CAREER: Geometric and Appearance Based Methods for Model Acquisition

该提案的目标是创造一个研究环境，并开发一个致力于研究和开发视觉系统原型的课程。重点将放在通过视觉感知捕捉物体和环境表征的基本问题上。目标包括增强机器人代理的自主性，赋予人类额外的能力，提高人与机器人互动的水平。所提出的自适应表示将在不同的细节层次上捕获环境的几何形状和外观，并将通过适当地利用上下文和任务相关的约束来健壮地获得。将我们的方法与现有方法区分开来的视觉系统的两个基本特性是：(1)它在扩展的动态3D环境中连续运行；(2)它根据环境和手头的任务进行适应的能力。所提出的研究的智力价值将体现在：(1)推进相机运动和场景结构在线估计方法的最新进展；(2)从连续视频流中对物体和环境的基于外观的表征进行独特的学习；(3)在视觉引导任务的背景下整合获得的模型。广泛的影响包括促进对生物系统的理解，为构建人工系统提供重要指导，并与新的技术趋势一起，将产生新的应用。开发的功能将使更复杂的机器人任务能够在较不结构化的环境中完成，并推进人机交互——这些功能对于部署（类人）机器人代理至关重要。它们还可以增强人类的导航和规划能力，并可能影响我们的生活和衰老方式。

The objective the proposal is to create a research environment and develop a curriculum dedicated to studying and developing prototypes of seeing systems. The emphasis will be on the fundamental issue of capturing representations of objects and environments by means of visual sensing. Goals include enhancing autonomy of robotic agents, empowering humans with additional capabilities and advancing the level of human and robot interaction. The proposed adaptive representations will capture both the geometry and the appearance of the environment at various levels of detail and will be acquired robustly by properly exploiting the context and task related constraints. The two essential properties of this visual system that distinguish our approach from previously existing methods is (1) its continuous operation in extended dynamic 3D environments and (2) its capability to adapt depending on the environment and task at hand. The intellectual merit of the proposed research will be in (1) advancing the state of the art in on-line methods for estimation of camera motion and scene structure; (2) unique learning of appearance based representations of objects and environments from a continuous video stream and (3) in the integration of the acquired models in the context of visually guided tasks. The broad impact includes advancing the understanding of biological systems, providing important guidance in building artificial systems, and together with new technological trends, will give rise to novel applications. The developed capabilities will enable more complex robotic tasks in less structured environments and advance human robot interaction -- features that are critical for deployment of (humanoid) robotic agents in service. They can also augment human navigation and planning capabilities and potentially affect the way we live and age.