传感技术学报2025，Vol.38Issue(11)：1990-1999,10.DOI:10.3969/j.issn.1004-1699.2025.11.009
基于知识蒸馏的CSI手势识别
Knowledge Distillation-Based Gesture Recognition Using CSI
摘要
Abstract
In the field of gesture recognition,current methodologies are challenged by inefficiencies in fine-grained feature extraction,high deployment costs for models,and suboptimal robustness.To address these issues,a wireless gesture recognition method based on knowledge distillation is proposed.The approach begins with data preprocessing of channel state information(CSI),employing conjugate multiplication to eliminate phase offset noise.This is followed by the use of a Butterworth filter to eradicate static and high-frequency noise,and then principal component analysis is applied to extract the primary components of the signal.This method enables a student model to learn from the output of a teacher model through knowledge distillation,resulting in an understanding of CSI features and the teacher model's interpretation of body coordinate velocity spectrum.This technique not only maintains accuracy in recognition but also reduces the size of the model.Tests conducted on the Widar3.0 dataset demonstrate that the model's parameters are reduced by approxi-mately 1.6 times,achieving a recognition accuracy of 94.75%.Compared to existing methods,the gesture recognition method proposed shows superior performance,offering a novel solution for wireless signal gesture recognition.关键词
手势识别/信道状态信息/WiFi/知识蒸馏Key words
gesture recognition/channel state information/WiFi/knowledge distillation分类
信息技术与安全科学引用本文复制引用
黄子非,朱海,龚浩成,杨明泽,吴飞..基于知识蒸馏的CSI手势识别[J].传感技术学报,2025,38(11):1990-1999,10.基金项目
国家自然科学基金青年科学基金项目(61902237) （61902237）
