 |
Center for Nonlinear Studies |
There are important security implications in being able to detect and describe human interactions in video streams. This relies on detecting humans in the first place, and so we start with a brief description of the current state of human tracking technology. We then present a model-based approach for finding intervals in which two actors are interacting, using their gaze direction and relative movement. Combining ideas from survival analysis and sequential Monte Carlo, we propose a framework to generate entire `stories', where each story is a sequence of cluster configurations that describes the actors' group memberships throughout the movie. Naturally, the ultimate objective is not just to cluster but to describe the social activities that are taking place. In a very recent extension of our work, we incorporate graphical models in the inferential engine which, using additional evidence, seek to determine the type of interaction that is occurring. We illustrate our results on scripted video sequences provided by DARPA.
Host: Joshua Neil, 665-2121, jneil@lanl.gov