In recent years, the maritime sector has been witnessing a growing interest in autonomous navigation due to its advantages in safety. Within this framework, it is challenging to have available an automatic collision avoidance system based on data coming from the onboard sensors. For instance, LiDAR (Light Detection and Ranging) sensors are often employed to obtain a virtualized model of the surrounding environment in order to obtain an increase of situational awareness. Deep Learning approaches are widely used in the automotive sector for obstacle detection, particularly, trained Convolutional Neural Networks. Still, they imply a training phase based on a large dataset of pre-labelled LiDAR scans. Unfortunately, an extensive training dataset is not yet available for the marine environment. For such a reason, this paper presents in detail a low-computational alternative procedure based on Unsupervised Learning which overcomes the lack of a training dataset; each step is investigated, from the data sampling up to the multi-target tracking of the detected obstacles. In particular, the proposed object detection framework has been tested by means of an extensive on-field data collection campaign carried out during sea trials by equipping a ship with a LiDAR. Furthermore, Euclidean distance-based clustering algorithm and a bounding-box construction method based on Principal Component Analysis have been adopted. Moreover, a specific tracking system with no prediction filters is proposed to fulfil the strict time constraint for fast reactions in complex scenarios where several objects need to be tracked. The algorithm has been tested versus the well-known Global Nearest Neighbour tracker; such a comparison includes the computational cost and the results’ accuracy. The whole approach has been tested on such a challenging and dynamic marine scenario, and the results obtained are presented and discussed. Such and outcome shows that the proposed approach can detect and track multiple objects with reasonable accuracy; moreover, the outputs are provided near real-time. To conclude, the pros, the weaknesses, and future developments are reported.