【作者】 李昕；

【导师】 赵健；

【作者基本信息】 西北大学 ， 信号与信息处理， 2022， 硕士

【摘要】 教学行为识别在智慧课堂中有着广泛的应用,是教育智能化的重要实现手段之一。目前主流的行为识别方式,主要分为基于摄像头的行为识别和基于专用传感器的行为识别两类。基于摄像头的识别方式具有潜在的隐私泄露风险,而专用传感器因其昂贵的造价和不适感也无法应用于现实课堂。基于Wi-Fi信号的信道状态信息(Channel State Information,CSI)的行为识别方案能够以低成本实现普适化的推广,并且完全规避了隐私泄露的风险。在基于CSI实现室内教学行为识别的过程中,由于CSI对实验环境和人体生理特征的敏感性,任何实验设置的改变都会对其识别结果造成影响。为提高在复杂场景下利用CSI实现室内教学行为识别的性能,针对实验设置的多样性,本文通过消除背景噪声来规避实验环境对CSI的影响,并提出了一种基于多特征融合M-LSTM的室内教学行为识别算法。该算法利用多层LSTM网络提取局部时序特征,并与统计得到的全局特征融合后送入注意力机制模块,提升了模型在复杂环境下的特征提取能力。针对模型在识别新用户时无法泛化,且新用户的标签样本在短时间内难以大量得到的问题。本文构建了一种基于SSGAN模型的新用户识别算法,通过Cycle GAN将新用户的样本迁移到已知数据集中,从而得到具有新用户生理特性信息的无标签样本。而后修改M-LSTM的输入输出作为SSGAN的鉴别器,以半监督学习的方式提升模型对新用户的识别性能。经不同实验数据集上的训练与验证结果表明,本文所提出的两种识别模型显著提高了在复杂环境下识别CSI的能力,并且对新用户具有的较好识别效果。更多还原

【Abstract】 Teaching behavior recognition has a wide range of applications in smart classroom and is one of the important methods of educational intelligence.At present,the widely used behavior recognition methods are mainly divided into two categories: camera-based behavior recognition and special sensor-based behavior recognition.Camera-based identification methods have potential privacy risks,and dedicated sensors cannot be used in real classrooms due to their high cost and discomfort.With CSI(Channel State Information)of Wi-Fi signal,behavior recognition can be performed at a low cost and completely avoid privacy risks.In the process of indoor teaching behavior recognition based on CSI,due to the sensitivity of CSI to the experimental environment and human physiological characteristics,any changes in experimental settings will affect the recognition results.In order to improve the performance of using CSI to realize indoor teaching behavior recognition in complex scenes,in view of the change of experimental settings,this paper eliminates background noise to avoid the influence of experimental environment on CSI,and proposes an indoor teaching behavior recognition algorithm based on multi-feature fusion M-LSTM.The algorithm uses a multi-layer LSTM network to extract local time series features,and fuses them with the global features obtained by statistics and sends them to the attention mechanism module,which improves the feature extraction ability of the model in complex environments.The model cannot recognize new users,and it is difficult to obtain a large number of label samples from new users in a short time.The paper builds an identification algorithm for new users based on SSGAN,and samples of new users are transferred to public datasets through Cycle GAN,so as to obtain unlabeled samples with information about the physiological characteristics of new users.M-LSTM outputs and inputs are modified as the discriminator of SSGAN,and the recognition performance for new users of the model is improved through semi-supervised learning.On different experimental datasets,training and validation results show that both recognition models significantly improve their ability to recognize CSI in complex environments as well as recognize new users.更多还原