Object Identification in Maritime Environments for ASV Path Planner

In this paper, we present real-time object identification algorithm, that can be implemented in the motion planning in Autonomous Surface Vehicles (ASV) at sea based on Faster Regional Convolutional Neural Networks (RCNN) model. These vehicles' control algorithms and path planning systems depend heavily on the surrounding area and the objects in sight, thus perception capability in near field is crucial for ASV's safety. Such systems can also be used for security applications, maritime traffic control and port management. The dataset we used consists of thirty six videos, captured in HD resolution, moreover, we shot another 3 videos for visual testing. We trained a Faster RCNN model in addition to a YOLOv3 model and compared the results. The first framework gave the best results while being the slowest, while the second one was fast and slightly less accurate. We showed that such a system can be implemented on a real-time camera feed and can be used as a part of the developments of an ASV as integral part of path planner for ASV.