基于单目摄像头的实时三维重建

发布时间：2018-10-05 19:54

【摘要】：三维重建是计算机视觉的重要研究方向,可以广泛应用于三维导航、在线购物、3D打印、计算摄影、视频游戏以及文化遗产保护。随着计算机的计算能力的不断增强和人们对于实时交互体验的追求,实时三维重建正在成为新的研究热点。基于单目摄像头的实时三维重建不仅需要对相机的姿态进行实时的估计,而且还要对图像的深度图进行实时估计和融合。为此,我们首先要准确地估计出相机的姿态,然后根据局部信息快速生成准确的深度图,最后采用适合实时交互的高效融合方法来进行融合处理。为了实现上述目标,本文研发了一个基于单目摄像头的实时三维重建系统。该系统可以直接从单目摄像头采集的实时视频数据实时地生成场景的三维模型。系统流程主要包括三个部分:首先是利用同时定位与地图构建(SLAM)系统,对视频序列进行相机姿态的估计,并抽取关键帧进行关键帧的姿态优化,然后以关键帧作为参考帧进行快速的深度恢复。本系统的深度恢复采用了平面扫描算法和基于概率的深度推断方法,并针对深度恢复的结果进行局部优化,最后将关键帧的深度信息进行融合和场景三维重建。为了满足实时的应用要求,我们对深度恢复进行了 GPU并行加速,并使用时空一致性约束和连通分量分析等多种方法来对恢复的深度进行后续优化处理,剔除深度恢复质量不好的区域,从而保证生成的三维模型的质量。该系统提供了良好的用户界面和交互操作,可以实时可视化当前重建的三维点云或网格表面,以指导用户完成对目标物体或场景的完整三维扫描,并可以将重建的三维模型保存到本地。实际拍摄的三维重建效果证明了该系统的有效性,能够满足实际应用需求。
[Abstract]:3D reconstruction is an important research field in computer vision. It can be widely used in 3D navigation, online shopping, 3D printing, computer photography, video games and cultural heritage protection. With the increasing computing power of computer and the pursuit of real-time interactive experience, real-time 3D reconstruction is becoming a new research hotspot. The real-time 3D reconstruction based on a single camera requires not only the real-time estimation of the camera's attitude, but also the real-time estimation and fusion of the depth map of the image. Therefore, we first estimate the camera attitude accurately, then quickly generate accurate depth map according to local information. Finally, we adopt an efficient fusion method suitable for real-time interaction to perform fusion processing. In order to achieve the above goal, a real-time 3D reconstruction system based on single-camera is developed in this paper. The system can generate 3D model of scene directly from real-time video data collected by monocular camera. The system flow mainly includes three parts: firstly, the (SLAM) system is constructed by simultaneous location and map, the camera attitude is estimated for video sequence, and the key frame is extracted to optimize the attitude of key frame. Then the key frame is used as the reference frame for fast depth recovery. The depth recovery of the system adopts plane scanning algorithm and probability-based depth inference method. The depth information of key frames is fused and scene 3D reconstruction is carried out based on the local optimization of depth recovery results. In order to meet the requirements of real-time application, we accelerate the depth recovery with GPU parallel, and use spatio-temporal consistency constraints and connected component analysis to process the depth of restoration. The quality of the generated 3D model is ensured by eliminating the areas with poor quality of depth recovery. The system provides a good user interface and interactive operation. It can visualize the current reconstructed 3D point cloud or mesh surface in real time to guide the user to complete the complete 3D scanning of the target object or scene. The reconstructed 3D model can be saved locally. The results of 3D reconstruction show that the system is effective and can meet the needs of practical applications.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】