交通场景下汽车车型图像检索技术研究

发布时间：2018-07-26 11:02

【摘要】：智能交通系统是目前交通领域的热点课题,而基于图像的车型自动检索是智能交通系统中的重要研究方向,可应用在公安侦缉工作中,能够协助公安部门对交通肇事逃逸或者其他犯罪活动的调查。交通场景下汽车车型图像检索的研究具有较高的理论价值和实际意义。本文重点研究了车型图像检索中的特征点提取与特征匹配技术,主要研究内容包括:(1)研究了车脸图像的SIFT底层特征描述与提取,构建了基于SIFT描述子的关键点数据结构。SIFT方法具有对图像旋转、缩放、平移和仿射等变换的不变性,可在一定程度上降低由于道路中行驶汽车与高清摄像头之间距离与角度不同,所导致的不同目标图像间尺度与角度变化的影响。(2)提出了一种基于车脸图像特征点关注度的SIFTKeyPre特征匹配算法,可以根据基准图像特征点对图像匹配的贡献度不同,赋予基准图像特征点不同的关注度。分析了该算法应用在车脸图像样本集上的检索效果,讨论了优选过程中阈值的确定,训练强度的收敛值以及车牌上特征点关注度对检索效果的影响。与未区分不同特征点对匹配结果影响的常用匹配算法Flann、Lowe相比,在不增加时间复杂度的前提下,所提出的SIFTKeyPre匹配算法具有更好的检索效果。(3)研究了一种基于视觉词袋模型和支持向量机相结合的车脸图像特征匹配算法。首先通过视觉词袋模型对SIFT特征点编码,利用K-means聚类算法将每幅车脸图像表示为一维长度固定的向量,再使用支持向量机进行匹配。实验结果和理论分析表明,其检索效果好于Flann、Lowe算法且仅次于SIFTKeyPre算法,但其计算耗时是最少的。
[Abstract]:Intelligent Transportation system (its) is a hot topic in the field of transportation at present, and the automatic retrieval of vehicle model based on image is an important research direction in Intelligent Transportation system (its), which can be applied to the investigation of public security. Be able to assist the public security department in the investigation of traffic hit-and-run or other criminal activities. The research of vehicle image retrieval in traffic scene has high theoretical value and practical significance. This paper focuses on the feature point extraction and feature matching technology in vehicle image retrieval. The main research contents are as follows: (1) the SIFT bottom feature description and extraction of Che face image is studied. The key point data structure based on SIFT descriptor is constructed. The method SIFT has invariance for image rotation, scaling, translation and affine transformation. It can be reduced to some extent because of the difference in distance and angle between the car and the high-definition camera on the road. (2) A SIFTKeyPre feature matching algorithm based on Che face image feature point concern is proposed, which can make different contribution to the image matching according to the reference image feature points. Give different attention to the feature points of the reference image. The retrieval effect of the algorithm applied to the Che face image sample set is analyzed. The effects of the threshold, the convergence value of the training intensity and the attention of the feature point on the license plate on the retrieval effect are discussed. Compared with Flannn Lowe, a common matching algorithm that does not distinguish the influence of different feature points on the matching results, it does not increase the time complexity. The proposed SIFTKeyPre matching algorithm has better retrieval effect. (3) A feature matching algorithm based on the combination of visual word bag model and support vector machine is studied. Firstly, the SIFT feature points are coded by the visual word bag model, and each face image is represented as one dimensional vector with fixed length by using the K-means clustering algorithm, and then the support vector machine is used to match the feature points. Experimental results and theoretical analysis show that the retrieval efficiency is better than that of Flannor Lowe algorithm and second only to the SIFTKeyPre algorithm, but the computation time is the least.
【学位授予单位】：石家庄铁道大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U495;TP391.41

【相似文献】