Classified Object Localization in Slam and Loop Closure Through Reinforcement Learning

Maps generated by many visual Simultaneous Localization and Mapping (SLAM) algorithms consist of geometric primitives such as points, lines or planes. These maps offer a topographic representation of the environment, but they lack semantic information about the scene and objects in the environment. Object classifiers leveraging advances in machine learning are highly accurate and reliable, capable of detecting and classifying thousands of objects. Classifiers can be incorporated into a SLAM pipeline to add semantic information to a scene. Frequently, this semantic information is conducted for each frame of the image, but semantic labeling is not persistent over time. Another element of SLAM is loop closure, which determines previously visited locations in the trajectory generated during the mapping and localization process. Identifying these loops in the trajectory is challenging due to the changes of viewing angles, illumination and environmental dynamics etc. In this dissertation, we introduce two novel approaches to address these problems. First, we present a non-parametric statistical approach to perform association/matching between detected objects over consecutive image frames. An unsupervised clustering method then localizes these associated classified objects in accrued map. We test our approach on several public data sets and our own data-set, which shows promising results in terms of objects correctly associated from frame to frame and localization of the existing objects in the map. We also have tested our algorithm on three data sets in our lab environment using tag markers to demonstrate the accuracy of classified object localization process. Second, we present a solution to the loop closure problem based on deep reinforcement learning. The framework is a reward-driven optimization process to learn loop closure detection. We demonstrate the framework in a simulated grid environment that generates data for a learning agent. The agent learns from data to perform loop closure in different environments. We demonstrate our results based on the rewards from the simulation and correct loop closure detection, and show that our outcomes are comparable to traditional loop-closure methods.