Unit:
Department of Informatics and TelecommunicationsΠληροφορική
Author:
IYAMU-PERISANIDIS SIMON
Supervisors info:
Ευάγγελος Σπύρου, Επίκουρος Καθηγητής, Πανεπιστήμιο Θεσσαλίας, Συνεργαζόμενος ερευνητής Εθνικό Κέντρο Έρευνας Φυσικών Επιστημών-«Δημόκριτος»
Παναγιώτης Σταματόπουλος, Επίκουρος Καθηγητής, τμήμα Πληροφορικής και Τηλεπικοινωνιών, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Original Title:
Human Action Recognition using Joint Trajectory Maps and Convolutional Neural Networks
Translated title:
Human Action Recognition using Joint Trajectory Maps and Convolutional Neural Networks
Summary:
Human action recognition is an important branch of human-centered research, which focuses on the automatic identification of human behaviours from images or video sequences. Motivated by the promising performance of Convolutional Neural Networks, and their applicability to the human action recognition task, in this thesis, we provide an effective method which is able to recognize human actions in a multitude of demanding scenarios.
Our approach is based on Joint Trajectory Maps and multi-stream Convolutional Neural Networks. The spatio-temporal information of the 3-D skeleton sequences are encoded into JTMs and fed into a deep three-stream ConvNet. Our architecture is trained and evaluated on the challenging PKU-MMD dataset and achieves competent results.
Main subject category:
Technology - Computer science
Keywords:
Convolutional Neural Network, Recognition, Action, Skeleton, Joint Trajectory Maps