Deep learning architectures for human activity recognition
Phivos Mylonas (supervisor)
Katia Lida Kermanidis
This PhD thesis focuses on human activity recognition (HAR). It combines the research areas of computer vision, signal processing and machine learning. It aims at automatically recognizing actions of humans through a series of observations either directly on them or within the area in which they operate. For activity recognition, new methodologies are studied in order to implement novel architectures based on deep learning techniques. The latter will be trained using measurements of physiological and kinematic parameters, as well as extracted audiovisual data recorded by sensors either placed on the users’ environment. Emphasis is given on visual representations of recorded data, so that recognition will be based on convolutional neural network architectures. Information fusion techniques using heterogeneous sensors are also investigated.