WiGNet:一种适用于无线感知场景的手势识别模型

WiGNet: A Gesture Recognition Model for the Wireless Sensing Scenario

针对现有Wi-Fi感知技术的手势识别研究中存在因感知特征辨识度较低及难以提取而导致识别准确率不高的问题,在对感知数据进行多维重构并施以二维离散小波变换的基础上,提出了一种基于卷积神经网络的高精度手势识别模型WiGNet。首先,在对信道状态信息影响因子分析的基础上,提取振幅数据作为手势识别的基础数据;其次,将感知数据重构为多维矩阵形式,以充分提取手势动作的时空特征;然后,对重构数据在时间和空间维度上进行二维离散小波变换,以实现感知数据的降噪和平滑;最后,使用神经架构搜索技术优化网络深度以及卷积核个数的合理化配置,并提出较好适配重构数据的手势识别模型WiGNet。实验结果表明:对感知数据进行二维离散小波变换后,特征辨识度和模型整体运算速度均有显著提升;WiGNet在自建数据集上和公共数据集上的平均识别准确率分别达到98.1%和96.0%,均优于同类模型。此外,WiGNet在实现高精度手势识别的同时,还能保持较快运算速度,并兼具一定的鲁棒性。

Existing research suggests low accuracy of gesture recognition based on the WiFi sensing technology because of difficulties in extracting hard-to-recognize perceptual features.To solve this problem,this paper,through multidimensional reconstruction of perceptual data and two-dimensional discrete wavelet transform,proposes a high-precision gesture recognition model WiGNet based on the convolutional neural network.Firstly,based on the analysis of influence factors of channel state information,the amplitude data is extracted as the base data for gesture recognition;secondly,the perception data is reconstructed into a multidimensional matrix to fully extract the spatio-temporal characteristics of gestures;next,the two-dimensional discrete wavelet transform is applied to the reconstructed data in the temporal and spatial dimensions to reduce noise and smooth the perceptual data;finally,a neural architecture search technique is used to optimise the network depth and rationalise the number of convolutional kernels.Thus WiGNet,a gesture recognition model that can be better adapted to reconstructed data,is proposed.The experimental results show that the model significantly improves the feature identification accuracy and has a higher overall operation speed after the perception data is processed by two-dimensional discrete wavelet transform.WiGNet achieves an average recognition accuracy of 98.1%and 96.0%based on the self-built dataset and the public dataset respectively,outperforming similar models.In addition,WiGNet can not only achieve high-precision gesture recognition,but also maintain a fast computing speed and have a certain degree of robustness.