Path pre-planning optimizing SLAM uncertainty using reinforcement learning

Abstract

3D mapping accuracy is affected by the data acquisition path of the mobile mapping system. In this thesis, we will propose a path pre-planning method that optimizes 3D mapping accuracy in terms of the trajectory error, using the reinforcement learning technique and an RGB image and a depth image of the target area. Our method first predicts the LiDAR measurement quality of the scene from an RGB and depth image. In this step, we calculate normal vectors for each pixel point and the scene with more diverse normal vectors is considered as a scene with better LiDAR measurement. We then convert location measurement into graph structure using predicted LiDAR measurement. Using the obtained graph, we finally calculate the optimal path using the reinforcement learning technique. Lastly, our experiment in the simulation environment shows our method lowers trajectory error compared to a randomly generated path and also faster than existing path planning algorithms on the graph structure.