基于深度线索的多假设变分场景流估计研究

In order to overcome the problem of estimation error caused by motion discontinuity and the stereo matching costs problem in the computation of scene flow, we focus on the research of depth cue based multi- hypothesis variational scene flow estimation. The research includes: 1) 2D optical flow smoothing based multi-hypothesis variational scene flow estimation, which utilizes depth information from binocular stereo sequences to constraint the unknowns and utilizes image and flow driven based scene flow diffusion smoothing to improve the estimation accuracy at motion boundary and robustness of the algorithm; 2) design of new functional based on 3D scene flow driven based anisotropic diffusion smoothing constraint and 3D data term for variational scene flow estimation, which keeps smoothing hypothesis consistent with 3D motion modal; 3) scene flow estimation based on RGB-D data from depth sensor and 3D scene flow driven based anisotropic diffusion smoothing constraint，which can save the computing costs, and at the same time provides the high quality initial value for variational optimizing; 4) scene flow estimation based on depth map anisotropic diffusion smoothing , which is used in the varivational scene flow computation framework to receive the dense scene flow and smoothing depth map. In conclusion, we aim to propose a novel depth cue based multi- hypothesis variational scene flow estimation frmawork, which can integrate various constraint principles flexibly. We believe the proposed methods will lay a sound foundation for the future research.

针对场景流计算在运动不连续处估计误差大及立体视觉深度测量方式计算量大的问题，研究基于深度线索的多假设变分场景流估计技术。研究内容包括：1）基于2维光流估计场景流，利用立体图像序列提供的深度信息约束运动求解，并将图像光流联合驱动各向异性平滑约束引入场景流估计，增强算法在运动边缘与遮挡位置处的估计精度与可靠性；2）基于3维场景流驱动各向异性平滑约束及相应的3维数据项构建新的能量泛函求解场景流，以使平滑约束符合3维运动模型；3）基于深度传感器获得的RGB-D数据及3维场景流驱动各向异性平滑约束估计场景流，在降低计算量的同时为变分优化提供高质量的深度初值；4）将各向异性深度图平滑约束引入变分框架构建深度图平滑项以约束场景流求解，在得到稠密场景流的同时对深度图初值进行光滑求精。本项目旨在构建变分场景流求解新方法，利用变分法灵活的多约束集成特性，为场景流计算探索新的途径，并为其应用奠定理论和技术基础。

为提高变分场景流估计精度，研究了基于深度线索的多假设变分场景流估计技术。研究内容包括：1）基于2维光流估计场景流，利用立体图像序列提供的深度信息约束运动求解，并将光流驱动各向异性平滑约束引入场景流估计，增强算法在运动边缘与遮挡位置处的估计精度与可靠性；2）基于3维场景流驱动各向异性平滑约束及相应的3维数据项构建新的能量泛函求解场景流，以使平滑约束符合3维运动模型；3）基于深度传感器获得的RGB-D数据及3维场景流驱动各向异性平滑约束估计场景流，将深度图驱动各向异性平滑引入变分框架以约束场景流求解，在得到稠密场景流的同时对深度图初值进行光滑求精。本项目构建了变分场景流求解新方法，利用变分法灵活的多约束集成特性，为场景流计算探索了新的途径，并为其应用奠定了理论和技术基础。