力机制的特征融合方法。(2) 针对用户在绑定惯性传感器时，不能保证每次的固定位置完全一致这一问题，提出一种利用旋转矩阵实现惯性传感器位置校正的人体动作识别方法。同时设计了基于熵权和模糊规则的分层融算法，以提高识别系统的识别精度。(3) 在人体情感识别与推断方面，研究了人体动作与情感的映射关系，同时设计基于深度模型的人体情感识别方法。另外本项目对算法模型优化和算法设计进行了细致的研究，进一步拓展了研究思路。