基于HOG特征的船舶识别跟踪算法

发布时间：2018-03-30 06:43

本文选题：HOG　切入点：降噪　出处：《大连海事大学》2017年硕士论文

【摘要】：在水路交通中,航行船舶类型日益多样化,船舶航行轨迹日益复杂化。监控船舶活动需要对船舶目标做到实时的跟踪,还要能够识别船舶目标,而传统的船舶视频目标跟踪方法存在跟踪计算时耗大、跟踪准确率有限、缺少识别能力等缺陷,这对于有效及时的指挥调度船舶带来了困难。因此有必要设计一种实时、误差率小、具有识别能力的船舶目标跟踪算法。本文的研究包括目标的特征提取、分类识别、跟踪三个方面。在目标特征提取方面,通过对各种图像特征的优缺点对比,本论文选定提取船舶目标的HOG特征以减少水域环境中其他干扰背景对船舶识别的影响。在目标分类识别方面,SVM模型对目标的原始HOG特征有一定的分类识别能力,但并不是所有的目标HOG特征位包含的都为有效特征,其中掺杂了噪声存在着冗余,并且模型复杂过高,因此本论文引进序列前向选取法对原始的船舶目标HOG特征进行了降噪和特征再选取。但是由于在处理训练样本数据集的时候采取的是交叉验证方法,并且序列前向选取法存在只能加入不能去除特征的缺陷,因此由其选取的最优特征具有不确定性并且关联性强。针对上述缺陷,在序列前向选取法的基础上,本文提出了一种特征位得分系统从而挑选出了船舶HOG特征中的最优特征位。在目标跟踪方面,本论文引进STC算法来对船舶目标进行定位。虽然STC跟踪算法计算速度快并且跟踪准确率高,但是当目标被遮挡时,会发生跟踪目标跳变的情况。本论文通过模板匹配算法对跟踪目标进行检测,改善了 STC的这一缺陷,从而提升了其跟踪性能。实验证实,本文船舶识别与跟踪算法能够实时、稳定且准确地识别跟踪船舶目标。
[Abstract]:In waterway traffic, the types of navigation ships are becoming more and more diversified, and the ship trajectory is becoming more and more complicated. Monitoring ship activities requires real-time tracking of ship targets and the ability to identify ship targets. However, the traditional ship video target tracking method has many disadvantages, such as large amount of tracking calculation, limited tracking accuracy, lack of recognition ability and so on, which brings difficulties to the effective and timely command and dispatch of ships, so it is necessary to design a real-time system. The research of this paper includes three aspects: target feature extraction, classification recognition and tracking. In the aspect of target feature extraction, the advantages and disadvantages of various image features are compared. In this paper, the HOG features of ship targets are selected to reduce the influence of other interference background on ship recognition. In the aspect of target classification and recognition, the HOG model has a certain ability to classify and recognize the original HOG features of the target. However, not all the target HOG feature bits contain valid features, in which the doped noise is redundant and the model is too complex. So this paper introduces the method of forward selection of sequence to reduce the noise and re-select the features of the original ship target HOG. However, the cross-validation method is adopted in the processing of the training sample data set. And the method of sequence forward selection can only add the defect that can not be removed, so the optimal feature selected by it has uncertainty and strong correlation. In view of the above defects, based on the method of sequence forward selection, In this paper, a feature bit scoring system is proposed to select the optimal feature bits in ship HOG features. In this paper, STC algorithm is introduced to locate the ship target. Although the STC tracking algorithm is fast and accurate, but when the target is occluded, In this paper, the template matching algorithm is used to detect the tracking target, which improves the performance of STC and improves the tracking performance. The experiments show that the algorithm of ship identification and tracking can be used in real time. Identify and track ship targets stably and accurately.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】