Obstacle Detection and Segmentation in Unmanned Surface Vessel Using Deep Learning

The goal is to improve the navigation safety of a semi-autonomous aquatic robot using computer vision and deep learning techniques. Three main approaches were evaluated: object detection with YOLOv8, instance segmentation with YOLOv8-Seg, and semantic segmentation with WASR-T, using images captured in real-life aquatic conditions. Two datasets were used for the experiments: a custom dataset for obstacle detection and segmentation (logs, boats, and plastics) and MaSTr1325 for semantic segmentation (sky, sea, and obstacles). The results show that YOLOv8 achieved the best balance between detection accuracy and speed, achieving a mAP of 88% on the custom dataset. On the other hand, instance segmentation with YOLOv8-Seg allowed for the identification and differentiation of individual objects with good accuracy, but with certain limitations in the detection of irregular edges in complex aquatic environments. While WASR-T provided more detailed semantic segmentation of environmental regions, its high computational cost made it less viable for real-time applications. Furthermore, the performance of YOLOv7-Tiny was compared, which showed faster processing but lower accuracy in obstacle detection. These findings highlight the effectiveness of YOLOv8 in aquatic environments and provide a detailed evaluation of the detection approach, facilitating model selection based on the computational and operational requirements of autonomous navigation.