基于机器视觉的移动机器人定位与三维地图重建方法研究

发布时间：2018-07-17 02:31

【摘要】：随着工业4.0的全面发展,机器人已经不单纯是机械领域的代表,人工智能、机器视觉和云计算的发展,使得机器人的功能和应用场景更加丰富,人类社会对于智能机器人的需求也变得多样化,在这些众多的需求中,对于机器人能够在空间中自主定位并对空间进行三维重构这一项需求来说,是很多功能的基础,而且直接决定了机器人能否出现在人类未来的生活中。本文所研究的方向就是基于RGB-D相机的机器人能够在空间中进行自主定位的同时,来完成对场景地图的重建。为了降低噪声对系统的影响使输出结果更加准确并提高系统的稳定性,本文主要从以下几个方面进行研究:首先对RGB-D所拍摄的图像进行分析,得出噪声在深度相中的分布形况,然后根据噪声的分布,为了降低噪声对运动估计的影响,提出了一种基于平面化参数的特征点降噪提取方法,该种方法能够利用平面化参数将特征点的深度值进行修正,从而降低噪声的影响,最后获得一个相对稳定的运动估计,使系统的鲁棒性得到提高,同时给出了与传统提取法的比较实验。然后,对SLAM(Simultaneous Localization And Mapping)非参数化方程进行了讨论,根据观测数据的形式导出了其概率模型,针对概率模型中的最大似然估计导出了最小二乘方程,将原问题转化为非线性问题。同时,在求解非线性方程的过程中,本文讨论了其增量方程的稀疏性,稀疏性能够保证方程的求解效率,满足了系统的实时性要求。其次,本文在系统中加入了回环检测模块,回环检测模块的作用是能够降低定位算法中产生的累积误差,这种累积误差是只考虑了连续时间上的数据关联而造成的,而回环检测将空间中的约束考虑进来,通过对同一场景不同时刻所拍摄的图像进行匹配,提供新的约束来达到消除累积误差的目的。最后,根据最后优化得到的运动轨迹数据,来对场景进行三维重建,本文中展示了两种建图方式:点云地图和八叉树地图。根据使用场景和应用需求的不同,对这两种构建地图的方式进行了对分析和讨论。
[Abstract]:With the development of industry 4.0, robot is not only the representative of mechanical field, but also the development of artificial intelligence, machine vision and cloud computing. The demands of human society for intelligent robots have also become diversified. Among these many needs, the need for robots to be able to independently locate in space and reconstruct space in three dimensions is the basis of many functions. And directly determines whether the robot can appear in the future of human life. The research direction of this paper is that the robot based on RGB-D camera can locate its own position in space and complete the reconstruction of scene map at the same time. In order to reduce the effect of noise on the system and improve the stability of the system, this paper mainly studies the following aspects: firstly, the image taken by RGB-D is analyzed, and the distribution of noise in the depth phase is obtained. Then, according to the distribution of noise, in order to reduce the influence of noise on motion estimation, a method of feature point denoising based on plane parameters is proposed, which can be used to modify the depth of feature point. Finally, a relatively stable motion estimation is obtained and the robustness of the system is improved. At the same time, a comparison experiment with the traditional extraction method is given. Then, the nonparametric equations of slam (simulated localization And mapping) are discussed, its probability model is derived according to the form of observed data, and the least square equation is derived for the maximum likelihood estimation in the probabilistic model, and the original problem is transformed into a nonlinear problem. At the same time, in the process of solving nonlinear equations, the paper discusses the sparsity of the incremental equations, which can guarantee the efficiency of solving the equations and meet the real-time requirements of the system. Secondly, this paper adds the loop detection module to the system. The function of the loop detection module is to reduce the cumulative error in the localization algorithm, which is caused by only considering the data association in the continuous time. Loop detection takes space constraints into account and provides new constraints to eliminate cumulative errors by matching images taken at different times in the same scene. Finally, according to the motion track data obtained from the final optimization, 3D reconstruction of the scene is carried out. In this paper, two kinds of mapping methods are shown: point cloud map and octree map. According to the different usage scenarios and application requirements, this paper analyzes and discusses the two ways of building maps.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】