CAREER: Real-Time 3D Reconstruction and Manipulation for Underwater Intervention - A Career Development Plan

像auv（自主水下航行器）和rov（遥控航行器）这样的水下机器人在深水科学、资源开采、海底建设和维护等许多任务中都有与环境互动的机会。这些任务的关键是解释操作环境的能力，以确保安全性和有效性。该项目提出了一种算法的发展，可以从水下相机获得的图片和视频中快速构建环境的3D模型。在短期内，这意味着增强人类控制auv和rov的方式。从长远来看，PI的职业规划旨在生产能够在没有人类直接控制的情况下安全地与环境互动的系统。除了技术目标之外，本提案还提出了教育和推广计划，以鼓励、培养和促进工程科学和海洋探索的广泛受众。实时水下三维重建是许多其他研究领域的使能技术，包括操作、导航、避障、智能采样和自适应测量。该职业发展规划通过分析水下领域所施加的约束来解决一般视觉问题。PI提出了一种在线3D重建的原则方法，该方法处理光在水中的传播和包含先前形状知识的公式。此外，场景理解框架允许水下光学数据的分割和分类基于三维结构的联合模型和使用最先进的降维方法的场景的光度特性。

Underwater robotic vehicles like AUVs (Autonomous Underwater Vehicles) and ROVs (Remotely Operated Vehicles) have occasion to interact with the environment in many tasks including deep-water science, resource extraction, and sub-sea construction, and maintenance. Crucial to these tasks is the ability to interpret the operating environment to ensure safety and effectiveness. This project proposes the development of algorithms for quickly constructing 3D models of the environment from pictures and video obtained with an underwater camera. In the short term, this means augmenting how humans control AUVs and ROVs. In the long term, the PI's career plan aims to produce systems that can safely interact with the environment without direct human control. Beyond the technical goals, this proposal lays out a plan for education and outreach to encourage, foster, and promote both engineering science and marine exploration to a broad audience.Real-time underwater 3D reconstruction is an enabling technology for many other research areas, including manipulation, navigation, obstacle avoidance, intelligent sampling, and adaptive surveying. This career development plan approaches the general vision problem by analyzing the constraints imposed by the underwater domain. The PI proposes a principled approach for online 3D reconstruction that handles both the propagation of light in water and a formulation that includes prior shape knowledge. Additionally, scene understanding is framed to allow for the segmentation and classification of underwater optical data based upon a joint model of the 3D structure and the photometric properties of the scene using a state-of-the-art dimensionality reduction approach.