A Machine Learning Approach for Predicting Docking-Based Structures

• Autor(a) principal: Maria Inês Rodrigues Pereira
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/132609
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

The maritime industry has only recently taken its first steps towards automation and Artificial Intelligence (AI). The development of Autonomous Surface Vehicles (ASVs) falls under this scope, uncovering the possibility of greatly reducing the role of human intervention in this industry and the inherent consequences of human error. Research concerning the operation of this type of vehicles has seen an upward trend in recent years, although its full-scale application still encounters various limitations, such as the need to ensure safety and develop the required technology. The docking and undocking processes are among the most challenging tasks for a maritime vehicle. In fact, most maritime accidents occur in sea ports and harbours. Such evidence, along with the scarceness of scientific endeavours in this topic, make the docking approach of an ASV an alluring matter to delve into. With this intention in mind, this dissertation aims to take one step further towards enabling a vessel to dock autonomously, by providing a perception tool necessary for this maneuver. The developed work comprises a Deep Learning (DL) network able to detect the existence of a docking platform in the vehicle's surrounding environment and classify the type of the structure. As data concerning this context is not widely available, it was necessary to conduct a data acquisition process in the 3D simulator Gazebo, in which models of five different types of docks were placed in a simulated maritime environment and data from several sensors, such as a LiDAR and a stereo camera, was captured. The gathered data was used for the training of the cascaded classifier, which was composed of two networks: a first one to ascertain whether there is a dock in the vicinity of the vehicle and a second one to categorize the structure. This dissertation also proposes a mechanism to perform an occupation analysis of the dock, based on a template matching approach. The developed work was tested on the gathered data and in a dynamic setup in the aforementioned simulator, taking into account different noise conditions to better replicate an authentic maritime environment. The detection model achieved an accuracy of 96.44% in optimal conditions and an average of 90.43% considering light to very severe noise conditions, with a deviation of 2.8% for the worst case. The categorizing model obtained a maximum accuracy of 86.70% and an average accuracy of 80.90% for noise tests. The occupation analysis algorithm is able to return the number and coordinates of the vacant spots of the dock. Both the cascade classifier and the occupation tool were developed through an approach that allows an easy adaptation to other types of docks, thus differentiating it from other works on the topic.