【作者】 李辉；

【导师】 杨晨晖；

【作者基本信息】 厦门大学 ， 计算机技术， 2014， 硕士

【摘要】 车辆图像检索是智能交通系统的重要组成部分,车辆检索在停车场智能管理、高速公路自动收费、道路监控、超时停车检测等方面有着非常好的应用前景。本课题来源于高速公路服务区车辆智能检测项目,是基于特定场景下的车辆图像检索问题。车辆图像检索主要是基于车辆检测、目标区域提取、特征提取和特征匹配的相关技术,目前关于图像检索的技术研究相对来说比较成熟。本文旨在进一步研究车辆图像特征在索引层面的融合,在有光照变化等情况的交通场景中能够进行准确的车辆图像检索。本文的主要的研究内容如下：1.研究图像检索的流程和经典算法,分析和总结了车辆图像检索流程中的关键要素、检索流程和特征提取。研究了不同检索流程和特征提取方法的优缺点,并分析了它们在实时车辆图像检索中的可行性。2.基于可变部件模型的车辆目标区域提取。为了减少检索时间、计算量以及车辆以外区域对检索的影响,本文首先对车辆目标区域进行了提取,把不同背景下的车辆目标部分作为特征提取区域,大大减少了处理时间,提高了检索效率。3.基于多视觉词汇索引的车辆图像检索。本文提出一种将局部对称性特征和局部颜色特征在索引层面上进行融合的多维索引方法。具体地说,互补的两种特征被集合到了一个多维的倒排索引中。首先我们选择局部对称性特征来表达车辆图像,其具有很好的光照适应性和不变性。为了进一步加强视觉词汇的分辨力,我们加入颜色特征来反映图像中局部颜色的分布,使得两种视觉词汇形成互补关系。本文应用基于TF-IDF加权的词袋模型,不同特征描述子被赋予不同权值。通过实验分析,基于多视觉词汇索引的车辆图像检索技术能很好的用于快速实时车辆图像检索。更多还原

【Abstract】 The vehicle image retrieval is an important part of intelligent transportation system. Vehicle retrieval has good application prospect in intelligent parking lot management, highway automatic charge, road monitoring and parking timeout detection. This paper comes from vehicle intelligent monitoring projects of highway service area, the problem bases on fixed background of vehicle image retrieval.Vehicle image retrieval has many relate technologies, such as vehicle detection, object area segmentation, feature extraction and feature matching, current technology in image retrieval is relatively mature. The purpose of this paper is to further study vehicle image feature fusion at indexing level, in response to the changes of illumination and luminance and to be able to make retrieval real-time. This article main research contents are as follows:1. Researching on the process and classical algorithm of image retrieval, analyze and summarize the key elements in the image retrieval process and feature extraction. Studying the advantages and disadvantages of different retrieval process and feature extraction methods, and analyze their feasibility in real-time vehicle image retrieval.2. Deformable part model based vehicle area segmentation. In order to reduce the retrieval time and workload of computional, first we need to carry on the vehicle localization, we extract feature only in vehicle regions, this greatly reduces the processing time, improving the efficiency of retrieval.3. The vehicle image retrieval based on multi-visual word index. We present a multi-dimensional index method to perform feature fusions of local symmetry feature and local color feature at indexing level. Specifically, two complementary features are coupled into a multi-dimensional inverted index. First we choose local symmetry feature to represent the vehicle image, it has strong illumination adaptability and invariance. To further enhance discriminative power of the visual word, we add local color feature to reflect distribution of local color in vehicle image. Thus, they form a complementary relationship with each other. We use bag of words as the retrieval model with TF-IDF weight, different feature descriptor is assigned to different weight.Through experimental analysis, vehicle image retrieval technique basing on multiple vocabularies index achieves good effect in accurate vehicle image retrieval.更多还原