Project i-RoCS: dirt detection system based on computer vision

• Autor(a) principal: Carapinha, Rui Filipe Santos
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/30482
• topic: Template matching; Robotics; ROS; Computer vision; Dirt detection; OpenCV; Camera calibration; Saliency; Deep learning

The ultimate goal of the i-RoCs project is to provide an e cient automatic robotic solution to clean industrial oors. The solution will integrate stateof- art computer vision algorithms for the navigation of the robot and for the monitoring of the cleaning process. Industrial oor cleaning is one of the most important tasks for the security of the personnel in a factory. In the worst case, a damaged/slippery oor can lead to the most various accidents. This is the main reason why the most advanced technologies should be involved in this area. In this thesis we pretend to give a step towards that goal. Digital cameras with the proper use and the proper algorithms can be one of the most rich sensors that can be used in the industrial environment due to the information they can capture. This information is a conversion of the real world into digital information that can be further processed. From this information, low-level computer vision algorithms can detect a lot of features from an image such as colors, lines, blobs, contours, edges, patterns, among others. In this thesis, we give an introduction of state-of-art technology to the cleaning task in a factory. For that purpose, we present a study about the implementation of cameras and digital image processing to detect dirt in industrial oors. We propose a method for automatic calibration of the camera parameters to tackle the di cult environment that can be found inside factories in terms of the light conditions. We developed algorithms for extraction of low-level characteristics to be used in the detection of dirt that obtained promising results in terms of detection results. However, they are not satisfactory in terms of performance if we consider them to be applied in real time on a mobile robot. The last step was the implementation of Deep Learning, one of the most promising technologies of the past few years used in image processing. This proposed solution is a segmentation network followed by a regression network. The segmentation will classify the several types of patterns existing on the ground and the regression will output the level of dirtiness of each area.