It is estimated that in 2014 there were over 100 million surveillance cameras in the world. Fueled by security concerns, this number continues to steadily grow. As monitoring of video feeds by human operators is not scalable, automatic surveillance tools are needed. In this talk, I will cover a complete video surveillance processing chain, developed over years at Boston University, from low-level video analysis to summarization of dynamic events. I will focus on three fundamental questions posed in video surveillance: “How to detect anomalous events in a visual scene? How to classify those events? How to represent them succinctly?’’ First, I will present “behavior subtraction’’, an extension of “background subtraction’’ to scenes with dynamic backgrounds (e.g., water surface that is notoriously difficult to handle), which can detect complex anomalies in surveillance video. Then, in order to classify activities within the detected anomalies, I will discuss activity recognition on covariance manifolds. Finally, I will describe “video condensation’’, a computational method to succinctly summarize activities of interest for efficient evaluation by human operators.

Speaker(s): Prof. Janusz Konrad,

Location:
Room: 229
Bldg: Wydział Elektroniki i Technik Informacyjnych
Nowowiejska 15/19
Warszawa, Mazowieckie
00-665