基于编码结构光的三维测量方法研究

发布时间：2018-02-28 05:08

本文关键词： 三维测量编码结构光特征点检测解码深度学习　出处：《五邑大学》2017年硕士论文　论文类型：学位论文

【摘要】：编码结构光投影三维测量方法具有无损、成本低、精度高、设备简单的优点,是目前三维测量、三维重建应用中主要采用的技术之一。但在编码的设计、特征点匹配、以及解码的鲁棒性方面仍然存在技术难点。由于测量对象表面颜色、纹理和亮度的不确定性,以及光照的影响,动态目标的时间相关性和空间相关性较弱,现有的空间编码结构光编解码算法容易受被测量对象表面颜色、纹理和光照的影响而使其三维测量鲁棒性不高。为此,本文以伪随机编码理论为基础,以二值几何符号为编码符号,设计了不易受表面颜色、纹理及光照变化影响、具有抗噪声能力、编码容量大、窗口尺寸小的编码图案。同时,根据编码图案的特殊结构,设计了一种多模板特征点检测算法,实验证明此算法抗噪性能强,对具有不同颜色、纹理、光照和曲率的表面仍具有较好的特征点检测效果。为最大限度降低畸变图像特征点检测的错误率,在基于极限约束优化的基础上扩展了平面约束和拓扑约束,提高了特征点检测精度。与编码方案相对应,本文设计了一种解码速度快、精度高的解码方法,此解码方法利用基于深度学习的卷积神经网络实现解码。首先采集大量的不同颜色、材质表面的三维测量目标在不同光照条件下的编码结构光投影的编码符号图像样本,用于训练卷积神经网络,然后在解码阶段利用训练好的卷积网络来对编码符号图像进行分类识别,实现对受目标物体表面调制而形成的畸变图案的解码,并进而完成三维信息测量。实验表明,本文解码算法对编码符号测试集识别率可达98.07%,对颜色变化、阴影、纹理结构变化或反射特性不同的动态目标、不同尺寸的样本都具有较好的解码效果,并且具有一定的抗噪性能。为验证本文设计的基于十字的多模板特征点检测算法、基于卷积网络的解码算法及优化机制的三维测量方法性能,选取表面特性不同的物体,分别对其特征点检测效果、点云数据、深度图、重建效果进行分析。实验结果表明,在结构光三维测量中,本文提出的编码模版、特征点检测算法、解码算法表现出了更好的精度和鲁棒性,并可以进一步应用于实际动、静态目标的高精度三维测量。
[Abstract]:The coded structured light projection 3D measurement method has the advantages of nondestructive, low cost, high precision and simple equipment. It is one of the main techniques used in the application of 3D measurement and 3D reconstruction. Because of the uncertainty of surface color, texture and brightness, and the influence of illumination, the temporal and spatial correlation of dynamic target is weak. The existing space-coded structured light codec algorithms are easy to be affected by the surface color, texture and illumination of the measured object, which makes the 3D measurement less robust. Therefore, based on the pseudorandom coding theory, Taking binary geometric symbols as coding symbols, a coding pattern with anti-noise ability, large coding capacity and small window size is designed, which is not easily affected by the changes of surface color, texture and illumination. At the same time, according to the special structure of the coding pattern, A multi-template feature point detection algorithm is designed. Experiments show that the algorithm has strong anti-noise performance and has different colors and textures. In order to minimize the error rate of feature point detection in distorted images, plane constraints and topological constraints are extended based on the optimization of limit constraints. The accuracy of feature point detection is improved. Corresponding to the coding scheme, this paper designs a decoding method with high decoding speed and high precision. This decoding method uses convolutional neural network based on deep learning to decode. Firstly, a large number of coded symbol images of different color and material surface 3D measurement objects are coded under different illumination conditions. It is used to train the convolutional neural network, and then use the trained convolution network to classify and recognize the coded symbol image in the decoding stage, so as to decode the distorted pattern formed by the modulation of the target object's surface. The experimental results show that the decoding algorithm can recognize the coded symbol test sets with a rate of 98.07, and the dynamic targets with different color changes, shadows, texture structure changes or reflection characteristics can be obtained. All samples of different sizes have good decoding effect, and have a certain anti-noise performance. In order to verify the multi-template feature point detection algorithm based on cross designed in this paper, Based on the decoding algorithm of convolution network and the performance of 3D measurement method based on optimization mechanism, the feature point detection effect, point cloud data, depth map and reconstruction effect of objects with different surface characteristics are analyzed respectively. The experimental results show that, In structured light 3D measurement, the coding template, feature point detection algorithm and decoding algorithm presented in this paper show better accuracy and robustness, and can be further applied to high precision 3D measurement of moving and static targets.
【学位授予单位】：五邑大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】