Unmanned surface vehicles (USVs) are distinguished by their intelligence, compactness, and absence of human casualties, making them a vital component of the maritime industry. The implementation of vision-based algorithms for sea surface target detection can enhance the autonomous perceptual abilities of USVs. In the present study, a sea surface target detection algorithm was proposed that fulfils the requirements of USVs marine environment sensing and sea area monitoring. Sea surface target detection faces unique challenges, such as highly variable target sizes and a complex and changing marine environments. The current state-of-the-art You Only Look Once (YOLO) model was selected as the baseline target detection model. Firstly, to improve the network’s ability to extract features of different sizes, a Cross Stage Partial Lightweight Spatial Pyramid Pooling-Fast (CSPLSPPF) structure was proposed. Additionally, for achieving the advantages of multiple feature maps to complement each other and output more judgmental feature maps, Path Aggregation Network Powerful (PANP) was proposed to more rationally fuse features of feature maps with different resolutions. Finally, lightweight convolution with fused attention(LCFA) was proposed to enable the network to selectively focus on crucial spatial and channel information while simultaneously reducing the model’s parameter count. Experiments were conducted on a self-made Ocean Buoys dataset and the open-source Seaships dataset. The results showed that the proposed method could efficiently and accurately detect objects such as ships and buoys in marine environments, which was of significant value for USVs to achieve intelligent environment perception.

无人水面车辆 (USV) 以其智能、紧凑和无人员伤亡等特点而著称，使其成为海事行业的重要组成部分。实施基于视觉的海面目标检测算法可以增强USV的自主感知能力。本研究提出了一种满足USV海洋环境感知和海域监测需求的海面目标检测算法。海面目标检测面临着独特的挑战，例如高度可变的目标尺寸和复杂多变的海洋环境。选择当前最先进的 You Only Look Once (YOLO) 模型作为基线目标检测模型。首先，为了提高网络提取不同尺寸特征的能力，提出了一种跨阶段部分轻量级空间金字塔池化快速（CSPLSPPF）结构。此外，为了实现多个特征图的优势互补并输出更多的判断特征图，提出了强大的路径聚合网络（PANP）来更合理地融合不同分辨率特征图的特征。最后，提出了具有融合注意力的轻量级卷积（LCFA），使网络能够选择性地关注关键的空间和通道信息，同时减少模型的参数数量。在自制的Ocean Buoys数据集和开源的Seaships数据集上进行了实验。结果表明，该方法能够高效、准确地检测海洋环境中的船舶、浮标等物体，对于USV实现智能环境感知具有重要价值。