基于红外视频的机器人夜间视觉三维显示研究

发布时间：2018-11-05 20:15

【摘要】：在机器人应用领域,许多人工难以胜任的恶劣环境下的操作通常都在夜间无光的情况下进行,机器人的夜间视野范围和场景识别能力直接关系到机器人的任务执行能力。在夜间模式下,通常采用红外成像系统获取夜间场景图像,红外图像具有信噪比低、对比度低和缺乏深度感等特点。从夜间红外视频中还原出场景或物体的三维模型是计算机视觉领域研究热点,它需要综合计算机科学、信号处理等多种科学知识,且在视觉导航、军事、工业等领域都有着非常广阔的应用前景。因此针对夜间红外视频进行场景三维重建具有重要的意义。论文主要内容分三部分:第一部分是本课题的研究背景意义现状以及三维重建算法的现有基本理论方法的概述;第二部分主要介绍了常用的相机标定算法和红外视频预处理算法以及直接法与稀疏法视觉里程计。第三部分基于直接法与稀疏法视觉里程计三维重建算法对红外视频进行场景三维重建。本课题的创新点在于首次实现了单目红外视频的场景三维重建,重建算法采用直接法与稀疏法视觉里程计,该方法首先通过对红外热像仪标定获得热像仪内参,然后构建直接法与稀疏法视觉里程计模型,视觉里程计前端执行帧管理和点管理的任务,利用滑动窗口并借助高斯-牛顿迭代对总光度误差进行优化,计算出直接法与稀疏法视觉里程计模型所依赖的所有变量,完成定位热像仪和建图的任务。通过实验结果分析证明该方法能够实时准确地对单目红外视频进行三维重建。
[Abstract]:In the field of robot application, many operations in bad environment are usually carried out at night without light. The range of night vision and the ability of scene recognition are directly related to the task execution ability of robot. In night mode, infrared imaging system is usually used to obtain night scene images. Infrared images are characterized by low signal-to-noise ratio (SNR), low contrast and lack of depth sense. The restoration of 3D models of scene or object from night infrared video is a hot topic in the field of computer vision. It needs to integrate computer science, signal processing and other scientific knowledge, as well as in visual navigation, military affairs, etc. Industry and other fields have a very broad application prospects. Therefore, it is of great significance to carry out 3D scene reconstruction for night infrared video. The main content of this paper is divided into three parts: the first part is the research background significance of this topic and the overview of the existing basic theory and methods of 3D reconstruction algorithm; In the second part, the camera calibration algorithm, infrared video preprocessing algorithm and direct and sparse vision odometer are introduced. In the third part, infrared video scene is reconstructed based on direct method and sparse method. The innovation of this paper lies in the realization of 3D scene reconstruction of monocular infrared video for the first time. The reconstruction algorithm adopts direct and sparse visual mileometer. Firstly, the inner parameters of thermal imager are obtained by calibrating the infrared thermal imager. Then the direct and sparse visual odometer models are constructed, the front end of the visual odometer performs the tasks of frame management and point management, and the total photometric error is optimized by sliding window and Gao Si Newton iteration. All the variables dependent on the direct and sparse visual odometer models are calculated to complete the tasks of locating the thermal imager and building the map. The experimental results show that this method can reconstruct monocular infrared video in real time and accurately.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】