Autonomous navigation framework for structural inspection using an unmanned aerial vehicle

Abstract

This dissertation relates to the entire framework and methods of coverage path planning, 3D mapping, localization, and high-level control for structural inspection using an unmanned aerial vehicle (UAV). Recently, various robots are being used for the purpose of structure inspection or safety diagnosis, and their needs are also rising rapidly. Among the structure inspection using a robot, a lot of research has recently been conducted on inspection of various facilities and structures using an unmanned aerial vehicle. Among them, one of the most necessary parts of the industry is the study of the autonomous bridge inspection system. Many institutions, universities, and corporations around the world spend a lot of time and money to develop the system, but they have not yet developed a system that is stable enough to be put directly into the field. In this study, the framework for inspecting structures such as bridges using an unmanned aerial vehicle is explained step by step, and finally, the results of experiments in the actual field are attached.

The framework proposed in this study is largely divided into pre-flight, on-flying, and after-flying steps, and the required work is further subdivided within each step. First of all, for safe autonomous flight and efficient bridge inspection of the UAV, preliminary work via pre-flight is required before actual flight. There are more scenarios to check the local part than to check the entire target bridge, and if the entire map of the target bridge is secured, there are many advantages for the operation of an autonomous flying drone, such as obstacle avoidance and rapid inspection. Therefore, for the sake of precise mapping, the 3D point cloud map is created using the graph SLAM method proposed in this study using the data obtained by flying the UAV by manual operation of the user. Based on the 3D point cloud map obtained through pre-flight, the user can pre-specify the parts to be inspected for the bridge and perform coverage path planning. Through this, it is possible to efficiently plan the inspection path and to avoid obstacles such as branches and bridge columns. One of the most important things in structure inspection is the method of covering the entire area without missing parts in the actual inspection. In this study, the structure is divided into multi-layers to extract viewpoints from each layer and solve the optimal path connecting the local paths. Eventually becomes the TSP (Travelling Salesman Problem), so the optimal path is calculated through the LKH (Lin-Kernighan heuristic) Solver.

After this preliminary preparation, the UAV can autonomously fly for actual inspection. The most important is the localization of the UAV. Since the GPS signal is denied at the bottom or near the bridge, in this paper, a graph-based SLAM algorithm combining IMU, camera, and 3D LiDAR is used. In graph-based SLAM, the generation of the basic node has a great influence on the stability and optimization of the overall graph structure. For robustness, visual-inertial (VI) state estimates and the corresponding LiDAR sweep are combined into `subnodes'. And then they are combined into `supernodes' which comprises state and `submap (accumulated scan data)'. The constraints are generated using LiDAR-based normal distribution transform (NDT) and generalized iterative closest point (G-ICP) matching. Before fully globally optimized, NDT locally optimizes the relationship between subnodes in the hierarchical scheme. This enables accurate localization and mapping of the UAV's surroundings for autonomous navigation without operator intervention. Then, using the nonlinear least squares method, we find an optimized solution that minimizes the error. Using the graph structure as described above, the position information of the unmanned aerial vehicle acquired from various sensors can be easily fused. In addition, by using a high-level control method for inspecting the actual structure, the UAV maintains a constant speed between observation points and can respond to instantaneous gusts to ensure the quality of the acquired image and the stability at the time of actual inspection work.

Finally, experiments by scenarios were conducted on various bridges in actuality to verify the results of this system, and the effect of its performance was verified through comparative analysis with the state-of-the-art algorithms.