Interpretation of visual models

点击翻译按钮获取中文摘要

We define a visual model as a representation of the 3D shape, structure, behaviour, etc., of complex objects or environments, derived from information obtained from sensors that reflect visual information. Like much of the research in computer vision over the years, our work has focused on how such models are computed from sensory data, often with the goal of determining a larger attribute or state. In recent years we have focused on the specific problem of modeling urban environments using information captured by mobile sensors and surveillance systems. The current proposal focuses on several aspects of this research that have become increasingly important in the field, as well as applications of this research that are of interest to our industrial partners. Of particular interest is the problem of situational awareness - the ability of a system to intentionally gather information from multiple sensor streams, fuse it into a consistent representation of the environment, and respond to queries about state in an asynchronous fashion. We will investigate two specific instances - monitoring changes to and activities within a large scale urban environment and ensuring worker health and safety on the factory floor. This will involve performing the research necessary to advance the state of the art in modeling complex environments, building more powerful tools for accessing temporal variations in the data, devising the means to control sensor networks and fuse information, and finally to develop computer representations sufficiently powerful to host these models. As in the past, we will work with industrial partners to provide grounding for the research, experimental validation, and technology transfer. In the long term we hope to develop algorithms that can not only infer models of complex environments, but also simulate them as well. Being able to make such predictions could dramatically alter the way in which we manage and remediate our urban infrastructure.