目标实体轮廓跟踪中动态高阶能量最小化问题的研究

Silhouette tracking is usually formulated as an energy minimization problem. Higher-order graph-cut is one of the most effective methods for minimizing the higher-order energy functions. However, in existing higher-order graph-cut methods there are certain limitations as follows: 1) they tend to be suited to specific forms of higher order energy functions and cannot charaterize complex visual information of objects, and 2) they are based on a single frame image processing strategy and are thus computationally inefficient. To address these shortcomings, we investigate how to represent a generalized higher order energy function using a uniform high-order s/t graph model, overcoming the limitation of the specific higher-order terms in the energy function and enriching the ability of graph-cut methods for characterizing complex targets and backgrounds. Furthermore, according to the coherence of targets and backgrounds in the video frame sequence, we develop a parameter transfer method for incrementally characterizing the changes between higher-order graphs for consecutive frames. Our method dynamically implements higher-order graph-cut algorithms and enables an efficient computation for high-order graph-cut in silhouette tracking. In summary, we will develop a novel framework based on dynamic higher-order graph-cut algorithm. Our work generalizes the theory of graph-cuts and provides an effective and efficient approch to the modelling and computation problems within the generalized higher-order enery minimization for silhouette tracking.

视频中的目标实体轮廓跟踪通常抽象为能量最小化问题。高维图割算法是计算高阶能量函数最小化的最有效方法之一。然而，现存高维图割方法存在以下局限：1）特定形式的高阶项无法刻画目标复杂视觉信息；2）独立帧处理方式导致目标跟踪运算效率低。针对上述缺陷，本课题研究广义高阶能量函数的规整化高维s/t图模型表达方法，克服特定高阶项的局限性，丰富高维图割刻画复杂目标和背景的能力。同时，根据视频中目标和背景的连贯性，研究连续帧中高阶能量的增量动态变化规律，设计高阶项参数转移方法，实现高维图割的动态运算，提升目标实体轮廓跟踪的计算效率。综上，本课题将设计新的动态高维图割框架，有机地拓宽图割类方法的理论外延，有效地解决目标实体轮廓跟踪中高阶能量的广义建模和高效运算的问题。

本项目就视频处理中目标实体轮廓跟踪的动态高阶能量最小化问题进行了一系列的研究。首先，研究了二维动态图割算法，提出动态图割算法运用到目标跟踪中的新方法，克服现有图割方法的不足，提高了目标跟踪效果和运算效率。进而研究一种基于增量式高次布尔能量最小化的视频前后景分割方法，既能利用高次布尔能量函数表示复杂的视觉信息，又能减少高次布尔能量最小化过程中的重复运算，从而达到比二次布尔能量函数更精确和比独立帧高次布尔能量最小化更高效的效果。后续在研究动态高维图割算法的基础上，提出了基于图像数据的对偶平滑海上溢油检测新方法，克服了传统图像分割方法无法完整检测长条状溢油形状的局限。针对能量最小化问题，提出新的医学3D图像针对刚性和局部配准新方法，提高3D图像的配准精度。将研究的算法应用到实际工程系统中，设计了无线智能双车联动系统。设计出一种车牌清晰显示方法，解决了在实际交通环境中，因抓拍的车辆成像太小而导致车牌不能识别或者识别精度较低的问题。在项目执行期间，美国佛罗里达大学访问学者一年，共发表论文10篇（其中SCI 4篇），授权中国发明专利4项，实用新型2项，为后续进一步的研究奠定了坚实的基础。