一种基于通道图卷积的手势识别框架

目前基于肌电信号的手势识别工作主要采用图像表示方法, 而对图像进行全局特征提取难以捕捉多个通道信号间的相互影响。文章设计一种基于通道图卷积(channel graph convolution, CGC)的手势识别框架, 用以挖掘多个通道信号间的关系。首先提出肌电信号的环形图表示, 将相邻通道的手势信号有机地联系在一起; 然后对环形图进行图卷积提取特征, 并通过图的剪枝自适应选取重要特征; 最后特征融合, 进行手势识别。该框架在 2 个公开基准数据集 Ninapro DB1 和 Ninapro DB2 上进行了评估, 分别取得 84.26% 和 86.42% 的识别准确率, 性能达到先进水平。

Currently, gesture recognition based on electromyography (EMG) signals primarily adopts image representation methods, making it difficult to capture the interactions between multiple channel signals through global feature extraction of images. This paper proposes a gesture recognition framework based on channel graph convolution (CGC), which aims to explore the relationships between multiple channel signals. Firstly, the paper proposes a circular graph representation of EMG signals, organically connecting signals from adjacent channels. Then, it performs graph convolution on the circular graph, and important features are adaptively selected through graph pruning. Finally, it conducts feature fusion for gesture recognition. The framework is evaluated on two publicly available benchmark datasets, Ninapro DB1 and Ninapro DB2, achieving recognition accuracies of 84.26% and 86.42%, respectively, demonstrating state-of-the-art performance.