【作者】 刘康；

【导师】 孙汉旭；

【作者基本信息】 北京邮电大学 ， 机械工程， 2018， 硕士

【摘要】 随着场景的不断增大,视觉SLAM系统中光束平差法的时间和内存消耗会逐渐的增加。大尺度场景下大量资源的消耗限制了 SLAM系统在低配置的电脑和ARM开发板中的应用,从而也限制了在智能机器人中的应用。同时,大尺度场景下视觉SLAM建立的三维地图无法指导机器人进行精确导航和路径规划。因此,本文针对大尺度视觉SLAM的光束平差法进行研究,降低SLAM光束平差法优化求解过程中的时间和内存的消耗,建立更加准确的机器人导航地图,对SLAM系统来说有着十分重要的意义。本文的主要内容如下:1.大尺度环境下视觉SLAM局部光束平差算法研究。对视觉SLAM框架下局部光束平差法进行分析,提出一种改进的局部光束平差算法。应用“五点法”初步筛选出局部地图中的关键帧,然后通过帧之间的距离进行筛选。提出一种高精度地图点牵引的优化方法,以提高局部光束平差法的精度。2.大尺度环境下视觉SLAM分段全局光束平差算法研究。对优化过程中变量太多的问题,提出改进的分段全局光束平差法。应用连续关键帧之间可视地图点数目作为初步分段方法,再利用Ncut(Normalized cut)的分段方法。给每一个分段赋予一个运动变量,然后优化求解每个分段的运动位姿变量,更新整个地图。应用该方法以达到减少全局光束平差法中的优化变量数目,保证系统精度的目的。3.大尺度环境下视觉与激光融合SLAM系统光束平差法研究。分析多种激光SLAM系统的优缺点,选择合适的激光SLAM系统。在视觉SLAM建立的离线地图基础上,应用激光SLAM辅助,采用离线的局部光束平差法优化方法,以获得更好的导航和路径规划地图。4.视觉SLAM系统光束平差算法实验研究。对改进的算法进行实验研究和分析,以验证实验中算法的可靠性和有效性。更多还原

【Abstract】 As the scene increases,the time and memory consumed of SLAM optimization will increase greatly.Therefore,real-time operation in the configuration of a low computer or arm platform will be very difficult in the large-scale visual Simultaneous Localization And Mapping(SLAM).At the same time,the map built by SLAM in large-scale scenes is not enough for robot navigation and control.It is important to reduce the time and memory consumption and to establish a better map though bundle adjustment in the large-scale SLAM.And it is very important for the real-time of SLAM system and robot’s navigation and path planning.The main contents of this article are as follows:1.Research on Local Bundle Adjustment(LBA)in large-scale scenes.The Local Bundle Adjustment method is analyzed,and an improved LBA is proposed.In creating a new local map,the "five point method" and the distance between the key-frames are used to determine the key-frames that have a closer relationship with the current key-frame to insert into the local map.On the other hand,the application of high-precision map point traction method makes the LBA method to improve the accuracy of optimization.2.Research on Global Bundle Adjustment(GBA)in large-scale scenes.In the Global Bundle Adjustment method,an improved segmentation GBA method is proposed.The number of visible map points between successive key frames is used as the initial segmentation method and then using Ncut(Normalized cut)segmentation method.According to the idea of segmentation,the key-frame sequence to be optimized is divided into a plurality of adjacent sub-sequences,and then only one sub-sequence of motion-position variables needs to be optimized and solved.This method can greatly reduce the optimization variables in the Bundle Adjustment method and ensure the high precision of the system.3.Research on bundle adjustment method of vision and laser fusion SLAM system in large-scale scenes.Analysis of the advantages and disadvantages of a variety of laser SLAM system,select the appropriate laser SLAM system.Laser SLAM has an advantage in creating a navigable map as to visual SLAM,so laser SLAM-assisted methods are used.Using off-line optimization of LBA method,it can get a new and better navigation map.4.Experimental on bundle adjustment based on RGBD visual SLAM System.The improved algorithm is studied and analyzed experimentally,and the experimental results verify the reliability and effectiveness of the algorithm.更多还原