复杂航行场景下的一体化水面目标检测与边界线分割算法

An integrated algorithm for water-surface object detection and boundary line segmentation in complex navigation scenarios

摘要:

为克服现有智能船舶目标检测算法在复杂航行场景下对连续性障碍物感知的局限性,并充分利用水面边界与目标在空间分布上的约束关系以提升检测精度,提出了一种检测与分割一体化的感知算法。首先,通过集成水面分割解码器构建多任务学习框架,使检测模型具备可航行范围感知能力；进而,将分割结果表征的水面边界线信息嵌入至检测解码器,显式引导模型更加聚焦目标分布密集的区域；此外,对骨干网络进行重参数化改进,以强化两个解码器对各自关键信息的建模。在多个真实水域数据集上的实验结果表明,该方法仅需较少的边界线标注成本,即可高效实现检测与分割一体化,并显著提升了水面目标检测精度,在多类场景中均优于现有主流方法,同时其水面分割的IoU指标达到99.43%。本研究通过利用场景结构先验来强化多任务协同学习,为提升复杂开放水域环境下智能船舶的环境感知可靠性提供了新的技术途径。

Abstract:

To overcome the limitations of existing intelligent ship object detection algorithms in perceiving continuous obstacles in complex navigation scenarios, and to fully leverage the spatial distribution constraints between water boundaries and objects to enhance detection accuracy, an integrated perception algorithm for detection and segmentation is proposed. Firstly, a multi-task learning framework is constructed by integrating a water segmentation decoder, which endows the detection model with navigable area perception capabilities. Subsequently, the water boundary line information, represented by the segmentation results, is embedded into the detection decoder to explicitly guide the model to focus more on regions with dense object distributions. Furthermore, the backbone network is improved with re-parameterization to strengthen the modeling of key information for both decoders. Experimental results on multiple real-world aquatic datasets show that the proposed method can efficiently achieve integrated detection and segmentation with minimal boundary line annotation costs. It significantly improves the accuracy of water object detection, outperforming existing mainstream methods in various scenarios, while its water segmentation IoU metric reaches 99.43%. This research provides a new technical pathway for enhancing the environmental perception reliability of intelligent ships in complex open-water environments by utilizing scene structure priors to reinforce synergistic multi-task learning.