RGB-D特征检测与描述方法及其应用研究
发布时间：2019-03-11 15:15
【摘要】：局部特征提取通常作为计算机视觉和图像处理等任务的第一步,例如:宽基线匹配,图像拼接以及图像分类等问题,因此局部特征性能的优劣直接影响整个系统最终性能的好坏。随着RGB-D(depth)传感器的快速发展,消费级别的RGB-D摄像机逐渐得到普及。相较于传统的RGB摄像机,RGB-D摄像机可以直接获取场景的深度信息,对于提高局部特征的性能有着天然的优势,因此设计一种高性能的RGB-D局部特征有着非常大的应用价值。局部特征提取主要涉及特征检测、特征描述和特征匹配,孤立研究其中某一内容对于局部特征性能的提升是有限的,因此如何合理地利用RGB-D信息构建局部特征是一项非常有技巧性的课题。本文主要针对现有的RGB-D描述算子LOIND(Local Ordinal Intensity and Nor-mal Descriptor)[1]进行优化,同时提出与其耦合度更高的RGB-D 检测算子,明显提升了RGB-D局部特征的性能。为了客观全面地评测RGB-D局部特征的性能,我们设计采集了标准的RGB-D局部特征评测数据集。本文主要内容和成果如下:1.将RGB-D摄像机固定在高精度的机械臂上,完成手眼标定后进行数据采集。设计采集了3大类,15小类的RGB-D数据集,包含尺度变换、旋转变换、视角变换以及光照变换,为局部特征研究者提供了极大的便利性。2.将自相关函数的思想分别作用于灰度图和点积图,从而设计了一种融合纹理信息和深度信息的RGB-D特征检测算子,解决了Harris检测算子[2]在光照变化剧烈时,检测失败的问题。同时该检测算子融合深度信息的方式与LOIND[1]比较一致,因此检测算子和描述算子的耦合度较高,明显地提高了 RGB-D局部特征的性能。3.基于点云重投影完成尺度估计和主方向估计,增强了RGB-D局部特征对于尺度变换和旋转变换的鲁棒性。同时从算法细节和代码两方面对LOIND[1]描述算子进行优化,进而提升RGB-D局部特征算子的时间效率和鲁棒性。
[Abstract]:The local feature extraction is usually the first step in the task of computer vision and image processing, such as wide base line matching, image stitching and image classification. With the rapid development of the RGB-D (depth) sensor, the consumption-level RGB-D camera is becoming more and more popular. Compared with the traditional RGB camera, the RGB-D camera can directly acquire the depth information of the scene, and has a natural advantage for improving the performance of the local feature, so that a high-performance RGB-D local characteristic is designed with very large application value. The local feature extraction mainly involves the feature detection, the feature description and the feature matching, and the isolation of some of the content is limited to the enhancement of the local feature performance, so how to use the RGB-D information reasonably to build the local feature is a very technical problem. In this paper, an RGB-D detection operator with higher degree of coupling is proposed, and the performance of the local features of the RGB-D is obviously improved. In order to evaluate the performance of the RGB-D local features in an objective and comprehensive manner, we design a standard RGB-D local feature evaluation data set. The main contents and achievements of this paper are as follows:1. The RGB-D camera is fixed on a high-precision mechanical arm, and after the hand-eye calibration is finished, the data acquisition is carried out. The design and acquisition of the RGB-D data sets of three categories and 15 small classes, including scale transformation, rotation transformation, visual angle transformation and illumination transformation, provide great convenience for the local feature researchers. The concept of self-correlation function is applied to the gray scale graph and the dot product graph respectively, so that an RGB-D characteristic detection operator for fusing the texture information and the depth information is designed, and the problem that the Harris detection operator[2] has failed to detect when the light change is violent is solved. At the same time, the fusion depth information of the detection operator is consistent with the LOIND[1], so the coupling degree of the detection operator and the description operator is high, and the performance of the local characteristic of the RGB-D is obviously improved. The robustness of the local features of the RGB-D to the scale transformation and the rotation transformation is enhanced based on the scale estimation and the main direction estimation of the point cloud re-projection. At the same time, the operator is optimized from the two aspects of the algorithm detail and the code, and the time efficiency and the robustness of the RGB-D local characteristic operator are improved.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2438391
[Abstract]:The local feature extraction is usually the first step in the task of computer vision and image processing, such as wide base line matching, image stitching and image classification. With the rapid development of the RGB-D (depth) sensor, the consumption-level RGB-D camera is becoming more and more popular. Compared with the traditional RGB camera, the RGB-D camera can directly acquire the depth information of the scene, and has a natural advantage for improving the performance of the local feature, so that a high-performance RGB-D local characteristic is designed with very large application value. The local feature extraction mainly involves the feature detection, the feature description and the feature matching, and the isolation of some of the content is limited to the enhancement of the local feature performance, so how to use the RGB-D information reasonably to build the local feature is a very technical problem. In this paper, an RGB-D detection operator with higher degree of coupling is proposed, and the performance of the local features of the RGB-D is obviously improved. In order to evaluate the performance of the RGB-D local features in an objective and comprehensive manner, we design a standard RGB-D local feature evaluation data set. The main contents and achievements of this paper are as follows:1. The RGB-D camera is fixed on a high-precision mechanical arm, and after the hand-eye calibration is finished, the data acquisition is carried out. The design and acquisition of the RGB-D data sets of three categories and 15 small classes, including scale transformation, rotation transformation, visual angle transformation and illumination transformation, provide great convenience for the local feature researchers. The concept of self-correlation function is applied to the gray scale graph and the dot product graph respectively, so that an RGB-D characteristic detection operator for fusing the texture information and the depth information is designed, and the problem that the Harris detection operator[2] has failed to detect when the light change is violent is solved. At the same time, the fusion depth information of the detection operator is consistent with the LOIND[1], so the coupling degree of the detection operator and the description operator is high, and the performance of the local characteristic of the RGB-D is obviously improved. The robustness of the local features of the RGB-D to the scale transformation and the rotation transformation is enhanced based on the scale estimation and the main direction estimation of the point cloud re-projection. At the same time, the operator is optimized from the two aspects of the algorithm detail and the code, and the time efficiency and the robustness of the RGB-D local characteristic operator are improved.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 ;Hand-eye calibration with a new linear decomposition algorithm[J];Journal of Zhejiang University(Science A:An International Applied Physics & Engineering Journal);2008年10期
，本文编号：2438391
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2438391.html
