Semantic Segmentation and Regions of Interest for Obstacles Detection and Avoidance in Autonomous Surface Vessels

• Rodrigo Aguiar UFF
• Luciano Bertini UNIFEI
• Alessandro Copetti UFF
• Esteban Walter Gonzalez Clua UFF
• Luiz Marcos Garcia Gonçalves UFRN
• Leonard Barreto Moreira UFF

Resumo

This paper proposes methodology for identifying and circumventing obstacles during the navigation of unmanned surface vessels (USVs). Our proposal leverages real-time visual data captured by onboard cameras to detect and avoid potential barriers. We start our pipeline processing with a convolutional neural network to provide semantic segmentation, aiming to reduce the number of features in the original images. Following we use a technique that we developed to determine the horizon line, which is identified in order to separate specific regions of interest in the image, thus reducing the computational cost. Through these results and the defined regions of interest, we apply traditional computer vision methods to detect obstacles that may be on a collision route with the vessel, as well as suggest a safe direction for the vessel to follow. We evaluate this work through a dataset of videos collected with a real USV operating under different lighting conditions, types of obstacles, and maritime traffic. Data augmentation is performed to provide a better training for the network. The experiments demonstrate the effectiveness of the proposed methodology, with accuracy which is above 90%.