Rethinking YOLOv5 with Feature Correlations for Unmanned Surface Vehicles

Reliable and robust obstacle detection is affected by the adverse environmental conditions and objects. Traditional object detection methods cannot cover the complex scenario especially in maritime environment thus hinder the progress of maritime robotics. Since the serials of YOLO algorithms have been proven superior in object detection, it becomes feasible to develop an real-time and accurate obstacle detection for USV. However, the complex maritime environment poses two intractable challenges for obstacle detection methods: 1) obstacles with various shapes, sizes, and types require detection algorithms learning features with large receptive field, and 2) multi-scale features should mutually communicate to improve the robustness of detection models. With these motivations, we propose an Feature Correlation YOLOv5 (FC-YOLO) for USV obstacle detection. To address the issues, we introduce two modifications to YOLOv5: 1) exploring intra-relationship within one feature maps to capture global information, and 2) modelling inter-relationship between different feature maps to highlight feature correlations. Extensive experiments demonstrate that our proposed method achieve average mAP of $$98.00\%$$ and $$83.70\%$$ on SeaShips and WSODD datasets, respectively. These results demonstrate the effectiveness of FC-YOLO on obstacle detection for USV.