Visual Place Recognition as a solution for the Kidnapped Robot Problem

The kidnapped robot is a well-known problem in mobile robotics; it is a special case of the global localization problem, which arises when the robot is physically moved to another location without its knowledge, or gets lost in the environment. Kidnapped Robot Problem is usually studied and tested to evaluate the performance of localization algorithms since none of them can guarantee not to fail. In this thesis, we propose a solution in order to solve the kidnapped robot problem in dynamic and similar environment based on Visual Place Recognition (VPR), Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to initialise the position of the robot and a particle filter to correct the pose. The VPR algorithm is based on using a Convolutional Neural Network (CNN) to extract global features from images and recognise the place where the robot is located, selecting a set of best poses using a similarity measurement algorithm. The DBSCAN is used to filter out isolated and small clusters of selected poses and initialise the position of the robot with the centroid of the largest cluster. A particle filter is then applied to correct the robot's pose. The proposed solution is fully implemented in Robot Operating System (ROS) using Python as the main programming language. It is first tested in a simulated environment in the Gazebo platform and then in a real scenario at PIC4SeR (PoliTo Interdepartmental Centre for Service Robotics). The results show that the proposed solution is able to correctly relocalise the robot after the kidnapping event.