Map
“Pattern Recognition in Video”
Rama Chellappa
Minta Martin Professor of Engineering
Department of Electrical and Computer Engineering and UMIACS
University of Maryland, College Park, MD.
With the ubiquitous presence of inexpensive video cameras, new challenges to video-based pattern recognition problems are emerging. Video-based pattern recognition problems have applications in homeland security, healthcare, battlefield awareness, and video indexing and anomaly detection. The single most important feature that distinguishes video-based pattern recognition problems from still-image based recognition problems is the dynamical nature of patterns in videos. This creates new intellectual challenges and provides opportunities for novel approaches. In this talk, I will first discuss some of the general principles for designing robust video-based pattern recognition systems using statistical, structural and syntactic methods. We first present the design of statistical parametric classifiers for face and gait-based human recognition, recognition of bee dances, human activity recognition and vehicle class recognition across non-overlapping cameras. Characterization of class-conditional densities using pattern appearance, shape, motion and behavior are discussed. A method for compensating for the variations in the rate at which patterns evolve and the role of quasi-invariants in activity recognition are then discussed. A non-parametric method based on a novel "human gait DNA" signature is then described for recognizing human motion patterns. We then present two examples of pattern recognition in video using attribute grammars and stochastic Petri nets. Finally, we discuss some theoretical issues and practical problems that remain to be addressed in this area.
Prof. Chellappa received the M.S.E.E. and Ph.D. Degrees in Electrical Engineering from Purdue University, West Lafayette, IN, in 1978 and 1981 respectively. Since 1991, he has
