面向“智慧工厂”的视觉目标检测和智能决策研究

"Smart Factory" is a transition stage from modern factory to the unmanned one. It means that the factory can fully understand all the information about itself and make intelligent decisions based on global consideration by combining information, network and decision-making technique together to insure the safety of workers, efficiency of manufacture and quality of products in the manufactories. Manufacturing is the pillar industry in our national economy. Moreover, advanced manufacturing is the leading direction of the upgrading for the whole industry, which has been confirmed by the "12th Five-Year Plan" of our country. Hence, the research on "Smart Factory", which stands for the advanced productivity, is very important for promoting the upgrading of industry and the development of economic. The main issues in the research of "Smart Factory" contains, 1)target detection based on vision in the special environment of factory with specifical task, 2)multi information extraction and its transmission in the network, and 3) the global intelligent decision making based on the previous steps. This project will research on these three issues, specially, focus on the effective detection and understanding of the human, environment and the manufactoring procedure in the factory, based on the previous research foundation of the chief applicant. The other two issues will also be researched based on cooperation of applicant and cooperator.

"智慧工厂"是现代工厂向无人工厂发展的过渡阶段，它通过"信息-网络-决策"的融合实现对工厂的全面理解和整体智能决策，从而进一步保障制造业中工人人身安全、生产效率和产品质量。制造业是我国国民经济的支柱，是国家"十二五"规划的重点布局产业。"智慧工厂"的研究对推动制造业转型升级、促进我国国民经济发展有重要的意义，也代表着先进的生产力。 "智慧工厂"研究的主要难点包括：1）工厂特定环境特定任务下的视觉目标检测、2）多种信息的有效提取与网络传输、以及3）因此获得的全面智能决策。 本项目拟从这三个方向开展研究，特别是基于申请团队的前期基础，实现在特定工厂场景中对人、环境和生产过程的有效视觉检测和理解，并为完成后两步工作提供了技术依据。

本项目针对当前“智慧工厂”在制造业企业技术改造、工业信息化提高方面的重要需求，研究了“智慧工厂”环境下基于任务的视觉理解、多种形式信息的融合和基于全局理解的智能决策三方面的研究内容，主要取得了以下研究成果。. （1）研究了高维、冗余的图像、视频数据的目标本质特征提取算法，提出工厂环境约束下工人目标的本质特征提取方法，基于特征结构性建立了目标组件的分割模型，对工人目标行为进行分析。. （2）研究了基于多源数据融合、具有在线学习能力的目标检测算法，提出基于包括微波雷达、RFID、视频等感知信息进行工人目标精确定位及跟踪的方法，建立工人的多源特征感知信息与系统状态之间的映射模型，实现系统状态的自适应切换。. （3）研究了基于全景理解的整体智能决策模型的学习方法等算法模型，基于贝叶斯模型以及图匹配模型分别建立了系统状态趋势的预测方法，实现工厂生产系统的短临决策；. （4）基于上述方法集成开发了基于视觉、微波雷达、RFID等感知模块的多源信息融合与分析智能终端，适用于智慧工业环境中工人位姿检测、设备状态监测、工序进度跟踪与预测等，具有很好的应用意义和市场前景。. 上述研究的算法模型能够在工厂车间环境中有效地对工人等目标进行全方位、多角度的监测，对生产系统状态进行实时跟踪与预测，基于特征和状态感知信息自适应做出系统决策，为提高智慧工厂生产效率提供了一种可行的方法。