Li Chuanjiang, Ding Xinhao, Tu Jiajun, Li Ang, Zhu Yanfei, Gu Ya, Zhi Erlei
The College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 200233, China.
The College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 200233, China.
J Neurosci Methods. 2025 Jul;419:110469. doi: 10.1016/j.jneumeth.2025.110469. Epub 2025 May 6.
Gait recognition based on surface electromyography (sEMG) signals has many applications in exoskeleton control. However, due to the irrelevance and redundancy of its features, how to extract features effectively and improve the recognition accuracy is a hotspot of current research.
This study proposes a progressive feature selection (PFS) gait recognition method based on sEMG. First, to solve the problem of inaccurate gait description, the stereo modelling projection and 3D dynamic capture are fused to capture the time and frequency domain features derived from the four muscles of the human lower limb according to the gait phase. Then, to address the problem of poor gait classification accuracy, a progressive feature combination optimization is performed based on the fitness evaluation to preserve the key information embedded in the features while eliminating features that contribute less to the model accuracy. Therefore, model accuracy is improved by determining the best combination of features.
The progressive feature selection method shows considerable performance in sEMG-based gait recognition, with the average accuracy of 98.54 % and the median accuracy of 98.67 %.
In order to verify the effectiveness of the proposed algorithm more comprehensively, the practical experimental dataset and the publicly available SIAT-LLMD dataset are adopted respectively. Compared with the state-of-the-art methods, the gait recognition accuracy of the proposed PFS algorithm can reach 98.91 % and 98.54 %.
The proposed PFS gait recognition method can significantly reduce unnecessary features, thus improving the recognition accuracy and safety of lower limb exoskeleton robots.
基于表面肌电图（sEMG）信号的步态识别在外骨骼控制中有许多应用。然而，由于其特征的不相关性和冗余性，如何有效提取特征并提高识别准确率是当前研究的热点。
本研究提出了一种基于sEMG的渐进特征选择（PFS）步态识别方法。首先，为了解决步态描述不准确的问题，将立体建模投影和3D动态捕捉相结合，根据步态阶段捕捉源自人体下肢四块肌肉的时域和频域特征。然后，为了解决步态分类准确率低的问题，基于适应度评估进行渐进特征组合优化，以保留特征中嵌入的关键信息，同时消除对模型准确率贡献较小的特征。因此，通过确定最佳特征组合来提高模型准确率。
渐进特征选择方法在基于sEMG的步态识别中表现出相当出色的性能，平均准确率为98.54%，中位数准确率为98.67%。
为了更全面地验证所提算法的有效性，分别采用了实际实验数据集和公开可用的SIAT-LLMD数据集。与现有最先进方法相比，所提PFS算法的步态识别准确率分别可达98.91%和98.54%。
所提PFS步态识别方法可显著减少不必要的特征，从而提高下肢外骨骼机器人的识别准确率和安全性。
