Um sistema de baixo custo para localização utilizando sensores posicionais e estereoscopia visual

• Autor(a) principal: Speroni, Eduardo Arrial
• Data de Publicação: 2016
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Manancial - Repositório Digital da UFSM
• Texto Completo: http://repositorio.ufsm.br/handle/1/12037
• topic: Robótica; Visão computacional; Estereoscopia; SLAM; Sensores; Android; Robotics; Computer vision; Stereocopy; Sensors; Android; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

One of the problems in robotics is called Simultaneous Location and Mapping (SLAM), and lies in the necessity of a robot to localize itself on the environment while simultaneously mapping it. The use of stereoscopic systems is one approach to solve this problem. Theses systems are composed by high cost cameras synchronized via hardware, while low cost cameras are more restrict to applications with low or no movement. This research proposes a low cost system by using stereoscopy with a low baseline and low horizontal field of view cameras, synchronizing them via software, along with a filter based on the density of the disparity map of the captured images, with the intent to discard badly rectified frames, which implies desynchronization. Additionally, an Android app capable of obtaining and transmitting sensory data from a smartphone, like GPS and orientation, was developed, reducing the cost and increasing the system’s accessibility. From these data, calibration and processing datasets were generated, so they could be analyzed afterward. The combination of visual odometry and the smartphone’s sensory data contained in the datasets resulted in a system capable of obtaining its localization without previous knowledge of the environment with a similar error to the ones obtained by well established high cost techniques. However, the GPS data was imprecise in low speed scenarios, while the high electromagnetic interference and the low amount of lateral points of reference harmed the device’s orientation data and the visual odometry calculation in the high speed scenario. The system isn’t capable of real time processing, given the need to analyze every frame so they can be filtered, discarding about 60% of them. It was demonstrated that the proposed low cost system was capable of keeping a low error in return of a high processing time, potentially reducing the cost and increasing the accessibility of VSLAM applications. Due to the system’s modularity, it’s possible to replace its components without many implementation changes, allowing the use of better precision devices in future work.