基于图和深度分层的前景物体提取研究

发布时间：2018-05-20 14:25

本文选题：计算机视觉 + 前景提取　；参考：《山东大学》2017年硕士论文

【摘要】：随着计算机视觉技术的快速发展,其成果如增强现实、虚拟现实、智能监控以及特征识别等被越来越多的用于实际生活中,为人们的生活带来了极大的乐趣与便捷。对于许多计算机视觉应用而言,场景中提取出的前景物体是应用进行处理的基础。因此,对图像中前景物体的分割提取越来越受到数字图像处理研究者们的重视。尽管目前研究者们在前景物体分割方面提出了大量巧妙的算法,但是仍然存在不少问题,如提取的前景物体轮廓不准确,无法给出独立个体的分割结果等。随着Kinect深度相机等廉价深度信息获取设备的出现,结合色彩和深度信息为提取前景物体提供了一条新的途径。针对以上存在的问题,本文提出了一种基于图和深度分层的室内场景前景物体提取算法。论文首先对图像中前景物体提取的研究背景、意义、国内外研究现状以及提取前景物体所面临的难点进行了综述,并介绍了论文的结构安排。然后对本文中提出的前景物体提取算法中涉及的关键技术进行了详细的阐述。最后对实验结果进行分析,并进一步用本文中算法得到的提取结果与其他先进算法进行对比评估算法的性能。本文通过将色彩信息与深度信息进行融合以改进基于图的图像分割算法,对给予的场景进行过分割用于后续的合并步骤。然后提出了一种深度图修复算法用以修复给定深度图中的空洞,基于室内场景中的前景物体在深度上起伏较小的假设,利用扩展的多阈值大津法将对应的深度图进行分层处理,断开各前景物体之间以及前景和背景之间在深度上的连续性。并提出了一种利用种子点自动选取合适深度分层的方案以减少交互操作。接着为了解决当场景中前景和背景的色彩和深度均相似时无法在两者之间生成有效边界的难点问题,同时也为了优化提取结果,本文利用深度信息生成法向图以提供一种新的约束条件。最后,利用深度分层、法向图、用户设置的种子点以及区域面积设计约束规则对过分割的场景图像进行区域合并,从而在色彩图和深度图中提取出具有清晰轮廓的前景物体。论文作者对算法的实验结果进行了统计分析并给出分析结果,同时进一步将实验结果与多种已发表的不同算法的实验结果进行对比。实验证明,本文算法更加稳定,所提取的前景物体更加完整,轮廓更加简洁清晰且用户交互更少。
[Abstract]:With the rapid development of computer vision technology, its achievements, such as augmented reality, virtual reality, intelligent monitoring and feature recognition, are more and more used in real life, which brings great pleasure and convenience to people's life. For many computer vision applications, foreground objects extracted from the scene are the basis of processing. Therefore, the segmentation and extraction of foreground objects in images are paid more and more attention by digital image processing researchers. Although researchers have put forward a large number of clever algorithms in foreground object segmentation, there are still many problems, such as the inaccuracy of the extracted foreground object contour, the inability to give the segmentation results of independent individuals and so on. With the appearance of cheap depth information acquisition equipment such as Kinect depth camera, combining color and depth information provides a new way to extract foreground objects. In view of the above problems, this paper proposes a foreground object extraction algorithm for indoor scene based on graph and depth stratification. Firstly, the background and significance of foreground object extraction in image, the research status at home and abroad and the difficulties in extracting foreground object are summarized, and the structure of the paper is introduced. Then, the key technology of foreground object extraction algorithm proposed in this paper is described in detail. Finally, the experimental results are analyzed and compared with other advanced algorithms to evaluate the performance of the algorithm. In this paper, the color information and depth information are fused to improve the graph-based image segmentation algorithm, and the given scene is over-segmented for subsequent merging steps. Then a depth map repair algorithm is proposed to repair the holes in the given depth map. Based on the assumption that the foreground object in the indoor scene has a small depth fluctuation, the extended multi-threshold method is used to delaminate the corresponding depth map. Disconnect the continuity of depth between foreground objects and between foreground and background. A scheme of selecting suitable depth layer automatically by using seed points is proposed to reduce the interactive operation. Then, in order to solve the problem that the color and depth of background and foreground in the scene are similar, which can not generate the effective boundary between them, and to optimize the extraction results, In this paper, we use depth information to generate normal graphs to provide a new constraint condition. Finally, the over-segmented scene images are merged by depth stratification, normal graph, user-set seed points and area design constraint rules, and the foreground objects with clear contours are extracted from the color map and depth map. In this paper, the experimental results of the algorithm are statistically analyzed and analyzed, and the experimental results are compared with the experimental results of various published algorithms. Experimental results show that the proposed algorithm is more stable, the extracted foreground objects are more complete, the contour is more concise and clear, and the user interaction is less.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】