导　　师： 刘伟铭

学科专业： 082302

授予学位： 博士

作　　者： ;

机构地区： 华南理工大学

摘　　要： 为了提高交通效率，改善交通状况，全世界范围兴起了使用智能交通系统/(ITS/)来管理交通的浪潮。要构建一个完整的ITS系统，首要任务是构建一个能准确、高效地获得交通特征参数的交通信息采集系统，而针对交通场景的实时车辆检测和识别技术是这个系统的核心，其中基于视频的车辆检测和识别是现在研宄的热点。基于视频的车型识别本质是车辆形状分析，形状分析主要分目标形状提取、形状特征表达、形状特征约减、形状分类识别等四个部分。本文的主要工作围绕这四个部分展开工作： 针对运动目标阴影经常被当成运动目标一起检测出来，如果不消除，将给后续运动目标的跟踪、识别造成很大困难的问题，本文提出一种基于阴影像素点与参考背景像素点多特征差异性的阴影消除方法，把阴影识别作为一个模式识别的问题，对当前帧图像的每个像素点抽取与背景帧的亮度差异性、色度差异性、纹理差异性，形成像素点的特征向量，把所得像素点特征向量与典型阴影点的特征向量进行匹配。并经过实验证明本文方法比近年来多种阴影消除方法的性能更好。 针对基于视频进行实时形状识别时，要求形状描述子既能准确的描述形状，又能快速地被提取出来的问题，本文研制一种极半径高度函数形状描述子，该描述子包含了形状轮廓采样点的相对位置关系，能够准确地捕捉物体的轮廓形状特征；并在多个标准数据库上进行了极半径高度函数形状描述子和多种近年流行的形状描述子比较实验，且在实际生产的工程中进行了检验，实验证明极半径高度函数形状描述子具有良好的形状检索性能。 针对基于视频进行实时形状识别时，如何找出数据集中不相互关联的、重要的特征，以降低算法的时间和空间复杂度，提高数据质量及数据泛化能力的问题，本文提出了一个基于学习向量量化/(LVQ/)分类低损失的特征选择评价准则，基于这个评价准则，提出了一个通过最小化损失函数来优化LVQ分类间隔的特征选择算法：LVQMFS。在多个UCI数据集之上进行了LVQMFS和近年来多种特征选择算法的对比实验。实验证明，LVQMFS的特征选择性能明显优于或达到近年来多种特征选择算法的性能。 针对欧氏距离度量作为向量相似性度量时忽视向量各特征的取值范围的差异性，从而影响学习向量量化算法（LVQ/)及其变种的分类精确度的问题。本文提出了一种面向特征取值范围的向量相似性度量函数，并基于该度量函数与泛化学习向量量化算法/(GLVQ/)研宄出一种新算法GLVQ-Range。使用UCI机器学习库中8组数据进行了GLVQ-Range和传统其它LVQ变种算法的对比实验，验证了算法的分类准确性和算法运算速度。使用视频车型分类数据，验证了GLVQ-Range在真实工程环境中的可用性。 针对基于视频的实时车型分类准确性抽检复核困难的问题，本文引入了聚类过程的全局性控制模糊矩阵，进而提出了基于模糊矩阵的蚁群聚类算法。实验结果证明了基于模糊矩阵的蚁群聚类算法的正确性和高效性，以及使用其对基于视频的实时车型分类数据抽检复核的可行性。 In order to improve the efficiency of traffic system, the Intelligent Transportation Systems /(ITS/) is used to manage traffic world-widely. To build a complete ITS, the first task is to build an accurate and efficient traffic information collection system that collects characteristic traffic parameters in real time. Vehicle detection and recognition technology is the core of the traffic information collection system. Vehicle detection and recognition based on video is a hot research topic now. Vehicle recognition based on video is vehicle shape analysis essentially. Shape analysis consists of four parts:target shape extracting, shape feature expression, shape feature subtraction, and shape classification etc. The main work of this thesis is around these four parts: The cast shadow is always detected as moving object. If not eliminated, the cast shadow will cause great trouble for tracking and recognition of moving object. To solve this problem, a shadow removal algorithm based on multiple feature differences between pixels and the reference background pixels is presented. The performance of the proposed method better than that of other shadow elimination methods in recent years is proved by experiments. In real time shape recognition based on video, the shape descriptor is desired to characterize the shape accurately, and can be extracted fast. Polar Radius Height Functions /(PRHF/) shape descriptor is proposed to solve this problem. The descriptor contains the relative position description of the shape contour sampling points, which is able to accurately capture the shape contour features. A number of comparative experiments are conducted on3standard databases for comparing shape retrieval performance of PRHF and many popular shape descriptors in recent years. The shape retrieval performance of PRHF achieves similar or even higher than that of shape descriptor s in recent years is proved by experiments. In real-time shape recognition based on video, feature selection play an important role that can find out the important features to reduce the algorithm's time and space complexity and to improve data quality and data generalization ability. A new vector similarity metric for learning vector quantization /(LVQ/) is advised, and a new feature selection evaluation criterion based on low loss function of LVQ classification is proposed. Based on the evaluation criterion, a feature selection algorithm named LVQMFS that optimizes the hypothesis margin of LVQ classification through minimizing its loss function is presented. Some experiments that compared with feature selection algorithms in recent years are carried out on several UCI data sets. The performance of new algorithm achieves similar or even higher than that of feature selection algorithms in recent years is proved by experiments. The feature data range is ignored when Euclidean distance used as a vector similarity metric, which affects the classification accuracy of the traditional learning vector quantization algorithm /(LVQ/) and its variants. To solve the problem, a novel vector similarity metric is proposed, and a new algorithm named as GLVQ-Range based on this metric and GLVQ is put forward. The classification accuracy and computation speed of the algorithm are tested in comparison to traditional alternative LVQ algorithms, using8datasets of UCI machine learning repository. The algorithm usability in real production environment is verified through the video vehicle classification data set. It is difficult to verify classification accuracy of vehicle classification based on video. To solve the problem, the fuzzy matrix for controlling of data clustering process is proposed, and then ant colony clustering algorithm based on fuzzy matrix is proposed. The experimental results show that fuzzy matrix ant colony clustering algorithm is correct and efficient, and verifying classification accuracy of real-time vehicle classification based on video is feasible.

关 键 词： 智能交通 车型识别 形状分析 阴影消除 形状描述子 特征选择 分类算法 蚁群聚类算法

分 类 号： [TP391.41]

领　　域： [自动化与计算机技术] [自动化与计算机技术]