Enhanced Object Detection in Maritime Vision: Modified YOLOv8 Architecture and SAHI-Based Inference

Object detection in maritime environments poses particular challenges due to dynamic environmental conditions, variability in object scale, and computational constraints of edge devices. In this study, a modified YOLOv8-based architecture optimized for small object detection is proposed. The proposal integrates Dynamic Convolution (DyConv) layers, Efficient Channel Attention (ECA) in the C2f modules, an additional P2/4-scale detection head, a multiscale fusion with spatial and channel attention, and an enhanced version of the SPPF module. In addition, the Slicing Aided Hyper Inference (SAHI) technique is incorporated to improve detection sensitivity during inference. Experiments performed on the Datasense@CRAS dataset demonstrate significant improvements, especially our proposal YOLOv8n-modified architecture obtained an average increase of +1.19 \% in mAP50 and +2.09 \% in mAP50:95, while the YOLOv8l-Modified variant achieved improvements of +1.88 \% and +3.60 \%, respectively. These gains are especially notable in small object classes, such as Small and Sail Boats, where increases of up to +7.6 \% and +10.5 \% were recorded in mAP50 and mAP50:95. The integration of SAHI increased by more than $2\times$ the number of objects detected in complex scenarios, evidencing its usefulness for maritime environments. Finally, the feasibility of implementation on edge devices such as NVIDIA Jetson AGX Orin is validated, where our model YOLOv8l-Modified in FP16 offers the best balance between accuracy, real-time performance, and energy efficiency. This shows that the proposed architecture is particularly effective in detecting small objects, increasing the number of objects detected and the confidence level to improve autonomous navigation. It addresses the unique challenges of maritime environments, including variable lighting, occlusion, and the need for efficient onboard processing in USVs and autonomous platforms.