Obstacle Detection in Real and Synthetic Harbour Scenarios

In the last decade, the autonomous vehicle has been investigated by both academia and industry. One of the open research topics is obstacle detection and avoidance in real-time; for such a challenge, the most used approaches are based on deep learning, especially in the automotive sector. Usually, trained neural networks are used to detect the obstacles by receiving the point clouds from LiDAR as input data. However, this approach is currently not feasible in the marine sector as there are no large datasets of LiDAR point clouds and relatively few RGB images available to train networks. For such a reason, this paper aims to present the first step for the design of an alternative approach that integrates unsupervised and supervised learning algorithms for the detection and tracking of both fixed and moving obstacles. A virtual scenario that can be customized according to the users’ purpose has been developed and used to collect data by emulating the LiDAR and camera behaviour. Moreover, the preliminary on-field LiDAR recording is presented and processed. The unsupervised clustering algorithms have been tested, and the pros and cons of the different clustering approaches are shown.