基于随机回归森林与多源数据融合的高精度三维动态形状获取

项目名称： 基于随机回归森林与多源数据融合的高精度三维动态形状获取

项目编号： No.61272049

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 吴怀宇

作者单位： 中国科学院自动化研究所

项目金额： 82万元

中文摘要： 高精度三维动态形状的获取在虚拟现实、游戏娱乐、影视工业等领域具有广阔的应用前景。传统的三维激光扫描仪采集速度慢，难以实时扫描动态运动的物体；而立体视觉系统对采集场景的要求高，在缺乏纹理的区域难以取得鲁棒的效果。作为一种近年来发展迅速的新型采集设备，飞行时间三维相机可实时获取动态物体的深度信息，造价低廉，具有传统扫描仪无法比拟的许多优点。但相应的本质缺点是传感器分辨率低，随机噪声大，不能直接生成高精度的形状。本项目基于随机回归森林理论，构建一个新的基于统计机器学习的几何处理框架，以获取高精度的三维动态形状。随机回归森林是一种有效的监督学习方法，可自动进行特征的选择且无需进行交叉验证。基于所估计的置信度矩阵，我们提出一种深度超分辨率方法，凸优化理论保证了鲁棒性和计算效率。我们进一步将时空几何平滑约束和高清视频纹理约束融合在贝叶斯统计框架中，并结合数字几何处理方法，来重建最终的高精度动态形状。

中文关键词： 多源数据融合；随机回归森林；TOF深度相机；动态三维形状获取；

英文摘要： High-quality dynamic 3D shape capture has abroad applications in virtual reality, video game, entertainment, film industry etc. Existing 3D laser scanner perform slowly, and can not scan dynamic objects in real-time. On the other hand, stereo vision system requires strict and controllable scenes and can not achieve good results in regions without texture. As a rapidly developed device, time-of-flight 3D camera is cheap, and can capture depth information of dynamic objects in real-time. But this kind of camera has low resolution, high noise, and can not recover high-quality shapes. In this project, based on random regression forests (RRF), we aim at building a statistic machine learning framework to capture high-quality dynamic geometry. RRF is an effective supervised method, which can automatically perform feature-selection while no needing cross-validation. Based on the confidence matrix built, we proposed a depth super-resolution method, which can guarantee robustness and effectivity due to convex optimization. By exploiting the digital geometry process methods, We further integrate space-temporal constraint and high-quality texture constraint into a Bayesian statistic framework, to build final high-quality dynamic shapes.

英文关键词： Multi-Data Fusion；Random Regression Forests；TOF Depth Camera；3D Dynamic Shape Capture；