基于各向异性点光源的近场光度学三维重建问题研究

Conventional Photometric Stereo (PS) methods usually adopt the assumption of parallel or distant point lighting conditions, so as to simplify the procedure of surface normal calculations. However, for the PS systems with nearfield non-isotropic point light sources, the reconstructed surface normal maps are usually lack of precision subject to the inhomogeneous lighting conditions. In this proposal, a novel PS framework is investigated for the accurate 3D reconstruction from nearfield non-isotropic point light sources. Major research contents and contributions are generalized as follows. The first research work is to establish an accurate lighting model for the adopted non-isotropic point source. Compared with conventional PS lighting models that only consider the incident light direction, the radiance model of light source is also considered in the proposed lighting model. To calculate the incident light intensity for each surface point more precisely, both the angular and distance attenuations are considered. The second research work is about the system calibration. Compared with conventional point light-based PS methods that only need to determine the position of light source, both light source position and principle optical axis of light source should be precisely calibrated in the proposed PS framework. To calculate the light source position, traditional multi-sphere-based method can be used. To determine the principle optical axis of light source, a reference plane-based method is proposed in this work. With the system calibration parameters and the proposed lighting model, for each surface point, its lighting condition can be fully determined. And the primary 3D model can be reconstructed via classical surface integration procedure. The last research work is to optimize the parameters in the primary lighting model. Based on the primary obtained 3D model, an iterative optimization procedure is introduced to improve the precision of lighting model with respect to the unknown distance factor. With above researches, a new framework for non-isotropic point light source-based PS systems can be established. And more accurate 3D model can be obtained based on a rather simple PS setup, which only consists of some single LED as light source and one camera. It provides an accurate and low-cost technical means for wide 3D acquisition applications.

现有的光度学三维重建方法为简化法向模型的求解问题，多采用平行光或远距离点光源假设，然而对于各向异性点光源条件下的近场光度学三维重建问题，目前尚缺乏系统而深入的研究。在本项目中，我们首先将光源自身发光模型引入到光度立体视觉框架中，通过综合考虑光源特性、入射光夹角、出射光线的角度及距离衰减等多种因素，对非标准点光源下的光照场进行精确建模。通过所提出的标定方法获取光源位置及主光轴方向等关键参数，根据光照场模型和系统标定参数建立任意场景点的光照条件数学模型，计算出每个点的初始入射光方向和强度。由于初始模型中包含未知距离参数，研究中拟基于初始三维重建结果采取迭代优化策略对相关参数进行优化，提高最终的法向场重建精度。通过本研究的开展，进一步完善了非标准点光源条件下的近场光度立体视觉方法理论框架，基于普通LED点光源组成的光度学系统实现高精度的三维重建，为低成本三维信息采集提供了一种新的方法和技术手段。

传统的光度学三维重建研究中，多采用平行光或者远距离点光源，从而避免光照场不均匀现象的产生，然而对于近场LED光源，这种非均匀光照现象是不可避免的，而且难以采用线性补偿策略来改善，在这种光源条件下，要实现精确的光度学三维重建，最大的挑战来自于如何对这种非标准光源条件下的非均匀光照场进行建模，准确计算出每个图像点的入射光方向及强度，从而实现精确的法向量计算过程。项目主要研究内容包括：1）各向异性点光源照射下的光照场建模；2）各向异性点光源参数标定方法；3）各向异性点光源光度学模型参数优化及有效性验证。. 项目所取得主要研究进展包括：1）搭建了由8个红外LED和一个720P摄像头构成的光度学3D重建实验装置，可在0.5秒内完成一组光度学图像的采集；2）针对LED这种非标准点光源，提出了一种包含LED光源特性、角度衰减、距离衰减等多个因素在内的精确光照场建模方法；3）提出了一种基于多个反射球体目标的LED光源空间位置参数计算方法；4）提出了一种基于参考平面等亮度分布曲线拟合的主光轴参数计算方法；5）针对所建模型中的距离未知参数，提出了一种迭代优化策略，进一步提升了光照场建模和法向计算精度。. 实验阶段，将所研究方法与三种常规光照场建模方法进行了对比，平面3D重构结果显示，所提出的QPLM模型法向重建平均误差约1.18度，方差仅有0.018度，明显优于其他三种光照模型，在人脸等复杂目标的3D重构实验中，基于所提出的近似点光源模型和光度学系统，能够获得满意的3D人脸重构效果。研究结果表明，通过所提出的近场近似点光源光度学重建方法，可以基于多个单LED光源实现较为精确的人脸等复杂目标的表面3D重构，从而为低成本精确3D信息采集提供了一种有效的技术手段，如果配合高速成像模块，该方法也适用于动态目标的高精度3D成像，因此有着广阔的技术应用前景。