Visual models and applications

点击翻译按钮获取中文摘要

This research program deals with the problem of constructing virtual models of real-world scenes within a computer from information acquired by sensors.  Sometimes referred to as virtualized reality, the goal of this research is to develop novel algorithms that can extract visual information from one or more sensors, determine the 3D shape and texture of objects within the field of view,  and render a facsimile of the real scene from an arbitrary viewpoint.  In recent years my research has been about computing such representations and using them to provide situational awareness to human operators performing difficult tasks under poor viewing conditions, e.g., using textured digital elevation maps to produce synthetic images of the ground for a helicopter pilot flying in fog.  Here the research emphasized how these computer representations could be updated in real-time based on external sensory data (e.g. infrared cameras) to reflect the actual state of the world.   However, taking these ideas forward to a more complex domain such as an urban street scene invokes a far more difficult acquisition task (multiple, distributed sensors), more complex representations (e.g. dynamic, articulated 3D shape models), and the evolution from static rendering to dynamic simulation.  Towards these somewhat ambitious goals, the proposed research extends our current work in active model acquisition to the multi-sensor case and new representations for complex 3D scenes.   In addition to supporting visualization, our work will also investigate the interpretation of evolving 3D scenes, such as the recognition of events taking place.  As part of a related project, this research will be initially applied to the problem of monitoring changes to and events within urban environments.