基于反残差自注意力模块与多模态、多特征融合的步态识别算法研究OA
A novel gait recognition algorithm based on inverted residual self-attention module and multi-modality,multi-feature fusion
基于可穿戴设备的步态识别是实现下肢穿戴设备人机交互的基础.考虑到单一模态、单一特征不能提供完整的信息来识别步态,且当前步态识别领域缺少结合注意力机制的轻量化多模态算法研究,本文提出一种融合深度学习特征与传统时域特征的轻量化多模态步态识别算法,探索多模态信号融合中2 种特征新的融合范式.该算法分别对姿态信号和肌电(electromyography,EMG)信号同时提取时域特征与深度学习特征,采取层级式融合方式将二者分步融合.基于自注意力机制和倒置残差结构设计了一个轻量化的特征提取模块,在保留卷积神经网络提取能力的同时也增加了远程信息交互能力.同时,设计了一个融合模块来融合不同模态的特征,增强特征中有用信息同时忽略次要信息.该模型在公开数据集 ENABL3S 和自建数据集上的准确率分别为98.16%和95.38%.此外,本文还分析了惯性测量单元(inertial measurement unit,IMU)和 EMG 这 2 种模态的 2种特征分别对步态识别精度的贡献,为后续步态识别算法的研究者提供了参考.
Gait recognition is fundamental to human-machine interaction for lower-limb wearable devices.Considering that a single modality and unitary feature cannot provide comprehensive information for gait recognition,and that there is limited number of researches incorporating attention mechanisms,this paper proposes a lightweight multi-modal gait recognition algorithm that integrates deep learning features with traditional time-domain features and ex-plores a new fusion paradigm of the two types of features.The proposed algorithm extracts time-domain features and deep learning features from both posture signals and electromyography(EMG)signals simultaneously and then adopts a hierarchical fusion method to integrate them step by step.A lightweight feature extraction module based on the self-attention mechanism and inverted residual structure is designed to retain the extraction capability of convo-lutional neural networks and increase the capacity for remote information interaction.Additionally,a fusion module is designed to integrate features from different modalities,enhancing the useful information in the features while ig-noring the unimportant information.The proposed algorithm achieves accuracies of 98.16%and 95.38%on a pub-lic dataset ENABL3S and a self-collected dataset,respectively.Moreover,the contribution of inertial measurement unit(IMU)and EMG features to the gait recognition accuracy is analyzed,providing references for researchers in the subsequent gait recognition algorithm research.
赵昶辰;刘凯;黄斌;冯远静
浙江工业大学信息工程学院 杭州 310023浙江工业大学信息工程学院 杭州 310023北京航空航天大学杭州创新研究院 杭州 310051浙江工业大学信息工程学院 杭州 310023
步态识别肌电惯性测量单元自注意力时域特征
gait recognitionelectromyographyinertial measurement unitself-attentiontime domain fea-ture
《高技术通讯》 2026 (7)
677-687,11
国家自然科学基金面上项目(62573171)和国家自然科学基金青年基金(62203037)资助项目.
评论