手持消费级深度相机稀疏序列融合三维扫描重建

Handheld scanning of the commodity depth cameras breaks through the limited sensing range of the professional 3D scanning devices and the strict constraint of scanning setting. How to generate the high-quality 3D models with a small amount of depth images is a critical problem for handheld scanning with commodity depth cameras. The project which we applied attempts to investigate the handheld 3D scanning method using commodity depth cameras. To compress the captured depth image sequence, our method first evaluates the variation of target shape and measures the indicators of camera motion. Then it constructs the supporting subset for the captured depth image sequence via analyzing the interrelationship of these evaluation indicators online. For each depth image, we extract its geometric features, which could substantially improve the registration of the scanned target. Our method combines the filtering operation and the geometry feature enhancement process to refine the coarse surface of a selected depth image. By fusing the refined depth images in the supporting subset we will propose a new handheld 3D scanning approach using commodity depth cameras based on the sparse-sequence fusion. It will effectively and robustly reconstruct the scanned target with a relatively high-quality. Our research possesses the theoretical significance on the compression and sparse-representation of depth image sequence. It will bring abundant 3D models for computer graphics community while benefit the robot scanning and the automatically 3D sensing of the scanned scene.

手持消费级深度相机三维扫描以低成本突破了专业三维扫描设备感知范围受限及对扫描环境要求严格的约束。如何用少量深度图生成高质量三维模型是实现手持消费级深度相机有效三维扫描的关键问题。本项目研究手持消费级深度相机三维扫描重建方法。通过量化序列中扫描目标形状变化与相机运动评估指标，动态分析各指标间相互关系，并依此构建深度图序列支撑子集，实现对序列的有效压缩；对深度图中目标进行特征提取，依据几何特征提高扫描目标的匹配精度；结合深度图滤波与特征增强优化选定的粗糙深度图表面；通过融合支撑子集中已优化的深度图而提出一种手持消费级深度相机稀疏序列融合三维扫描重建方法，真正用少量深度图有效、鲁棒、高质量地重建目标。手持消费级深度相机稀疏序列融合三维扫描对深度图序列的压缩与稀疏表示具有重要理论意义，将为计算机图形学领域带来丰富的3D模型，对机器人扫描及自动感知场景3D信息具有实际应用价值。

手持消费级深度相机三维扫描以低成本突破了专业三维扫描设备感知范围受限及对扫描环境要求严格的约束。如何提高消费级深度相机手持式扫描的鲁棒性是该技术得以推广的关键问题。本项目研究一种基于稀疏深度图序列融合的三维扫描重建新框架，实现用少量深度图有效、鲁棒、高质量地重建扫描对象的目标。具体而言，本项目系统而深入地研究了如下内容：.首先，研究稀疏深度图序列融合三维扫描重建基本方案。通过量化深度图序列的相机运动评估指标，动态分析各指标间相互关系，并依此构建深度图序列支撑子集，实现对序列的有效压缩。基于支撑子集的稀疏序列融合方法同时在物体级目标和场景级目标上进行了测试，取得了预期的效果，能够在剔除抖动帧干扰的同时降低冗余帧造成的相机匹配累积误差。在以24fps的物体级目标扫描测试实验中，支撑子集对原始深度图序列的压缩比例超过了80%。.其次，提出一种近邻重新权重局部质心（NRLC）的点云表面几何特征识别新方法。该方法首次将点云表面的凹、凸和边界特征纳入统一的框架，通过将特征区域点的质心推至远离局部表面而增大特征点和非特征点的对比度，实现对特征点的识别。NRLC方法被用于场景扫描测试中，显著提升了扫描片段间的匹配精度，同时NRLC几何特征检测结果用于场景点云的实例分割中也收到了良好的效果。.最后，通过对深度图表面执行多尺度滤波和几何特征增强实现对捕获深度图表面的优化。深度图表面优化为稀疏深度图序列融合生成高质量目标表面提供了有效支撑。.手持消费级深度相机稀疏序列融合三维扫描对深度图序列的压缩与稀疏表示具有重要理论意义，将为计算机图形学领域带来丰富的3D模型，对机器人扫描及自动感知场景3D信息具有实际应用价值。