Automatic human action recognition in videos is a significant research problem and has many applications in surveillance, human-computer interaction and video retrieval. Depth cameras have become popular for this problem because depth videos do not suffer from the uncertain attributes induced by variations in illumination and clothing texture. However, the presence of occlusion, sensor noise and most importantly viewpoint variations still make action recognition a challenging task. This thesis proposes algorithms for efficient modelling of depth and RGB videos with particular emphasis on automatic learning of the complex structures of human actions without making prior assumptions about the camera viewpoint.
|Qualification||Doctor of Philosophy|
|Award date||2 Nov 2016|
|Publication status||Unpublished - 2016|