面向电子巡航的内河视频智能分析算法研究

CCTV system, in combination with vessel transportation system, automatic identification system, global position system, weather information system, etc., serves as a key part of "electronic cruise" and an important safety monitoring method in the field of inland waterway marine surveillance. However, because of the related low intelligent level, the use of current CCTV leads to the expenditure of human resource. Besides, being susceptible to weather conditions, current CCTV systems can hardly get clear video information for waterways under foggy navigation state which, according to related statistics, accounts for more than 60 percent in inland waterway. In this project, based on the features of inland waterway images, we will focus on the video defogging and vessel target detection and tracking algorithms of inland waterway CCTV systems, principally as follows: first, an intelligent dehazing algorithm which consists of both image restoration and enhancement in order to get clear inland waterway video surveillance images; second, the disturbance mechanism and subtracting algorithm of waterway shadow, waterway background modeling and foreground area extraction algorithms, as well as the vessel target detection algorithm in complicated background scene; third, the disturbance mechanism of light reflection on the water surface, the features extraction algorithm for vessel targets, and multi-targets tracking based on Tracking-Learning-Detection (TLD) algorithm. Over all, this project is aimed at improving the intelligent level of current CCTV surveillance system, realizing intelligent video dehazing, and realizing intelligent analyzing, detecting, identifying and tracking vessel targets. It will provide theoretical foundation for maritime surveillance, vessel auxiliary driving etc.

内河航道安装的CCTV系统，与船舶交通管理、自动识别、全球定位等系统相结合构成的"电子巡航"，成为当前内河海事监管的一种重要手段。然而目前CCTV系统智能化程度不高，人工参与的工作量大。而且内河有雾天数达全年60%，CCTV监控系统亦无法获得清晰视频，不能发挥"眼"的作用。本项目基于内河图像特征，研究内河CCTV系统的视频去雾与船舶检测、跟踪算法。主要内容包括：1.研究图像复原与增强相结合的去雾方法来获取清晰视频；2.研究河道中阴影干扰机理和消除方法，建立船舶背景模型，提取前景区域，实现复杂背景下船舶目标检测；3.研究水面光照反射干扰机理，提取船舶特征，采用智能计算方法改进TLD(Tracking-Learning-Detection)算法，实现多目标跟踪。本项目旨在提高现有CCTV监控水平，实现视频智能去雾，智能分析、检测、识别与船舶跟踪，为海事监控、船舶辅助驾驶等提供理论基础。

电子巡航是整合船舶交通管理系统、船舶自动识别系统、无线甚高频、闭路电视监控系统等的先进监管手段，为提高其闭路电视监控系统的智能化，项目组基于计算机视觉理论研究了内河视频去雾、运动船舶检测和跟踪的算法。主要内容包括：.1）基于雾图退化模型,分别用图像增强和复原两类方法对内河航运单幅图像去雾算法开展了研究；通过深入分析雾图特征，项目组总结出内河雾天水面景深先验知识，提出一种基于景深估计的内河视频去雾算法，实现时空一致的内河航运视频清晰化。为实现视频智能去雾，项目组还提出了基于图像频谱特征和灰度共生矩特征的雾图自动检测算法。.2）深入分析了内河航道背景特点和船舶前景目标特征，用视觉注意机制和改进的混合高斯背景建模来解决水波纹干扰问题；用改进的ViBe和Subsense算法来解决慢速船舶的漏检问题；用压缩感知理论与背景差分算法结合来解决内河场景清晰度差的问题；由此提出了一种在线学习的运动船舶视频检测系统，实现了复杂环境下内河航运视频中运动船舶的鲁棒检测。.3）从内河航运视频环境中的原始信号属性、随机观测矩阵、低维特征选择等角度对50多种随机投影类跟踪算法进行系统地、全方位地理论分析研究后，提出了一种正交粒子滤波低秩约束随机投影内河船舶跟踪算法来实现对运动船舶的鲁棒跟踪；针对运动船舶尺度的变化、环境变化和干扰，提出了一种跟踪检测协同的内河船舶跟踪算法，实现重点船舶准确、鲁棒、实时的视觉跟踪；结合HOG特征和模糊SVM的方法提取运动船舶的外观模型，进一步提出了基于可变形部件模型的多船舶跟踪方法。.4）开创性地建立了内河航运视频分析库，为同类研究人员开展算法测试、比较和适应性分析提供了标准视频。设计了一个开放式的、用于科学研究的内河航运船舶视觉检测和跟踪软件平台。.本项目的研究成果不仅丰富了机器视觉的应用领域，而且对提高水上交通安全监管智能化，对提高通航安全和效率提供了理论和技术支持。