[pt] Robôs móveis são cada dia mais importantes na sociedade moderna, realizando tarefas consideradas tediosas ou muito repetitivas para humanos, como limpeza ou patrulhamento. A maioria dessas tarefas requer um certo nível de autonomia do robô. Para que o robô seja considerado autônomo, ele precisa de um mapa do ambiente, e de sua posição e orientação nesse mapa. O problema de localização e mapeamento simultâneos (SLAM) é a tarefa de estimar tanto o mapa quanto a posição e orientação simultaneamente, usando somente informações dos sensores, sem ajuda externa. O problema de SLAM visual consiste na tarefa de realizar SLAM usando somente câmeras para o sensoriamento. A maior vantagem de usar câmeras é a possibilidade de resolver problemas de visão computacional que provêm informações de alto nível sobre a cena, como detecção de objetos. Porém a maioria dos sistemas de SLAM visual assume um ambiente estático, o que impõe limitações para a sua aplicabilidade em cenários reais. Esta tese apresenta soluções para o problema de SLAM visual em ambientes dinâmicos e em mudança. Especificamente, a tese propõe um método para ambientes com multidões, junto com um detector de pessoas customizado baseado em aprendizado profundo. Além disso, também é proposto um método de SLAM visual para ambientes altamente dinâmicos contendo objetos em movimento, combinando um rastreador de objetos robusto com um algoritmo de filtragem de pontos. Além disso, esta tese propõe um método de SLAM visual para ambientes em mudança, isto é, em cenas onde os objetos podem mudar de lugar após o robô já os ter mapeado. Todos os métodos propostos são testados com dados públicos e experimentos, e comparados com diversos métodos da literatura, alcançando um bom desempenho em tempo real. / [en] Mobile robots have become increasingly important in modern society, as they can perform tasks that are tedious or too repetitive for humans, such as cleaning and patrolling. Most of these tasks require a certain level of autonomy of the robot. To be fully autonomous and perform navigation, the robot needs a map of the environment and its pose within this map. The Simultaneous Localization and Mapping (SLAM) problem is the task of estimating both map and localization, simultaneously, only using sensor measurements. The visual SLAM problem is the task of performing SLAM only using cameras for sensing. The main advantage of using cameras is the possibility of solving computer vision problems that provide high-level information about the scene, such as object detection. However, most visual SLAM systems assume a static environment, which imposes a limitation on their applicability in real-world scenarios. This thesis presents solutions to the visual SLAM problem in dynamic and changing environments. A custom deep learning-based people detector allows our solution to deal with crowded environments. Also, a combination of a robust object tracker and a filtering algorithm enables our visual SLAM system to perform well in highly dynamic environments containing moving objects. Furthermore, this thesis proposes a visual SLAM method for changing environments, i.e., in scenes where the objects are moved after the robot has already mapped them. All proposed methods are tested in datasets and experiments and compared with several state-of-the-art methods, achieving high accuracy in real time.

No description available.

An autonomous robot that acts in a goal-directed fashion requires a world model of the elements that are relevant to the robot's task. In real-world, dynamic environments, the world model has to be created and continually updated from uncertain sensor data. The symbols used in plan-based robot control have to be anchored to detected objects. Furthermore, robot perception is not only a bottom-up and passive process: Knowledge about the composition of compound objects can be used to recognize larger-scale structures from their parts. Knowledge about the spatial context of an object and about common relations to other objects can be exploited to improve the quality of the world model and can inform an active search for objects that are missing from the world model. This thesis makes several contributions to address these challenges: First, a model-based semantic mapping system is presented that recognizes larger-scale structures like furniture based on semantic descriptions in an ontology. Second, a context-aware anchoring process is presented that creates and maintains the links between object symbols and the sensor data corresponding to those objects while exploiting the geometric context of objects. Third, an active perception system is presented that actively searches for a required object while being guided by the robot's knowledge about the environment.

Methodology and issues of data transformation and its information value estimation during the integration of the heterogenous information sources PhDr. Ivan BARTOŠ Abstract This study focuses mainly on the data and information transformation issue. This topic is currently critical in several scientific and commercial areas. Information value, information quality and the quality of the source data differs between the various systems. This is not only due to the different topologies of the information sources but also because of its different understanding and a manner of storing the information describing the entity of the enterprise. Such information systems, respectively database systems in the scope of the thesis, could perform well as the stand alone systems. The issue appears in the moment when such heterogeneous systems are required to be integrated and the information shall be migrated between each other. The thesis is logically divided into four major parts based on these issues. The first part describes the methods that can be used to classify the data quality of the source system (the one to be integrated) from which the information can be extracted. Based on assumption of the common lack of project and system documentation hereby introduced methods can be used for such qualification even when the...