University of Technology, Sydney

Staff directory | Webmail | Maps | Newsroom | What's on

Seminar: Human Activity Recognition with depth cameras - Dr Zicheng Liu, Microsoft Research, Redmond, USA

Presenter: Dr Zicheng Liu, Microsoft Research, Redmond, WA USA (invited by Prof Dacheng Tao)

Abstract: Understaning human activity has been a long-standing problem in computer vision. It has many potential applications, such as surveillance, health monitoring, entertainment, and human computer interaction. Recently, the availability of commodity depth cameras such as Kinect has introduced a new level of excitement to this field. Rapid progress has been made addressing new technical issues in activity recognition with depth cameras. This talk will describe some of our recent work in this direction (one appeared in CVPR’12, and the other is accepted by ECCV’12).

Part1: A common choice when using Kinect for activity recognition is to use skeletons, which has two limitations. First, the skeletons are sometimes missing, or very noisy, seriously degrading the recognition performance. Second, many human activities involve human object interactions, and the skeleton does not contain any information on the object that the person is interacting with. To address these issues, we have developed a skeleton-based feature by using Fourier Temporal Pyramid, and a local shape feature called Local Occupancy Pattern. A data mining method is used to learn discriminative sub-sets of joints, called action-let ensembles, to further improve the recognition performance. This method is evaluated on a daily activity dataset that consists of activities involving human-object interactions. 

Part2: We propose a type of semi-local feature, called a random occupancy pattern (ROP), for activity recognition from depth maps. We employ a sampling scheme that effectively explores an extremely large sampling space, and use a sparse coding approach to robustly encode these features. It does not rely on skeleton tracking, and it is quite robust for occlusions. We have applied it in action recognition, hand gesture recognition, and object recognition.

Brief Bio: Dr Zicheng Liu is a Senior Researcher at Microsoft Research, Redmond, USA. His current research interests include human activity recognition, face modeling and animation, and multimedia collaboration. He received a PhD in Computer Science from Princeton University. He has published over 80 papers in peer-reviewed international journals and conferences, and holds over 50 granted patents. He co-authored the book, “Face Geometry and Appearance Modeling: Concepts and Applications”, Cambridge University Press, 2011. He has served on technical committees for many international conferences. He is a technical co-chair of both 2010 and 2014 ICME, a co-organizer of 2011 and 2012 CVPR Workshops on Human Activity Understanding from 3D Data, and a general co-chair of 2012 IEEE Visual Communication and Image Processing. He is an associate editor of Machine Vision and Applications Journal, and an associate editor of the Journal of Visual Communications and Image Representation.



6 July 2012
10:30 - 12:00
City - Broadway CB02 Level 4, room 20
All Welcome
Barbara Munday

Back to list of Past Events