Creation and maintenance of visual incremental maps and hierarchical localization
Dr. Vicente Román Erades

Over the last few years, the presence of the mobile robotics has considerably increased in a wide variety of environments. It is common to find robots that carry out repetitive and specific applications and also, they can be used for working at dangerous environments and to perform precise tasks. These robots can be found in a variety of social environments, such as industry, household, educational and health scenarios. For that reason, they need a specific and continuous research and improvement work. Specifically, autonomous mobile robots require a very precise technology to perform tasks without human assistance.
 
To perform tasks autonomously, the robots must be able to navigate in an unknown environment. For that reason, the autonomous mobile robots must be able to address the mapping and localization tasks: they must create a model of the environment and estimate their position and orientation.
 
This PhD thesis proposes and analyses different methods to carry out the map creation and the localization tasks in indoor environments. To address these tasks only visual information is used, specifically, omnidirectional images, with a 360º field of view. Throughout the chapters of this document solutions for autonomous navigation tasks are proposed, they are solved using transformations in the images captured by a vision system mounted on the robot.
 
Firstly, the thesis focuses on the study of the global appearance descriptors in the localization task. The global appearance descriptors are algorithms that transform an image globally, into a unique vector. In these works, a deep comparative study is performed. In the experiments different global appearance descriptors are used along with omnidirectional images and the results are compared. The main goal is to obtain an optimized algorithm to estimate the robot position and orientation in real indoor environments. The experiments take place with real conditions, so some visual changes in the scenes can occur, such as camera defects, furniture or people movements and changes in the lighting conditions. The computational cost is also studied; the idea is that the robot has to localize the robot in an accurate mode, but also, it has to be fast enough.
 
Additionally, a second application, whose goal is to carry out an incremental mapping in indoor environments, is presented. This application uses the best global appearance descriptors used in the localization task, but this time they are constructed with the purpose of solving the mapping problem using an incremental clustering technique. The application clusters a batch of images that are visually similar; every group of images or cluster is expected to identify a zone of the environment. The shape and size of the cluster can vary while the robot is visiting the different rooms. Nowadays. different algorithms can be used to obtain the clusters, but all these solutions usually work properly when they work ‘offline’, starting from the whole set of data to cluster. The main idea of this study is to obtain the map incrementally while the robot explores the new environment.
 
Carrying out the mapping incrementally while the robot is still visiting the area is very interesting since having the map separated into nodes with relationships of similitude between them can be used subsequently for the hierarchical localization tasks, and also, to recognize environments already visited in the model.
 
Finally, this PhD thesis includes an analysis of deep learning techniques for localization tasks. Particularly, siamese networks have been studied. Siamese networks are based on classic convolutional networks, but they permit evaluating two images simultaneously. These networks output a similarity value between the input images, and that information can be used for the localization tasks. Throughout this work the technique is presented, the possible architectures are analysed and the results after the experiments are shown and compared. Using the siamese networks, the localization in real operation conditions and environments is solved, focusing on improving the performance against illumination changes on the scene. During the experiments the room retrieval problem, the hierarchical localization and the absolute localization have been solved.