摘  要

随着“冰上丝绸之路”的不断发展，北极航道中季节性通航的商业船舶数量逐年增多，进一步深化了世界经济和贸易格局。但是由于通航环境复杂、沿岸基础设施不完善、船舶技术条件限制等问题，导致船-冰碰撞事故时有发生，为北极航道的安全通航带来重重阻碍。尽管通航船只均配备有雷达、卫星通讯等设备，但仍不能及时有效地获取到航行水域的海冰信息，为船舶驾驶员提供避碰决策参考，避免船-冰碰撞事故发生。针对上述问题本文从瞭望为主的冰区航行方式出发，采用船基海冰图像采集，在线图像拼接和图像语义分割技术，研究基于船基海冰图像处理的船-冰避碰决策方法。本文的工作主要研究工作如下：

（1）基于拼接缝驱动的图像拼接算法。针对实际航行过程中船载单目相机得到的船基海冰图像视场范围不足的问题，采用多组不同视角相机组合拍摄的方式，获取更大范围的冰场信息。将所摄自然图像之间存在的重合区域，通过尺度不变特征变换（Scale Invariant Feature Transform，SIFT）特征提取和随机抽样一致性（Random Sample Consensus，RANSAC）误点筛除算法进行图像配准，在此基础上进一步提出改进拼接缝驱动的梯度域图像融合算法，在对齐的重叠区域中寻找一条最佳拼接缝，以消除拼接图像中的错位和重影，得到主观视觉评估和客观质量评价更优的大视场合成图像。

（2）基于深度学习的海冰图像语义分割模型研究。针对以瞭望为主的冰区船舶航行过程中船员在海冰观测时存在主观疏忽的问题，提出一种可训练的图像语义分割海冰识别与分类模型，实现船基海冰图像中海冰对象的快速检测。首先，该模型通过制定一套海冰数据集制作标准，分别制作海冰图像语义识别和分类数据集，以适用于不同任务的端到端深度学习网络参数训练。其次，模型所采用的IceNet网络结构以SegNet网络骨架为基础，在不增加大量模型参数的前提下，融合残差注意力模块，解决因网络深度带来的训练难和网络上下文信息不关联问题，最终分别实现大视场船基海冰图像中海冰目标的识别与分类。

（3）多阶段划分的船-冰避碰决策方法。针对船基海冰语义图像与船舶平面的视角差异问题，根据船载相机倾斜成像条件下的几何关系，将语义图像中的海冰像素经逆透视变换（Inverse Perspective Mapping，IPM）后映射到船舶所在的世界空间坐标系中，再进行栅格化处理统一船-冰时空坐标系，进一步实现海冰参数分析与船-冰碰撞物理模型构建。在此基础上，结合船舶运动概率模型，采用基于KUNZI模型的船-冰碰撞危险度分析方法，设定船-冰碰撞风险概率阈值实现对危险海冰的动态预警，再进一步依据碰撞危险程度进行碰撞态势分析制定有效的避碰策略。最后，运用上述方法模拟两种典型船-冰碰撞态势下的碰撞危险程度判断和避碰决策流程。

关键词：船基海冰图像处理；图像拼接；图像语义分割；船-冰碰撞危险度；避碰决策

Abstract

With the continuous development of the Ice Silk Road, the number of seasonally navigable commercial ships in the Arctic waterways has increased year by year, which has further deepened the world economy and trade pattern. However, due to the complicated navigation environment, imperfect coastal infrastructure, and limited ship technical conditions, ship-ice collision accidents have occurred from time to time, which has brought numerous obstacles to the safe navigation of the Arctic waterways. Although navigable vessels are equipped with radar, satellite communications and other equipment, they still cannot effectively obtain sea ice information in the navigating waters to remind ship pilots to pay attention to high-risk sea ice areas and provide collision avoidance decisions to avoid ship-ice collision accidents. Aiming at the above problems, this article starts from the observation-based navigation method in the ice area, makes full use of ship-based sea ice images, and uses image stitching and image semantic segmentation techniques to deeply study ship-ice collision avoidance decision-making methods based on ship-based sea ice image analysis. The main research work of this thesis is as follows:

(1) Image stitching algorithm based on stitch driving. Aiming at the problem of insufficient field of view of the ship-based sea ice image obtained by the ship-borne monocular camera during actual sailing, multiple sets of camera equipment with different viewing angles are set to shoot to obtain a larger range of ice rink information. The coincident areas between the natural images taken are used for image registration based on the SIFT feature extraction method and the RANSAC error point filtering algorithm, and an improved stitch-driven gradient domain image fusion algorithm is further proposed to find the best one in the aligned overlapping areas. Better stitching seams to eliminate the problems of misalignment and ghosting in stitched images, and to obtain composite images with better subjective visual evaluation and objective quality evaluation.

(2) Sea ice image semantic segmentation recognition and classification model based on deep learning. Aiming at the problem of subjective negligence in the observation of sea ice by the crew during the navigation of ships in the icy water where the observation is the mainstay, a trainable image semantic segmentation sea ice recognition and classification model is proposed to realize rapid detection of sea ice objects in ship-based sea ice images. First, this model develops a set of sea ice data set production standards to produce sea ice image semantic recognition and classification data sets, which are suitable for end-to-end deep learning network parameter training for different tasks. Secondly, the IceNet network structure adopted by the model is based on the SegNet network skeleton. Without increasing a large number of model parameters, the residual attention module is integrated to solve the problem of difficulty in training due to the depth of the network and the non-correlation of network context information. Finally, the recognition and classification of sea ice targets in the large field of view ship-based sea ice image are realized respectively.

(3) Multi-stage decision-making method for ship-ice collision avoidance. Aiming at the problem of different perspectives between ship-based sea ice semantic image and ship plane, the sea ice pixels in semantic image are mapped to the ship's world space coordinate system after inverse perspective transformation according to the geometric relationship of ship-based camera in oblique imaging condition. Then the rasterization process is performed to unify the ship-ice space-time coordinate system, and further realize the analysis of sea ice parameters and the construction of the ship-ice collision physical model. On this basis, combined with the probability model of ship's own motion state, the risk analysis method of ship ice collision based on kunzi model is adopted, and the risk threshold of ship ice collision is set to realize the dynamic early warning of dangerous sea ice, and then the collision state is analyzed according to the collision risk degree, and the effective collision avoidance strategy is formulated. Finally, the above method is used to simulate the collision risk judgment and collision avoidance decision-making process under two typical ship ice collision situations.

Key words: ship-based sea ice image processing, image stitching, image semantic segmentation, ship-ice collision risk, collision avoidance decision-making