Vision based localization and mapping for UAV navigation

Abstract

To extend the UAV application scenario from the open-air environment into the indoor one, the UAV must have the capacity to sense the environment. In detail, the UAV needs to localize itself and map the surroundings with the equipped sensors and the onboard computer. This thesis presents methodologies, system designs, and experiment results, focusing on localization and mapping tasks that enable the UAV to percept in the complex environment with its onboard resources. For the localization task, a novel stereo visual-inertial pose estimation method is present. Designed feedback or feedforward loops are introduced to achieve the stable control of the system, which includes a gradient decreased feedback loop, a roll-pitch feedforward loop and a bias estimation feedback loop. Moreover, in this framework, the frontend and backend are further decoupled with the keyframe correction mechanism. Covisibility-based bundle adjustments are moved to the backend so that the frontend is kept simple and fast. For the mapping task, a newly navigation oriented designed mapping kit is present. The design extends the conventional occupancy voxel map and probability-based sampling model, with separated maintained cylindrical coordinate based localmap and cartesian coordinate based globalmap. This design is suitable for a group of navigation strategies, which combine global planning and local planning. Also, by considering the sensor model and applying the visibility check, the globalmap can automatically filter out the dynamic obstacle while the localmap presents both dynamic and static obstacles nearby. The above two kits were first tested in the benchmark and simulation environment and verified in the real environment. Extensive experimental results and detailed system configuration are presented throughout the thesis. Furthermore, the thesis also presents an end-to-end UAV vSLAM simulation platform. The simulation platform lowers the barrier to carry out the algorithm testing and validation before field trials. The proposed localization kit and mapping kit, together with a planning kit, were integrated into the simulation platform. In this end-to-end simulation, we achieved click and fly level autonomy UAV navigation. Finally, the conclusion contains the on-going relevant works and proposing future research opportunities.