【作者】 郑伟；

【导师】 汪增福；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2014， 博士

【摘要】 微小型四旋翼无人飞行机器人(Quadrotor)是目前无人飞行机器人(Unmanned Aerial Vehicle, UAV)领域研究的热点问题,具有重要的理论意义和广阔的应用前景。相关的微小型四旋翼无人飞行机器人运动及位姿估计和定位导航,是其中极富挑战性的基础性和关键性研究课题。目前为止,世界上还没有一个适用于微小型四旋翼无人飞行机器人的视觉定位和导航系统可以满足实际的应用需求。本文的工作,围绕有限负载、动力及计算资源的微小型四旋翼无人飞行机器人平台展开,主要研究了视觉传感器主导,多传感器融合的微小型四旋翼无人飞行机器人运动及位姿估计和定位导航问题。微小型四旋翼无人飞行机器人的运动估计和位姿估计问题,是微小型四旋翼实现各种任务和应用的基础和关键。利用视觉传感器的运动及位姿估计方法,可以大致分为采用外部视觉的方法,或机载视觉的方法。外部视觉的方法通常更可靠更精确,但工作范围受到限制,可以用于微小型四旋翼无人飞行机器人的自主起降及精细控制。基于机载视觉的运动及位姿估计方法,可以使微小型四旋翼摆脱有限环境的限制,具有更好的灵活性,是当前研究的热点问题,可以广泛工作于更多样更复杂的任务和应用。实际中的众多相关应用,需要微小型四旋翼无人飞行机器人能够有效工作在一片较大范围的区域内。因此基于前面的相关技术积累,将有限范围内的基于视觉的位姿估计、运动估计及相关算法扩展到一定的较大的范围内,我们进行了基于机载视觉的微小型四旋翼无人飞行机器人定位导航问题的研究。本文的主要工作和创新点如下：(1)提出了基于外部视觉的,针对微小型四旋翼无人飞行机器人的鲁棒精确的位姿估计方法,并进行了实验验证。不同于现有大多数仅适用于室内环境的系统,我们的系统可以有效工作于室外环境。我们的一大特色是,以四旋翼飞行机器人的四个旋翼电机为视觉特征进行相关的位姿估计,在室外光照环境下,取得了比包括LED发光标签在内的彩色标签更加稳定和可靠的检测和定位效果。此外我们还研究了相关的位姿估计问题,提出一种解决共面点问题的快速精确的EMRPP位姿估计算法。该算法首先利用EPnP算法获得初始位姿估计,然后利用结合初步视觉结果的改进RPP算法得到精确的位姿估计结果。更进一步,针对微小型四旋翼的视觉特征的不完全观测情况,现有的方法都没有考虑这一问题,也得不到正确的位姿估计结果。我们结合视觉和机载IMU信息,提出了IMU+3P和IMU+2P位姿估计算法,有效解决了视觉特征不完全的问题,并得到准确的位姿估计结果。综合利用以上提出的方法,我们研发的位姿估计系统可用于微小型四旋翼无人飞行机器人的自主起降、机动控制和其它精细控制。(2)针对微小型四旋翼无人飞行机器人指定位置降落的特殊应用,我们提出一种基于平面地标的EIRPP位姿估计算法。该算法将IMU提供的飞行机器人部分姿态角信息,作为相关参数的初始估计代入EPnP算法中得到更精确的初始估计,并有效降低相关RPP算法迭代估计的未知量数目,实验结果表明,该算法可以获得快速精确的位姿估计结果。(3)充分利用微小型四旋翼无人飞行机器人的飞行特性,结合机载视觉和机载IMU,我们提出了一种基于自然环境特征的BRISK匹配的快速运动估计算法。针对目前微小型四旋翼的悬停实现多是基于光流方法实现,只能得到速度信息,且存在悬停点漂移等问题。我们引入基于特征匹配的快速运动估计算法,成功实现了微小型四旋翼无人飞行机器人的快速悬停功能。我们还针对实际应用中的一般情况,提出一种利用自然环境特征,结合机载视觉和机载IMU信息的机载情况下IMU+3P位姿估计算法。该算法可有效工作于非平面及平面场景情况,解决了单目视觉的初始化问题和绝对尺度估计问题,利用IMU提供的部分姿态角信息,有效降低了位姿估计问题的维度,获得了快速可靠的位姿估计结果。(4)在有限负载、动力及计算资源的条件下,探讨了微小型四旋翼无人飞行机器人平台的构建问题,提出了基于机载单目视觉、IMU和声纳的多传感器融合的单目视觉定位导航方案。该系统主要应用于GPS不可用的环境,以及无标识及无先验知识的环境。不同于其它大多数基于关键帧的系统,我们的系统同时利用了关键帧和关键点。为了保证精度和计算速度,使用基于GPU加速的SURF算法选择关键点。为了及时准确更新关键帧和关键点,提出基于快速运动估计和多级运动判决的更新方案,其在系统的较长时间执行过程中可以有效地减少误差累积。通常的单目视觉系统缺少绝对度量尺度,利用声纳传感器的距离信息完成初始化步骤,并较好地解决了绝对尺度估计问题。最后,通过综合利用特征点选择排序、RANSAC、局部光束平差法(Local Bundle Adjustmnet)等技术有效减少了系统误差及累积观测误差,实现了单目视觉主导的微小型四旋翼无人飞行机器人定位与导航功能。更多还原

【Abstract】 Nowadays, the research about quadrotor has got much attention in the UAV(Unmaned Aerial Vehicle) field. It has great significance and wide prospect of applications. Vision based motion and pose estimation, localization and navigation for a quadrotor are the fundamental and key problems in research. Up to now, there have no satisfied monocular visual localization and navigation systems for quadrotors in the world. The main work in this paper is to have a research on the problems, which are vision based multi-sensor fusion motion and pose estimation, localization and navigation problems for a quadrotor with limited payload, power and computational resource.The motion estimation and pose estimation of the quadrotor are the foundation and key for various missions. The vision based motion estimation and pose estimation approaches are mainly external vision based algorithms and onboard vision based algorithms. The external vision based algorithms are more reliable and robust, but have limited region. They could be used for the autonomous take-off and landing of a quadrotor, or the precision control. Onboard vision based approaches for the quadrotor are paid much attention to because of the flexibility and getting rid of limited region. They could be used for various and more complex missions.In real applications, the quadrotor might work in a large-scale region. Based on our former algorithms, we extend the existing approaches in limited region to a larger region. We have some research on the vision base localization and navigation for the quadrotor. The main work and contribution of this paper are as follows:(1) We give the external vision based robust and accurate pose estimation system for the quadrotor, and perform in real experiments. Our system could work well in outdoor environments, while most existing systems could only perform in indoor environments. One key characteristic of our system is that we only use the quadrotor’s own four rotors as the visual features, which is more reliable and robust than colored blobs and LEDs in outdoor environments. When all the four rotors of the quadrotor are observed rightly, we present the fast and accurate pose estimation algorithm EMRPP for the coplanar point problem. It gets the initial pose guess by non-iterative EPnP algorithm. By using the preliminary position result calculated by former vision step, we have modified the RPP algorithm and got the fast and accurate results of pose estimation. When the four rotors are observed partly, most of the existing approaches don’t mention this case and could not get right results. By using the vision data and the onboard IMU, we propose the IMU+3P and IMU+2P algorithms which could resolve this case and get fast and right pose estimation results. Taking full advantage of our former proposed algorithms, the pose estimation system could be used for the autonomous take-off and landing of a quadrotor or the precision control.(2) Considering the pinpoint landing of the quadrotor, we present the landmark based fast and accurate pose estimation algorithm EIRPP. It makes use of onboard vision and IMU data, utilizing these data in EPnP to get initial guess and improve the RPP algorithms. We get fast and accurate pose estimation results in the end.(3) Making the best of the omni-directional flight characteristic of the quadrotor, we propose the BRISK based fast motion estimation algorithm which uses the natural features and the onboard IMU. Considering the hovering flight of the quadrotor, most approaches utilize the optical flow algorithms. The optical flow algorithms could only obtain the velocity information and the hovering point might drift along the time. Using the BRISK based fast motion estimation algorithm, we realize the fast spot hovering of the quadrotor. For the general conditions, we present the natural features based fast and accurate pose estimation algorithm for the quadrotor which has limited pay load and computational resource. This algorithm makes use of the onboard camera, IMU and sonar. It could work for both non-planar and planar scenes and solves the initialization problem and the metric scale estimation problem of the monocular system effectively. By using the data from the IMU, the pose estimation problem is simplified and obtains more fast and accurate results of pose estimation.(4) Considering the limited payload, power and computational resources, we discuss the hardware platform construction for the quadrotor. We have presented a multi-sensor fusion based monocular visual localization and navigation system for the quadrotor. This system with an IMU, a sonar and a monocular down-looking camera as its main sensor is able to work well in GPS-denied and markerless environments. Different from common keyframe-based system, our visual localization and navigation system is based on both keyframes and keypoints. Considering the accuracy and computational time, GPU-based SURF is performed for feature detection and feature matching. The fast motion estimation algorithm and the multilevel motion judgment rule are presented for updating the keyframes and keypoints. This is beneficial to hovering or near-hovering conditions and could reduce the error accumulation effectively. The general monocular visual systems usually lack the metric scale. By using sonar data, we solve the metric scale estimation problem and get good initialization of the navigation system. The good features selected, RANSAC, Local bundle adjustment and some other measures are performed to reduce the error accumulation and optimize the results.In the end, we have realized the monocular localization and navigation system for the quadrotor.更多还原