【作者】 杨名；

【导师】 张豹山； 阮雅端； 陈启美；

【作者基本信息】 南京大学 ， 电子与通信工程（专业学位）， 2018， 硕士

【摘要】 航道运价便宜,大宗物资大都利用航道运输。内河航道的船舶智能视频监控系统,对运输安全、提高运输效率有着重要的意义。而在实际场景中,存在水面倒影、船行波纹等干扰,而基于帧间像素差的传统高斯背景建模检测方法,囿于其原理,难免存在误检、漏检等情况。为此,论文结合船联网国家专项,对此展开专题研究,其具有重要的应用价值与理论意义。论文描述了国内外船舶视频检测的现状,包括阴影检测、运动目标检测等;分析了船舶检测的难点;解析了常用的几种阴影检测算法的机理,并分析比较了各自的优缺点;提出了一种基于超图的随机阴影分区优化算法;分析了 Faster RCNN存在的误检缺陷,描述了对Faster RCNN两个过程改进的细节,给出了融合样本取舍的船舶检测算法;研究了融合多层特征HyperNet网络的算法原理,对比了与Faster RCNN检测到目标船舶的位置信息、检测精度,验证了 HyperNet网络在实际应用中的可行性。针对传统的阴影检测算法中,因船舷边界颜色与阴影颜色相近导致的检测失败,论文中解析了超图分割的算法原理,改进优化了颜色空间的阴影检测算法,给出了基于超图的随机阴影检测分区算法。针对Faster RCNN算法在实际检测中出现的误检问题,论文提出了融合样本取舍的船舶检测算法。该算法解析了CFAFT模型的原理,首先用Fast RCNN分类器对RPN产生的候选区域进行筛选,然后分类阶段在原有分类器的基础上又加入了1个二值分类器以进行更精细的精测。实验效果表明,该算法能够较好地解决误检问题。针对Faster RCNN在实际检测中存在的目标框位置信息精度问题,以及进一步提高Faster RCNN算法的检测精度的需要,论文提出了融合多层特征的船舶检测算法。该算法将分辨率高的浅层、中间层、丰富语义信息的深层特征图的特征聚集并统一到一个空间,在实际交通数据集下,具有很好的位置信息。相对于Faster R-CNN,平均检测精度mAP与召回率分别提高15%和6.2%。综上,论文的创新点和特点在于:●提出了基于超图的随机阴影检测分区算法,去除了运动船舶的阴影,提高了运动目标检测精度;●优化了运动目标检测的两个过程,提出了融合样本取舍的船舶检测算法,解决了 Faster RCNN算法在实际检测中出现的误检问题,提高了运动目标检测与内河航道船舶参数的精度;·提出了融合多层特征的船舶检测算法,融合了浅层、中间层、高层信息,解决了 Faster RCNN在实际检测中存在的目标框位置信息精度问题,提高了目标检测的精度。更多还原

【Abstract】 Fairway freight rates are cheap,and most of the bulk goods use waterway transportation.The intelligent videosurveillance system for ships in inland waterways is of great significance for transportation safety and improving transportation efficiency.In the actual scene,there are interferences such as reflection of the water surface and ripples of the ship.However,the traditional Gaussian background modeling and detection method based on the pixel difference between frames is inconsistent with its principle,and inevitably there are cases of misdetection and missed detection.For this reason,the paper combines the special projects of the shipbuilding network and carries out special research on this topic.This has important application value and theoretical significance.The paper describes the research status of ship detection at home and abroad,including shadow detection,moving target detection,etc.It points out the difficulties in ship detection.The mechanism of several commonly used shadow detection algorithms is analyzed,and their advantages and disadvantages are analyzed and compared.A random shadow partition algorithm based on hypergraph is proposed.The flaws in the false detection of Faster RCNN are analyzed.The details of the two process improvement of Faster RCNN are described.The ship detection algorithm with fusion sample sclection is given.The algorithm principle of the multi-layered HyperNet network is studied,and the position information and detection accuracy of the target ship are compared with the Faster RCNN,and the feasibility of the HyperNet network in practical application is verified.For the traditional shadow detection algorithm,the detection failure due to the similarity between the color of the ship’s side and the color of the shadow.In this thesis,the algorithm principle of hypergraph segmentation is analyzed,the shadow space detection algorithm in color space is improved and optimized,and the random shadow detection partition algorithm based on hypergraph is givenFor the problem of false detection in the actual detection of the Faster RCNN algorithm,the paper proposes a ship detection algorithm that combines sample selection and rejection.The algorithm analyzes the principle of the CFAFT model.Firstly,the Fast RCNN classifier is used to filter the candidate regions generated by the RPN.Then in the classification stage,two binary classifiers were added on the basis of the original classifier for finer precision measurement.Experimental results show that the algorithm can solve the problem of misdetectionAiming at the accuracy of the target frame position information in the actual detection of the Faster RCNN,and the need to further improve the detection accuracy of the Faster RCNN algorithm,this paper presents a multi-featured ship detection algorithm.The algorithm gathers and integrates the characteristics of the high-resolution shallow layers,middle layers,and deep-seated feature maps that are rich in semantic information into a single space.With the actual traffic data set,it has good location information.Compared with Faster R-CNN,the average detection accuracy of mAP and accuracy were improved by 15%and 6.2%,respectivelyIn summary,the innovation and characteristics of the paper are● A hyper-graph-based random shadow detection partition algorithm was proposed to remove the shadow of the moving ship and improve the detection accuracy of the moving target● The two processes of moving target detection are optimized,and the ship detection algorithm with fusion sample selection is proposed.The false detection problem in the actual detection of the Faster RCNN algorithm is solved,and the precision of the moving target detection and the ship parameters of the inland waterway are improved● A multi-featured ship detection algorithm is proposed,which integrates shallow,intermediate,and high-level information,solves the accuracy problem of the target frame position information in the actual detection of the Faster RCNN,and improves the accuracy of target detection.更多还原