Heterogeneous robots localization based on convolution neural network

A complete simultaneous localization and mapping system will care about the global consistency of camera trajectory and map, but it also means that more computing resources are needed to calculate the global optimization. When resources are limited and global path is not being concerned (such as UAV landing, short-term control), only visual odometry can be used instead of complete SLAM. Hence, for hybrid team robots of Unmanned Aerial Vehicle (UAV) and Unmanned Ground Vehicle (UGV), one important goal is to reckon the relative transformation between several kinds of robots. However, after reviewing several kinds of earlier studies on point cloud registration problem, most of the previous researches only concerned with the geometric inference, which did not include the semantic information. In this thesis, I intend to carry on collaborative semantic and geometric inference to improve point cloud registration problem under larger initial offset by making use of a novel algorithm which based on expectation maximization (EM) technique. The point associations and semantic labels between two point clouds are treated as latent variables. Moreover, the evaluation and visualization have been done on the several subsets of KITTI dataset. The results show that we can successfully acquires the accurate rigid body relative transformation between two dense 3D point clouds. Furthermore, compared with standard Generalized ICP (GICP) available in the Point Cloud Library, the registration accuracy of Semantic ICP (SICP) is improved and SICP is more robust.Master of Science (Computer Control and Automation