The aim of this project is to understand image and video content by exploiting the projective geometry between the two views of a scene captured via moving cameras. To achieve this, we extend the well-studied projective geometry between images to video by regularizing the geometric relation in consecutive time instants. The geometric relation between two videos is a powerful representation which has shown potential to solve the shortcomings of the current methods for understanding video content. An important advantage of analyzing the geometry between the videos is treating the video as an object in spatio-temporal coordinates compared to individual processing of video frames. Practically, this property not only overcomes error accumulation related to individual processing of each frame, but also generalizes the standard view geometry between stationary cameras to moving cameras.

We have tested the developed techniques on various challenging problems including recognition of human actions, tracking objects across cameras, and recovering geo-spatial trajectories of mobile objects from video.