Human action recognition has wide applications in security, entertainment, sports, and medicine. It is very challenging to develop robust action recognition models for real-life scenarios, where human movements are associated with numerous irrelevant variations such as viewpoints, backgrounds, and illuminations. This thesis proposes novel data-driven methods to address the fundamental problems in robust action recognition, including human data synthesis, spatio-temporal representation, action classification, and human pose recovery. Besides, the adversarial robustness of human action recognition is investigated in this thesis.
|Qualification||Doctor of Philosophy|
|Award date||26 Jan 2020|
|Publication status||Unpublished - 2020|