动态场景深度场高效感知与计算

Depth field is the core of stereo vision, and it has great potential in the application of military, aerospace, medical, machinery, digital content generation, and other fields. Sensing and computation of dynamic scene depth field acts an important role in opening a new paradigm from static to dynamic for the development of above stereo vision based applications. The research of sensing and computation of dynamic scene depth field is a multidiscipline topic, including computational photography, image sensing, optimization, etc. It is the frontier and highlight topic in international academia, facing the challenges from low accuracy sensing, low spatial quality, and difficulties in temporal reconstruction. Crucial research topics of it include coding methodology of light, imaging model of object surface and illumination, temporal-spatial variation rules of depth field data, etc. In this project, we dedicate to reveal the intrinsic relationship between the sensing and computation of dynamic scene depth field. Our works include (a) realize accurate sensing for dynamic scene depth field, in which temporal-spatial multiplexing structured light coding method, temporal-spatial joint multi-resolution unique structured light pattern and computational model for decoding will be involved; (b) spatial quality optimization method with surface-illumination imaging model, where object surface and geometric features, and optical features of lens will be involved in modeling; (c) realize dense temporal robust reconstruction for dynamic scene depth field, in which the intrinsic relationship between object motion and the temporal-spatial variation rule of depth field data will be uncovered. We attempt to have breakthroughs in the research of both theory and core techniques in this field, and realize an effective sensing and computational platform for dynamic scene depth field. We can contribute the promotion of stereo vision based applications with the core of depth field.

以深度场为核心的三维视觉在军事、航天、医学、制造、数字内容等领域有着广泛的应用前景，动态场景深度场的感知与计算是其实现从静态场景向动态场景跨越式发展的关键，涉及计算摄像、图像感知与优化等多个学科方向的交叉，是当前国际研究的前沿与热点，面临感知精度低、空域质量差、时域重构难的挑战，核心问题包括光信息编码方法、物体表面-光成像模型、深度场数据时-空变化规律等。本项目力图揭示动态场景深度场其感知与计算之间的本质联系，研究时-空联合复用的光信息编码方法，建立时-空域多分辨唯一性结构光模板与解码计算模型，实现动态场景深度场精确感知；研究物体表面与几何特性、光学特性下的表面-光成像模型，构建该模型约束下的空域质量优化方法；研究物体运动特征与深度场数据的时-空变化规律，建立时域密集鲁棒重构方法。在理论与关键技术研究上取得突破，实现动态场景深度场的高效感知与计算平台，推动深度场为核心的三维视觉的广泛应用。

以深度场为核心的三维视觉在军事、航天、医学、制造、数字内容等领域有着广泛的应用前景，动态场景深度场的感知与计算是其实现从静态场景向动态场景跨越式发展的关键，涉及计算摄像、图像感知与优化等多个学科方向的交叉，是当前国际研究的前沿与热点，面临感知精度低、空域质量差、时域重构难的挑战，核心问题包括光信息编码方法、物体表面-光成像模型、深度场数据时-空变化规律等。本项目针对上述挑战，开展了动态场景深度场精确感知、动态场景深度场空域优化、动态场景深度场时域重构等研究，在光照编码的方法、高质量深度图的多视点联合重建方法、空域多视点耦合关系的建立等方面形成突破，研制多维视觉采集装置与系统1套，发表学术论文14篇（包括6篇SCI期刊论文，其中2篇入选ESI高被引论文，8篇EI检索的国际会议论文，其中6篇为CCF推荐的A/B/C类论文）、申请发明专利2项（其中1项已授权）；培养研究生10人（博士1人，硕士9人），项目主持人2018年当选英国工程技术学会会士（IET Fellow）；受邀在国际会议、专题讨论中做学术报告10次，举办国际学术会议1次，出国参加学术会议3次，邀请国际同行前来做学术交流4次。项目成果形成立体视频内容生成与关键技术，在海尔、数码视讯等国内多个企业的产品中得到集成，取得了显著的社会效益，并作为主要完成人之一与海尔共同申报2020年国家科技进步奖。