A Robust Dual-Modality Perception Framework for Unmanned Surface Vessels: Registration and Decision Fusion

Accurate detection of surface objects is pivotal for the safe autonomy and operation of Unmanned Surface Vessels (USVs). However, the drastic scale variations and clutter interference found in complex marine environments often compromise the reliability of single-sensor systems. To address this, this paper proposes a dual-modality perception framework integrating visible and infrared (IR) sensors. First, this paper introduces a novel maritime image registration method leveraging central target matching and enhance edge detection to ensure robust spatial alignment between modalities. Second, to mitigate the scarcity of public multi-modal maritime data, this paper presents a newly constructed dataset comprising 4,008 pairs of aligned images across diverse operational scenarios. Finally, a decision-level fusion architecture based on YOLOv10-S is developed. Utilizing a CIoU-based association strategy combined with adaptive thresholding, the method performs intelligent screening of simultaneous detections. Experimental evaluations across various maritime scenarios demonstrate that the proposed framework significantly outperforms single modality baselines, improving mAP50-95 by 14.91% and 9.21% compared to visible-only and IR-only methods, respectively. This enhanced precision significantly reduces detection errors, providing essential reliability for autonomous maritime navigation.