Monocular-SLAM-based Autonomous Flight of a Quadcopter in Unknown Environment

論文翻譯標題: 未知環境下單視角同步定位與地圖構建的四旋翼全自動飛行
  • 譚 宇廷

學生論文: Master's Thesis


In this thesis we try to develop a system capable of navigating a modified Parrot AR Drone 2 0 in an indoor unknown environment using only onboard hardware like sensors controller and single board computer The quadcopter should be able to accomplish three tasks: 1 Constructing the 3D map of the surrounding and determining the position of itself in that environment 2 Three-dimensional position control and following a given path consisting of way points 3 Planning feasible path between user-specified points and avoiding the obstacles using the 3D map we construct AR Drone is very safe and lightweight but also has very limited payload so the particular challenge is dealing with the limited sensor variety and quality available by designing an effective data fusion framework To achieve the goal of positioning and mapping we decide to use normal webcam and perform monocular SLAM to lower the CPU loading because single board computer only has limited computing power Furthermore we choose the LSD SLAM because it can reconstruct the 3D environment Although the depth map is semi-dense only but this is a good trade-off between map quality and CPU load: If you choose other keypoint-based monocular SLAMs for their light CPU loading the map consisting of keypoints will not dense enough for obstacle avoiding But even laptop cannot do a dense 3D reconstruction fast enough for online usage using common algorithm Monocular visual SLAM cannot navigate a quadcopter all by itself That is it can’t determine an absolute attitude w r t any local coordinate system like NED or ENU frame and also the distance it get has to be multiplied by an unknown scaling factor The PX4 firmware has a reliable complementary-filter-based AHRS so we can estimate an attitude offset by comparing the attitude from PX4 AHRS to the one from LSD SLAM For the estimation of the unknown scaling factor we need the height reading from sonar sensor and use it to perform a slightly modified recursive least-squares method Sonar doesn’t drift like barometer and the noise is relatively low Finally for the path planning part we use ROS inside our single board computer and choose Moveit a very powerful and flexible planning package only available in ROS as our solution The results of the software framework and the scale estimation are shown Although the fully autonomous flight didn’t work because a lack of time this thesis has proved the concept and the development of the system proposed here should be carry on
獎項日期2015 八月 31
監督員Shau-Shiun Jan (Supervisor)