| Title | Automated Human Activity Recognition from Video Clips |
|
Prof. K. K. Biswas Professor, Computer Science Engineering Department IIT, Delhi, India Email: kkb@cse.iitd.ernet.in |
Prof K. K. Biswas did his Btech in Electrical Engineering from IIT Madras, followed by Mtech in Control systems and Phd in signal estimation from IIT Delhi. After a brief stint at University of Roorkee, he joined the EE deptt of IIT Delhi. He later shifted to Computer science engineering department where he is currently serving as a professor. His teaching career spans over 35 years. He has been a visiting professor at the University of Auckland, New Zealand and at the University of Central Florida, USA. He has also acted as UNESCO expert for development of curriculum at university of Nigeria.
He has been collaborating with University of Oxford and University of Texas at Austin. He has been an active researcher with 15 Phd students, and more than 60 publications in reputed journals and international conferences. His current area of research interest is image, video processing, machine learning with applications in activity recognition and salient object detection. His other main research interest is handling fuzzy models in probabilistic domain. He is also working in the area of logic based knowledge representation in scientific domains.
Abstract
Human action recognition for automatic understanding of video clips is becoming increasingly important. With the growing need of surveillance related applications, the research in the field of action recognition has been fueled in past few years. Action recognition systems can be used for surveillance purpose and can trigger alarm whenever some suspicious activity takes place. Action recognition systems also help in creating interactive environment which can respond to the actions performed by human actors. Assisted care applications to the elderly can also make use of action recognition techniques. Other possible applications are video summarization, and content based video retrieval, etc.
This talk will be in two parts. The first part will deal with action recognition from RGB video clips. It will be shown how shape based features and optical flow based motion features are extracted from the video frames. A short introduction to machine learning will be given to justify use of Support Vector Machine based approach for training the system. Results based on lab activities such as walking, sitting down, writing on board, opening a door, sliding on chair will be presented.
The second part of the talk will deal with small scale actions performed while sitting down on a chair, such as typing, reading, attending phone, engaged in discussion, drinking tea, texting, stretching and dozing. The objective will be to show how a depth based camera can be effectively used for identifying the body parts. The features used are depth features, spatio-depth features, temporal depth features, motion history image. RGB data is used to identify face and hands to support the depth data. The system is trained using a Support Vector Machine. Testing is done on data set collected at IIT delhi.
