Search

This thesis proposes methods for global localization in urban and natural environments. The proposed approaches focus on segment-based localization systems for loop closures in the context of SLAM, alleviating the need for explicit fine-tuning or parameter fitting. Initial analysis of current state-of-the-art systems are conducted and their performance in urban and natural environments is evaluated. We propose a hybrid feature descriptor method that does not rely only on the geometry of objects, but also considers the distribution of points within those objects. The segmentation differences encountered in natural environments are addressed by a novel key pose extraction technique. Inspired by the ability of deep networks to generalize to multiple terrains, we propose a novel deep architecture capable of generalizing between natural and urban scenes. We design a state-of-the-art point cloud based network with efficiency in mind that is also deployable on computationally constrained platforms, such as drones, legged robots, or UGVs. The proposed objective function recognizes similar objects without any semantic information. We also extract a matchability score between segments that we use in a probabilistic geometric verification stage for coarse registration. The analysis of the point-based architectures led to exploration of the bot- tlenecks in the computational complexity of these models. We propose a sub- sampling method allowing us to shrink an existing network architecture with- out sacrifice in performance. In addition, substantial analysis are presented on different training and test data variation for all the localization methods. The evaluation was performed on data from over 300 km of driving, that we manually pre-processed and made available to the community. The final contribution is a method for learning keypoint regions that can be applied to any differentiable descriptor. The proposed framework is a substitute for the first step in segment-based localization systems. This thesis also presents two applications of the above contributions. First, motivated by the expressive power of the geometric features, we propose a verification system that seeks to predict when localization failures occur in the environment. Specifically, we leverage some of the geometric features from our localization system to propose a classifier that predicts alignment risk based on overlap and degenerate geometric constraints. Second, we en- capsulate the proposed loop closure method into a full SLAM system which is also deployed on the ANYmal quadrupedal robot. The proposed deep learned loop closure method shows improved scalability when compared to geometric-based loop closure systems.