自然场景的3D深度恢复及应用研究

发布时间：2018-05-25 16:53

本文选题：立体显示 + 深度恢复　；参考：《天津大学》2015年博士论文

【摘要】：随着显示技术的快速发展以及人类日常生活需求的增长,3D立体显示技术已然掀起了图像图形显示领域的一场新技术革命,成为影视及影像行业的最新、最前沿的高新技术,它以新、特、奇的表现手法,真实而强烈地视觉冲出力,良好优美的环境感染力吸引着人们的目光。同时,3D显示技术在各行各业中都得到了实际有效的应用。但是现存的大部分图片及视频仍然是2D的,而且市场上仍然没有非常廉价的适用于个人使用的3D采集设备。怎么恢复这些2D图片或视频的3D深度信息成为3D立体显示领域的一项重要任务。在获得这些2D场景的深度信息之后即可轻易的将2D场景转换成3D场景。本文围绕2D场景的深度恢复及应用,研究了利用散焦线索对单幅图像进行深度图的恢复、基于图像结构的深度图平滑修正、结合深度图的RGB图像显著性目标检测、利用3D深度信息的在线人类动作识别等重要问题。主要创新点包括:1.提出了一种利用图像散焦线索恢复场景3D深度信息的方法,通过图像的局部区域频谱幅度对比度来建立场景深度恢复模型。考虑到恢复出的深度图中可能存在噪音点,本文提出一种基于总变分的图像保边缘平滑算法,用以平滑恢复出的初始深度图。使得最终恢复出的深度图纹理区域更加平滑,因此本文恢复的深度图更加适合于2D到3D转换以及其它方面的应用。2.利用图像的背景先验和颜色的空间分布,首先提出了一种RGB图像的显著性检测。另外,本文将深度图信息融入到矩阵的低秩恢复模型,提出了一种RGBD图像的显著性检测方法。相比于以往的主流算法,我们在不同的RGB图像数据集和RGBD图像数据集上都取得了更好的结果。本文的结果有更高的准确率和召回率,因此,本文检测出的显著图结果能够更好的用于图像的后续处理,如图像分割、基于内容的图像编辑等。3.从实际应用的角度出发,本文提出了一种基于深度信息的在线人类行为识别算法,通过协方差矩阵对每一帧进行特征描述相,利用核化的SVM和最邻近搜索算法实现分类。相比于以往基于片段的行为识别方法,本文提出的方法更具实用性。相比于以往的在线行为识别方法,本文的方法具有更高的准确率、更低的时延。
[Abstract]:With the rapid development of display technology and the growth of human daily life demand, 3D stereoscopic display technology has set off a new technological revolution in the field of image and graphics display, and has become the latest and most advanced technology in the film and video industry. It attracts people's attention with its new, special, strange expression, real and strong visual impact, good beautiful environment appeal. At the same time, 3D display technology has been effectively applied in various industries. But most of the existing images and videos are still 2D, and there is still no very cheap 3D collection device for personal use. How to restore 3D depth information of 2D images or videos becomes an important task in 3D stereoscopic display field. After obtaining the depth information of these 2 D scenes, you can easily convert 2 D scenes into 3 D scenes. Based on the depth restoration and application of 2D scene, this paper studies the restoration of a single image by defocusing clues. The depth map is smoothed based on the image structure, and the salient target detection of RGB image is combined with the depth map. Online recognition of human motion using 3D depth information and other important issues. The main innovations include: 1. In this paper, a method of restoring 3D depth information of scene by defocusing cues is proposed, and the model of scene depth recovery is established by contrast of local region spectrum amplitude. Considering that there may be noise points in the restored depth map, this paper presents an image edge preserving smoothing algorithm based on total variation to smooth the restored initial depth map. So the depth map restored in this paper is more suitable for 2D to 3D conversion and other applications. Based on background priori and color spatial distribution of images, a salience detection method for RGB images is proposed. In addition, the depth map information is incorporated into the low rank recovery model of the matrix, and a significance detection method for RGBD images is proposed. Compared with the previous mainstream algorithms, we have obtained better results on different RGB image datasets and RGBD image datasets. The results of this paper have higher accuracy and recall rate. Therefore, the salient map can be better used in image processing, such as image segmentation, content-based image editing, and so on. From the point of view of practical application, an online human behavior recognition algorithm based on depth information is proposed in this paper. Each frame is characterized by covariance matrix, and the kernel SVM and nearest neighbor search algorithm are used to classify each frame. Compared with the previous segment-based behavior recognition methods, the proposed method is more practical. Compared with the previous online behavior recognition methods, the proposed method has higher accuracy and lower delay.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】