Human action recognition is one of the challenging issues under Computer Vision and most of the effort on understanding the human actions involves video analysis.The majority of the works done so far focuses on action recognition over 2D images taken by a single camera. However, single camera systems have some drawbacks: A single camera view looses the information of body configuration called self-occlusion. Moreover, the presence of articulation in human body makes the problem much harder.Real world actions are in 3D space. In this study, we work on understanding actions from multiple cameras. Using volumetric data can resolve the challenges coming from single camera view. On the other hand, viewpoint variations create fundamental challenges in recognizing actions from multi-camera videos. We propose a view invariant pose descriptor over volumetric data. First, we divide the volume data into 2D slices. Then, we fit circles to gain a more compact representation. Lastly, the proposed descriptor is used to present a view-invariant representation for actions.

DATE: 10 November, 2008, Monday@ 16:40
PLACE: EA 409