Vision-based detection and tracking of marine vehicles using convolutional neural networks

Abstract

Marine vehicle tracking is gaining relevance nowadays in the research community. This has to do with the fact that current development in technology has propitiated a tendency towards autonomous navigation, with all the potential advantages this implies. A solution to this problem will be the focus of this dissertation. Based on previous work in the field, a tracking system using computer vision with a monocular camera arrangement is proposed. In this system, some possibilities of improvement have been identified and implemented. They can be grouped in two blocks: those related to perception and those related to the model of tracking. In the former, modern techniques based on deep-learning have been applied, whereas in the latter, mathematical models have been developed aiming to improve and complement the existing system. The results of the proposed methods show improvements in the robustness and accuracy of the tracking of marine vehicles in the tested scenarios.