多视环境下基于变分水平集的场景流估计

项目名称： 多视环境下基于变分水平集的场景流估计

项目编号： No.61502382

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 计算机科学学科

项目作者： 李秀秀

作者单位： 西安理工大学

项目金额： 20万元

中文摘要： 人机交互是产业智能化的关键环节，其核心内容之一是获取分析人与机器的运动及交互行为，而场景流因能够细致地描绘三维场景的运动场，因而研究意义巨大。本课题拟综合多个深度、可见光摄像机的优点，研究多视环境下基于变分水平集的场景流获取框架，以获取连续、逼真的场景流，内容包括：针对深度图像中信息缺失、外点等问题，研究在深度和影像信息驱动下，基于水平集函数的多深度视图融合机制，以获得连续、完整的外形曲面。针对场景流估计中的计算量大且易出错问题，研究局部刚性运动指导下的全局场景流估计模型：通过局部区域属性一致性分割及刚性运动估计，缩小场景流估计的搜索空间，降低计算中的空间、时间复杂度；使用水平集函数演化模拟三维目标外形运动以实现连续的场景流的获取。本课题的成果可作为三维场景中的目标分割、跟踪和基于运动场的行为分析的依据和数据源；在机器人导航、智能监控和智能汽车等方面可显著提高其分析理解场景的能力。

中文关键词： 多视点；场景流；深度图像

英文摘要： Human-computer interaction and three-dimensional technology are one of the three themes of Industry 4.0, and the key aspects of the industry intelligence. A core element of human-computer interaction is to acquire and analyze the motions of human and interactions between human and computers. Since scene flow is able to describe the three-dimensional motion field in detail, it is very significant to study the acquisition of scene flow. In this project, multiple depth cameras and color cameras are combined to estimation the scene flow which should be continuous and vivid. The deficiency of depth information and the existence of outlies can affect the completeness and continuity of the shape of three-dimensional object. Therefore, the fusion mechanism of multiple depth images are studied to obtain a continuous and complete shape of three-dimensional object. Further, a model for estimating dense scene flow will be proposed with the guidance of local rigid motion to reduce the complexity of computation and avoid errors: the segmentation of the three-dimensional object shape and the calculation of rigid motion parameters of segmented regions are studied, which can narrow the search space of scene flow estimation; then the dense and continuous scene flow is estimated by simulating the motion of three-dimensional object shape with the evolution of variational level set. The achievements of this project will be the dependence and data source of segmentation, tracking and behavior analysis in the three-dimensional scene, and can improve the ability to analyze and understand the scene in robot navigation, smart monitoring, and smart car etc.

英文关键词： multi-view;scene flow;depth image