基于胶囊网络的WiFi手势识别方法

WiFi Gesture Recognition Method Based on Capsule Network

在WiFi手势识别中,针对深度学习模型训练成本高、准确率低的问题,提出了一种基于胶囊网络迁移学习的GRU-CapNet方法。针对提取手势特征过程中重采样操作导致特征与数据直接映射关系变弱的问题,提出基于动态信道状态信息(Channel State Information,CSI)分割模式的身体速度谱(Body-coordinate Velocity Profile,BVP)估计算法,通过动态计算分割模式建立数据与特征间的直接映射关系。在公开室内数据集试验中,经优化后特征提取算法所得数据集上GRU,LSTM,CNN-GRU,3D-CNN和GRU-CapNet模型的识别准确率提升1.57%~3.34%,特征计算时间减少超过6 s;使用胶囊网络迁移学习所得识别准确率较未迁移提升7.87%,识别准确率在训练迭代仅10次后达到90%。在小型家用轿车场景应用试验中,针对驾驶员4种手势的平均识别准确率达到97.95%。

In WiFi gesture recognition,a GRU-CapNet gesture recognition method based on capsule network transfer learning is proposed to solve the problem of high training cost and low accuracy of deep learning model.To address the problem that the direct mapping relationship between features and data becomes weak due to resampling operation in the process of gesture features extraction,a Body-coordinate Velocity Profile(BVP)estimation algorithm based on dynamic Channel State Information(CSI)segmentation pattern is proposed,which establishes direct mapping relationship between features and data by dynamically computing the segmentation pattern.In a public indoor dataset test,the recognition accuracy of GRU,LSTM,CNN-GRU,3D-CNN and GRU-CapNet models on the dataset obtained by the optimized feature extraction algorithm is improved by 1.57%~3.34%respectively,while the feature computation time is reduced by over 6 s.The recognition accuracy obtained by using capsule network transfer learning is improved by 7.87%and the recognition accuracy exceeds 90%after only 10 training iterations.In an application of a sedan scenario,the average recognition accuracy of the driver’s four gestures reaches 97.95%.