水下敏捷机器人抓捕性能智能评测模型与融合平台

The mechanization, automation and intelligentization of underwater fishing are critical enablers for the fast and sustainable development of marine ranching. Scientists in China and abroad have started laying the groundwork for the foundational theories and key technologies for underwater agile robots. However, no one has explored the terrain of measuring fishing performance in shallow water in a scientific and comprehensive fashion. This project focuses on the challenging problems in underwater fishing, such as limited visibility in an underwater environment and waves and ocean currents on the surface level. Our goal is to build an evaluation model and platform that enables the visualization of different scenarios, evaluation of results and sharing of data. Specifically, we aim at eliminating noises caused by bad visibility and ocean currents by applying the idea of data driven, to accurately recreate and demonstrate competition scenarios that combine virtual and real-time information. We plan to work on the elimination of noise caused by surface waves by generating signal transmission models upon statistical analysis, to enable reliable field data transmission by using remote wireless network, and to build a data-driven, coordinated fishing-performance evaluation model by using machine learning, human evaluation and the objective measure of historical data. And lastly, we plan to establish a learning platform that allows sharing of evaluation data and feedback by standardizing data environment and task data interface and marking. The completion of the project will fill in the blank of performance evaluation for fishing missions for underwater agile robots, and will form a complete research loop for “data sample to artificial intelligence technology to evaluation” upon environmental and task data. The new method adapted to underwater environment and the data-sharing platform will also lay important groundwork for future theoretical investigation, technical breakthrough and industry development.

国内外对水下生物抓捕敏捷机器人的基础理论与关键技术研究方兴未艾。然而，缺乏综合评测近浅海特定环境下抓捕性能的方法与系统。项目聚焦水下环境能见度差、海面波浪、水流等带来的挑战问题，研究达到场景可呈现、结果可评价与数据可共享要求的评测模型与平台。运用数据驱动思想克服能见度差及水流带来误差，精确重建与实时展现虚实结合的抓捕场景；利用数据统计分析建立信号传播模型抵消海面多波浪影响，实现现场数据远程无线网络可靠传输；借助机器学习结合人工评测、客观指标与历史数据，构建数据驱动的抓捕性能智能评测模型；规范化数据接口与标注，形成评测数据共享与反馈学习的融合平台。项目研究有望填补水下敏捷机器人抓捕任务评测的空白，并以数据为核心，形成“样例数据<->智能方法<->评测知识”的研究闭环，所获得适应水下环境的智能评测新方法与数据共享平台，也将为水下敏捷机器人的理论研究、技术突破和产业拓展提供有力的支撑。

本项目聚焦水下敏捷机器人抓捕性能智能评测模型与融合平台研究，针对能见度差、水流湍急和环境光干扰等水下恶劣环境存在的挑战，提出一系列视觉增强、目标检测、场景虚/实重建、可靠信息传输和水下抓捕智能评测等方法，研发了场景可呈现、结果可评价与数据可共享的评测模型与数据融合平台。在水下抓捕虚/实重建方面，提出了基于自动学习的水下环境适应的视觉增强方法、基于通道锐化的鲁棒快速多目标检测方法和基于局部和全局几何结构约束的全景图拼接方法，精准呈现水下场景。在海洋环境信息传输方面，首次将圆偏振光引入到水下光通信系统设计中，实现复杂水下环境鲁棒信息传输。在水下抓捕性能评测方面，构建了真实海洋环境下多尺度融合的水下智能机器人测评方法和基于在线持续学习的智能评测模型，客观评测抓取性能。在评测数据生成平台方面，构建了真实水下环境抓捕多任务标准数据集，建立场景与任务共享平台，生成并共享评测数据。发表国际权威期刊/会议论文40余篇，其中CCF/CAA 推荐A类期刊论文18篇（如TPAMI、TIP等），CCF推荐A类会议论文13篇（如CVPR、IJCAI等）。已授权发明专利16项，其中5项关于恶劣环境视觉感知的专利群完成100万的成果转化。本项目所研发的水下场景呈现与任务共享平台自2020年起应用于国家自然科学基金委主办的全国水下机器人大赛，为央视等多家媒体的赛事报道提供了重要技术支持。利用该系统采集构建了大规模水下视觉数据集，已作为鹏城国家实验室云脑数据面向全球研究者共享，国内外40多个研究团队已利用该数据集研发和测试感知新算法。项目组自2018年连续五年承办全国水下机器人大赛，该赛事为国际上首个在真实海域进行的机器人竞赛，被央视新闻、央视国际频道累计新闻报道10余次。本项目的成果为水下敏捷机器人的理论研究、技术突破和产业拓展提供了有力的支撑，助力国家智慧海洋、经略海洋相关领域发展。