基于结构化低秩表示的运动目标分割研究

Effective motion object segmentation is the premise and basis of intelligent video analysis. Owning to the complexity of background, the dynamic variations and the non-rigid deformation of motion objects, how to effectively segment the video object is a difficult problem. This proposal aims to construct a new structured low rank representation model for video data, which is then applied to the video motion object segmentation problem. The model decomposes the background and foreground as the low rank and sparse part of the video matrix respectively. The low rank prior of background and the temporal-spatial continuity prior of foreground motion objects is fully used to segment the motion objects robustly. We first study the discriminative dictionary learning method for low rank representation of scene background and 3D graph structured sparse model for foreground motion objects. Then video object segmentation model is constructed based on the structured low rank and sparse matrix decomposition of video matrix. Optimization algorithm is also proposed to solve the model fast. Lastly, increment and recursive processing mechanism is also designed. This proposal adapts to the dynamic change of the background by learning the background low rank representation dictionary adaptively. 3D graph structured sparsity model is used to utilize temporal and spatial context information adequately, which is beneficial to locate the motion object accurately. In addition, incremental segmentation model enables it to be suitable for the online segmentation and the long-time video processing. The proposal not only has important theoretical meaning to develop the existing matrix low rank decomposition theory, but also has wide application prospect, such as video analysis and video object coding, et al.

有效的运动目标分割是实现视频智能分析的一个重要前提。由于视频场景的复杂性以及背景的动态变化、目标的非刚体形变等因素影响，如何实现高效的目标分割是一个难点问题。本项目旨在于建立视频数据的结构化低秩表示模型，并将其应用于运动目标分割问题，将背景和前景分离为视频数据的低秩与稀疏部分，充分挖掘背景的低秩先验与前景运动目标的时空连续性先验，实现运动目标的鲁棒分割。项目首先研究背景的鉴别性低秩表示字典学习与前景目标的3D图结构化稀疏建模方法，进而构建基于图结构化低秩稀疏分解的分割模型及其快速算法，最后建立该模型的增量递推式处理机制。本项目通过在线学习背景字典使其自适应于复杂背景的动态变化，应用3D图融合时空上下文信息准确定位运动目标，增量分割模型适用于在线处理以及长时段视频的处理，不仅对拓展现有的矩阵低秩分解方法具有重要的理论意义，且在视频分析、视频对象编码等领域具有广泛的应用前景。

由于视频场景的复杂性以及背景的动态变化、目标的非刚体形变等因素影响，如何实现高效的目标分割是一个难点问题。本项目建立了一套基于结构化低秩稀疏分解的运动对象分割模型与求解方法, 将背景和前景分离为视频数据的低秩与稀疏部分，充分挖掘背景的低秩先验与前景运动目标的时空连续性先验，实现运动目标的鲁棒分割。针对稀疏低秩字典学习、运动对象的时空结构化稀疏性度量、运动目标分割模型、分割模型的优化求解算法等关键问题进行了深入研究，取得了多项成果，录用发表期刊论文14篇，包括IEEE Trans. on Image Processing、IEEE　Trans. on Geoscience and Remote Sensing、IEEE Journal Of Selected Topics In Applied Earth Observations And Remote Sensing、International Journal of Remote Sensing等国际知名期刊，以及电子学报、计算机辅助设计与图形学学报等国内一级学会期刊，会议论文3篇，申请发明专利8项，授权发明专利2项，授权实用新型专利3项,获得软件著作权2项。圆满完成了项目计划书规定的研究任务，发表(录用)的论文数和专利数均满足验收指标。项目研究成果不仅对拓展现有的矩阵低秩分解方法具有重要的理论意义，且在智能视频监控、智能交通以及视频检索、人机交互等领域具有广泛的应用前景。