【作者】 叶波；

【导师】 许超；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2017， 硕士

【摘要】 本论文关注融合单目视觉和惯性传感的里程计方法研究,旨在实现能在四旋翼飞行器平台上实时运算,并且具备高精度和鲁棒性的自主定位算法,主要分为两个场景。首先在笔者参加的国际空中机器人大赛(the International Aerial Robotics Competition,IARC)中要求飞行器在不依赖外界辅助导航的条件下实现自主定位。针对该比赛的特殊场景(地面具有丰富纹理信息和规则网格特征),提出了一种基于光流法和网格信息,同时融合惯性测量单元(Inertial measurement unit,IMU)的定位方法。.首先在传统光流法上作了改进,基于固定块匹配方法,实时获取相机平动速度,然后将其积分作为初始估计,并通过地面网格信息来校正光流积分得到的位置信息,最后融合IMU数据进行信号平滑,确保飞行器位姿无累积误差。该方法在2016年IARC亚太赛区比赛中得到成功应用。其次针对一般场景,选择基于稀疏直接法的单目视觉里程计算法,无需计算每帧图像的特征描述子,计算速率提高,并且设计了模块化的扩展卡尔曼滤波(extend Kalman Filter,EKF)框架,融合单目视觉里程计计算得到的相机位姿和IMU数据。对于预测部分,基于IMU驱动系统的误差状态运动学实现,对于测量部分,由视觉里程计提供的位置和姿态作为量测,另外进行了测量量和状态量的时间同步处理,以及视觉算法位姿检测失败时的校正处理。在开源数据集测试本算法的准确性和鲁棒性。更多还原

【Abstract】 This thesis is aimed at proposing a precise,robust visual inertial odometry tech-nology which is real-time computed on the quadrotor onboard.Then,it is applied in two scenes basically.First,we realize a visual inertial odometry based on optical flow and grids infor-mation for the International Aerial Robotics Competition(IARC),in which the UAV is required to localize itself without any external global positioning assistance.This work can be summarized into three parts.First,given attitude and height information measured by the inertial measurement unit(IMU)and altimeter,a rough location can be generated by the optical flow with attitude compensation and metric scaling.Then,we note that the grid structure of the ground is formed by straight lines,which can be used to refine the localization result in the previous stage.Finally,a modular Kalman filtering framework is applied to assimilate the IMU and visual localization result.The effectiveness of the proposed algorithms in this work is validated both in the simulation package and real experiments.The proposed technique has been successfully imple-mented to the competition in 2016.It helped our team winning the first place of the IARC Asian-Pacific venue and remaining the best performance among all the competi-tion teams in 2016.Then,we propose a mono visual odometry based on sparse direct methods.It does not need to extract and match the features on each frames,which leads to the low computer power.Next,we seek to fuse the visual and IMU data in a filtering framework for optimal pose estimation.The filter frame work is based on error state kinematics for IMU-driven systems for the propagation part,and we include the camera as an additional sensor with the mono visual odometry to compensate for the temporal drift of the IMU.Moreover,we treat the pose estimation part as a black box,which result in the independence of the ubderlying pose estimation.Validation results tested on the open source data sets indicate the accuracy and robustness of the algorithm.更多还原