【作者】 陈宇；

【导师】 李建成； 闫利；

【作者基本信息】 武汉大学 ， 大地测量学与测量工程， 2021， 博士

【摘要】 近年来,自主移动设备在多方领域受到了极大地关注,自主无人系统的服务范围也获得了极大的拓展,比如无人机、无人车、四足机器人甚至无人船等。通过视觉传感器获取光学影像并进行分析是自主无人系统实现自主导航与自主感知的重要途径,视觉自主导航技术作为自主无人系统实现自身智能化与自主运动的基础技术,支撑了自主系统进行后续目标跟踪、自主避障以及多机协同任务。本文围绕视觉自主导航技术及其主导下的多传感器融合导航技术这一主题开展研究,取得的创新性成果包括:(1)提出一种基于多级匹配策略的点线特征增强匹配算法。自主无人系统在依靠自主移动设备所搭载的视觉传感器进行特征追踪时,往往会遇到复杂场景中弱纹理或重复纹理区域导致的特征匹配可靠性低、匹配精度差,以及移动设备路径突变导致的影像动态模糊和特征不连续问题。为了实现视觉导航系统的特征追踪需求,本文提出了一种基于多级匹配策略的点线特征增强匹配算法。首先,在待匹配影像上提取FAST角点特征与线特征,并通过BRIEF描述子粗匹配两帧影像间的特征点,继而计算并评估两帧影像间的变换矩阵,最后利用合适的变换矩阵实现点特征与线特征的几何对应匹配。与现有的ORB特征匹配算法相比,本文算法在符合自主导航系统实时性要求的前提下,能够有效提高算法匹配结果的正确率以及有效匹配特征对的总体数量。(2)提出一种基于改进PWC-Net补帧网络的关键帧填补与选取算法。自主移动设备在运动路径突变时所进行的高机动变向与传感器晃动会造成影像动态模糊与特征不连续,甚至小范围内数据丢帧等问题,但现有的视觉导航系统中相应的关键帧选取算法只能根据移动设备运动状态或者系统轨迹跟踪质量来从影像流中选取现有影像作为关键帧,影像的数量与质量极大地制约了系统的轨迹跟踪连续性与最终完整轨迹的精度。为了维持视觉自主导航系统的连续性与稳定性,本文提出了一种基于改进PWC-Net补帧网络的关键帧填补与选取策略。具体地说,本算法首先采用两步分解法计算连续影像帧之间姿态的变化,然后根据姿态变化预测的系统运动状态通过改进PWC-Net网络增加关键帧或者在特定间隔内自动选取关键帧。在KITTI公开数据集中多种场景数据上的实验表明,本文改进的PWC-Net网络能够生成优于现有补帧网络结果质量的中间帧结果,其所支持的视觉自主导航系统能够有效地实现轨迹跟踪任务。(3)提出一种基于因子图的单目视觉、IMU与GNSS融合的自主导航算法。单目视觉传感器在移动设备高机动自主运动时,往往会受到载荷平台变向过快与剧烈光照变化的影响。而视觉传感器采集到的影像也会存在大量的动态模糊,与弱纹理环境一起,降低了视觉自主导航系统的连续性与精度。为了增强视觉自主导航系统的稳定性,本文提出了一种视觉主导下的多源数据融合导航算法。首先给定点、线特征相应的权值,同时参与系统的位姿估计,其次将IMU测量信息共同加入到滑动窗口中进行优化,最后通过因子图优化的方式,将GNSS数据作为再约束条件,进一步优化系统轨迹精度。在包含多种不同场景类别的公开数据集上的实验表明,利用本文算法所生成的轨迹跟踪结果更为完整和稳定,可以较好地实现系统的自主导航需求。更多还原

【Abstract】 In recent years,autonomous mobile devices have received great attention in many fields,and the application of unmanned automous systems has become more and more extensive,such as unmanned aerial vehicles,unmanned vehicles,quadruped robots and even unmanned ships.Obtaining and analyzing optical images through visual sensors is an important way for unmanned autonomous systems to realize autonomous navigation and autonomous perception.This paper focuses on the subject of visual autonomous navigation technology under its leadership.The innovative results obtained include:(1)This paper proposes a point-line feature-enhanced matching algorithm based on multi-level matching strategy.When unmanned autonomous systems rely on the visual sensors carried by autonomous mobile devices for feature tracking,they often encounter problems such as low feature matching reliability and low matching accuracy caused by weak texture or repeated texture regions in complex scenes,as well as the problems of dynamic blur and feature discontinuity caused by path mutation of mobile device.In order to achieve the feature tracking requirements of the visual navigation system,this paper proposes a point-line feature-enhanced matching algorithm based on a multi-level matching strategy.Firstly,FAST corner features and line features are extracted from the image,and the feature points between the two images are roughly matched by BRIEF descriptors,then the transformation matrices between the two images are calculated and evaluated,and finally the geometric matching of point features and line features are realized by using the appropriate transformation matrix.Compared to the existing ORB matching strategy,this algorithm can highly increase the accuracy rate and the number of matching points on the premise of meeting the real-time needs of autonomous navigation system.(2)This paper proposes a key frame filling and selection algorithm based on improved PWC-Net.The high maneuverability and sensor shaking of the autonomous mobile device during the sudden change of the motion path will cause the image dynamic blue and feature discontinuity,and even the data frame loss in a small area.However,the corresponding key frame selection algorithm in the existing visual navigation system can only select the exist ing image from the image stream as the key frame according to the motion state of the mobile device or the tracking quality of the system trajectory.The quantity and quality of the image greatly restrict the continuity of track tracking and the robustness of the final complete track.For the sake of keeping the stability of visual autonomous navigation system,this paper proposes a key frame filling and selection strategy based on improved PWC-Net.Specifically,this algorithm first uses two-step decomposition method to calculate the attitude changes between consecutive image frames,and then adds key frames by improving PWCNet,or automatically selects key frames in a specific time interval according to the platform motion state predicted by the attitude changes.Experiments on a variety of scene data in KITTI public dataset show that the improved PWC-Net can generate intermediate frame results due to the quality of the existing complementary frame network results,and the visual autonomous navigation system supported by the improved PWC-Net can effectively achieve the trajectory tracking task.(3)We proposes an autonomous navigation algorithm based on factor graph-based version,inertial measurement unit and GNSS fusion.The monocular vision sensor is often affected by the rapid change of the load platform and the severe illumination change when the mobile device moves autonomously with high mobility.The images collected by the vision sensor will also have a large amout of dynamic blur,which,together with the weak texture environment,decrease the robustness of the visual autonomous navigation system.For the sake of enhancing the results of the system,we designed a new algorithm.Firstly,the corresponding weights of the points and line features are given to jointly participate in the pose estimation of the system.Secondly,the IMU measurement information is added to the sliding window for optimization.Finally,the GNSS data is used as a re-constraint to further optimize the trajectory accuracy of the system through factor graph optimization.Experiments on open datasets with different scene categories show that the trajectory tracking results generated by this algorithm are more complete and stable,and the autonomous navigation requirements of the system can be better realized.更多还原