Object detection in maritime environments is particularly challenging due to dynamic conditions, variable object scales, and the computational constraints of edge devices. This work proposes a modified YOLOv8-based architecture optimized for small object detection. Our approach integrates Dynamic Convolution (DyConv), Efficient Channel Attention (ECA) in C2f modules, an additional P2/4-scale detection head, multiscale fusion with spatial and channel attention, and an enhanced SPPF module. Additionally, the Slicing Aided Hyper Inference (SAHI) technique is incorporated to boost detection sensitivity during inference. Experiments on the Datasense@CRAS dataset show consistent gains: YOLOv8n (ours) improves mAP50 and mAP50:95 by +1.19% and +2.09%, while YOLOv8l (ours) achieves +1.88% and +3.60%, respectively. In particular, small object categories such as Small boat and Sail boat achieve improvements of up to +7.6% and +10.5% in mAP50 and mAP50:95. The integration of SAHI more than doubled the number of objects detected in complex scenarios. Finally, the feasibility of implementation on edge devices such as NVIDIA Jetson AGX Orin is validated, where our YOLOv8l (ours) model in FP16 offers the best balance between accuracy, real-time performance, and energy efficiency. Compared to other lightweight detectors, our model demonstrates superior accuracy while maintaining a similar computational cost and strong cross-domain generalization. These results confirm the effectiveness of the proposed architecture for maritime applications, which addresses challenges such as variable lighting, occlusions, and onboard processing constraints in USVs and autonomous platforms.
Mela et al. (Fri,) studied this question.