An improved photometric stereo through distance estimation and light vector optimization from diffused maxima region
An improved photometric stereo through distance estimation and light vector optimization from diffused maxima region
复制标题
DOI：
10.1016/j.patrec.2013.09.005
复制
发表时间：
2014-12
期刊：
影响因子：
--
通讯作者：
Jahanzeb Ahmad;Jiuai Sun;Lyndon N. Smith;Melvyn L. Smith
中科院分区：
文献类型：
--
作者：
Jahanzeb Ahmad;Jiuai Sun;Lyndon N. Smith;Melvyn L. Smith
虽然光度立体提供了一个有吸引力的技术，用于获取3D数据，使用低成本的设备，固有的局限性的方法，以限制其实际应用，特别是在测量或计量任务。我们在这里解决这个问题。传统的光度立体假设每个像素处的照明方向是相同的，这在真实的应用中通常不是这种情况，并且特别是在被观察对象的大小与工作距离相当的情况下。照明的这种缺陷可能使得用于获得场景的3D形状的后续重建过程易于出现低频几何失真和系统误差（偏差）。此外，对象的3D重建导致具有未知比例的几何形状。为了克服这些问题，开发了一种新的方法来估计物体与相机的距离，该方法采用光度立体图像而不使用其他额外的成像方式。该方法首先确定朗伯扩散极大值区域来计算物体到相机的距离，从该距离中能够导出校正的每像素光矢量，并且随后可以估计物体的绝对尺寸。我们还提出了一个新的校准过程，允许动态（作为一个对象在视场中移动）计算每个像素的光矢量，几乎没有额外的计算成本。合成以及真实的数据进行的实验表明，该方法提供了改进的性能，实现了减少估计的表面法线误差高达45%，以及平均高度误差重建表面高达6毫米。此外，当与传统的光度立体相比，所提出的方法减少了平均角度和高度误差，使得它是低的，恒定的，并且独立于正常工作范围内的对象放置的位置。
Although photometric stereo offers an attractive technique for acquiring 3D data using low-cost equipment, inherent limitations in the methodology have served to limit its practical application, particularly in measurement or metrology tasks. Here we address this issue. Traditional photometric stereo assumes that lighting directions at every pixel are the same, which is not usually the case in real applications, and especially where the size of object being observed is comparable to the working distance. Such imperfections of the illumination may make the subsequent reconstruction procedures used to obtain the 3D shape of the scene prone to low frequency geometric distortion and systematic error (bias). Also, the 3D reconstruction of the object results in a geometric shape with an unknown scale. To overcome these problems a novel method of estimating the distance of the object from the camera is developed, which employs photometric stereo images without using other additional imaging modality. The method firstly identifies Lambertian diffused maxima region to calculate the object distance from the camera, from which the corrected per-pixel light vector is able to be derived and the absolute dimensions of the object can be subsequently estimated. We also propose a new calibration process to allow a dynamic(as an object moves in the field of view) calculation of light vectors for each pixel with little additional computation cost. Experiments performed on synthetic as well as real data demonstrates that the proposed approach offers improved performance, achieving a reduction in the estimated surface normal error of up to 45% as well as mean height error of reconstructed surface of up to 6 mm. In addition, when compared to traditional photometric stereo, the proposed method reduces the mean angular and height error so that it is low, constant and independent of the position of the object placement within a normal working range.