Research On Human Activity Recognition Algorithm Based On Deep Learning

With the rapid development of embedded technology and artificial intelligence,ubiquitous computing and intelligent perception have emerged as the fundamental building blocks in the field of internet of things(Io T)applications,thus attracting significant research attention.Real-time and accurate recognition of human activities is essential for providing personalized services in the context of internet of things(Io T).Wearable devices are widely used in ubiquitous computing scenarios due to their small size,variety,low energy consumption,and good privacy protection.Human activity recognition based on wearable devices has become a hot topic in fall monitoring,human-computer interaction,and smart homes.The development of deep learning provides opportunities and challenges for the progress of human activity recognition.Therefore,this thesis focuses on studying the application of deep learning in wearablebased human activity recognition,exploring human activity recognition from the two directions of supervised and self-supervised learning.The main objectives of this research are as follows:(1)This thesis proposes a supervised learning framework for HAR named TTN,based on self-attention mechanism and two-stream structures.Previous supervised learning methods can effectively extract the temporal features of activities but ignore the spatial relationship between sensor streams.TTN uses the self-attention mechanism and two-stream structure to extract the spatial-temporal features of multi-sensor data.The features extracted from the two streams are complementary.The spatial features carry additional information that cannot be directly captured from the time series.In a multi-sensor environment,sensors placed at different positions on the body contribute differently to the recognition of different activities.TTN dynamically assigns weights to each sensor stream based on the importance of each stream in the spatial channel.More important streams are assigned larger weights.Finally,extensive experiments on four public datasets(PAMAP2,Opportunity,USC-HAD,and Skoda)show that TTN has higher recognition accuracy and lower time complexity than previous supervised learning methods.(2)This thesis proposes a self-supervised learning method for HAR based on Time-Frequency Contrasting(TFC),inspired by the idea of structured time series.The most pressing challenge faced by supervised learning methods is the need for largescale,manually annotated data.Self-supervised learning provides an effective solution to this pressing challenge,as this learning paradigm can leverage unlabeled data in ubiquitous computing scenarios.This thesis assumes and proves that learning representations of unlabeled data from the perspective of time-frequency decomposition is a more effective learning method.TFC can separate the general representation of activity signals from irrelevant noise,thereby further learning the time-frequency representation of activity signals.Based on the idea of time-frequency learning,a contrastive loss function is designed,which continuously optimizes the learned time-frequency representation using a sequence cross-prediction strategy.TFC is robust and can adapt well to various backbone encoders and downstream classifiers.Extensive experiments on three public datasets and a self-built dataset show that this method can effectively improve self-supervised HAR tasks.Especially in the absence of labeled data,TFC can significantly improve recognition performance.