Transfer Learning With Small Data Samples For Human Activity Recognition Based On Wireless Sensing

Human activity recognition is one of the important research directions in the field of pattern recognition and has important theoretical and application values.Traditional human activity recognition based on visible light images has made important progress and has been successfully applied in the field of intelligent monitoring.But,it has the disadvantage that the recognition rate is greatly affected by light and the risk of privacy and identity leakage,which limits its application scope.Radio-sensing-based human activity recognition by analyzing the effect of human activity on wireless signal as Wi-Fi,the detection and recognition of human activity is achieved.It does not rely on cameras and wearable devices,is not limited by visual distance and lighting conditions,and can achieve effective human activity analysis,detection and recognition in real scenarios.In addition,because wireless sensing is a contactless working mode,there is no need to collect faces or human appearance,and this process does not involve user privacy,which has the advantages of easy to deploy at scale and low cost.Currently,Human activity recognition based on radio signals mainly includes two types of methods based on Received Signal Strength Indication(RSSI)and based on Channel State Information(CSI).Compared with RSSI,CSI provides more fine-grained information,which is conducive to matching human activities from multiple dimensions,thus effectively improving the recognition accuracy.Due to the influence of human activities on the propagation characteristics of wireless signals,it is difficult to carry out accurate modeling,and it is difficult to achieve better results based on traditional statistical signal processing.Since the training and recognition of deep learning do not depend on accurate modeling,scholars at home and abroad have explored the use of deep learning methods for wireless sensing in recent years,and has been made some important progress.However,wireless sensing suffers from the problems of lack of training samples and poor scene generalization performance of the training model,which hinder its wide application and promotion.To address this problem,this thesis studies human activity recognition based on transfer learning with small sample radio vision,and proposes two deep transfer learning models with higher accuracy and generalization ability,which effectively improve human activity recognition accuracy and scene transfer and generalization performance.Firstly,a human activity recognition method based on model retraining of small sample sensing images is proposed to address the problems of small sample size and poor generalization ability of deep neural networks in the current publicly available CSI dataset.First,the data preprocessing is performed to convert the source domain and target domain datasets into wireless sensing image set;the DenseNets pre-trained model is obtained by training on the ImageNet dataset with ten million images,and then it refers to the WAR data collection method,The data collection and data preprocessing are carried out in the three different laboratory environments,and wireless sensing image small sample datasets are obtained.Finally,the small sample dataset is loaded on the pre-trained model for retraining,and the optimized model is obtained.Through experimental verification,we can obtain the model with excellent performance and stability under the small sample wireless sensing data,which can effectively improve recognition accuracy.In addition,the prediction performance of the model is experimentally verified,and the results show that the proposed model has a recognition accuracy of more than 93% in three different laboratory environments,which verifies that the model has good generalization ability.Secondly,we propose a human activity recognition method based on domain adaptation for small sample wireless sensing images.To improve the generalization performance of the neural network model,converting the CSI dataset of the used source and target domains into wireless sensing image dataset;the parameters of the pre-trained residual network ResNets model trained on the ImageNet dataset are loaded on the layer where the overall network structure for model fine-tuning,and the local maximum mean difference(LMMD)constraint is introduced.Domain adaptation is achieved by aligning the data distribution of the relevant sub-domains in different domains and calculating the LMMD loss.Finally,experimental validation is performed on different indoor environment datasets,and the results show that the recognition accuracy of the domain adaptation method of transfer learning is above 94%,which verifies that the model has good scene generalization ability and good application prospects.In conclusion,converting one-dimensional CSI into two-dimensional wireless sensing images,and expands the wireless sensing image dataset,and two models of deep transfer learning with good generalization performance based on small sample wireless sensing images,which effectively improve the recognition accuracy of human activities in different environments and are of great value for the promotion and application of human activity recognition.