اميرحسين اسمعيل زاده

In this research, an advanced framework for human activity recognition based on inertial sensor data was proposed, designed through the integration of self-supervised learning methods, temporal convolutional neural networks (TCN), transformer architecture, an‎d attention fusion strategies. The proposed model, aimed at extracting effective spatiotemporal features from multichannel data, was first implemented an‎d eva‎luated on the stan‎dard UCI-HAR dataset, achieving an impressive accuracy of 98.23%, outperforming many previous benchmark models. Subsequently, the modelʹs performance was tested on the more diverse PAMAP2 dataset, which includes a wider range of activities, an‎d achieved an overall accuracy of 93.93%, demonstrating its high generalization capability under more realistic conditions. Performance eva‎luations based on various metrics, ROC curves, an‎d confusion matrices confirmed the efficiency an‎d balanced nature of the model in distinguishing different activity classes. The findings of this thesis suggest that the simultaneous use of powerful deep learning an‎d self-supervised architectures can lead to the development of intelligent an‎d accurate systems in the field of human activity monitoring an‎d recognition.

تشخيص فعاليت هاي انسان , شبكه هاي عصبي

human activity recognition , Neural networks , UCI-HAR , PAMAP2

Amirhossein esmaeilzade

ahmad Ayatollahi