基于多视角特征的车型识别方法

发布时间：2018-01-14 10:31

本文关键词：基于多视角特征的车型识别方法　出处：《北京交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：车型识别是智能交通系统(Intelligent Transportation System, ITS)的关键技术之一。纵观国内外关于车型识别的研究,多数基于某一个或一种特征对车型进行分类,其识别准确率仅在特定情况下具有稳定性,并且为获得较高识别精度,需进行大量的数据分析,但冗余信息较多,影响到了目标识别的实时性。同时,识别分类方法的自适应优化学习机制也有待进一步完善。针对上述问题,本论文提出了一种基于多视角特征的车型识别方法,旨在更加快速、准确地完成车型的识别。论文的主要研究内容包括： 1、多视角多维特征参数体系的建立。本文建立了包含前视角、侧视角、尾视角的多视角的多维特征混合树形结构体系,并提出了一种基于自适应显著性水平集的轮廓模型用于对体系中不同视角的区域分割,该模型采用基于二维凸包的显著性初始轮廓曲线自适应定位算法来获取演化曲线的初始位置,同时采用正则化的P-M方程替代原Li模型中的高斯滤波。在此基础上,完成了对不同视角的优化特征参数的定义及量化处理。 2、特征参数优化方法的研究。本文研究了一种基于改进型核独立成分分析的特征参数降维优化方法,该方法通过KICA算法获取图像多维特征的独立基元以构造独立基子空间,采用2DPCA算法完成图像去二阶相关和进一步降维处理。同时,本文提出了基于Amari误差和平均相关度作为评价标准的降维效果评价方法。对比仿真实验表明,该参数优化方法能够完成对多维特征参数的有效降维约简。 3、基于改进型支持向量机分类识别模型的提出。本文提出了一种基于组合核函数的自适应支持向量机分类模型,该模型研究了组合核函数以及组合超参数组的确定,在此基础上,采用双角度约束以提高分类识别的效率和精度,即一方面设计基于马氏距离和“aσ-原则”实现对样本数据进行自动优化分选,并结合加权判别算法加快支持向量机的训练测试速度；另一方面设计了基于先验知识的迭代最优参数自适应搜索算法用于核函数参数的设定,以提高分类器的分类识别精度。仿真实验结果表明,基于自适应显著性水平集的轮廓模型分割方法的准确率稳定在95%以上；基于改进型KICA模型的特征参数优化方法的Amari误差低于6%,平均相关度稳定在97%以上；基于组合核函数的自适应支持向量机分类模型对不同车型的识别率为97.926%,其训练、测试时间分别为1.9s和44.7ms。证明本文改进模型能够满足车型识别分类的需求,具有识别速度快、准确率高等优点,这对于智能交通系统及车型识别系统的发展具有重要的理论及实际意义。
[Abstract]:Vehicle recognition is the intelligent transportation system (Intelligent Transportation System, ITS) is one of the key technologies. Research on vehicle recognition at home and abroad, the majority of one or a feature based on the classification model, the recognition accuracy rate only has stability under specific circumstances, and in order to obtain a higher recognition accuracy, need analysis a large amount of data, but more redundant information, affect the real-time target recognition. At the same time, the classification method of adaptive learning mechanism is to be further improved.
To solve the above problems, a vehicle recognition method based on multi view features is proposed in this paper, aiming to identify vehicle types more quickly and accurately.
1, establish the parameters system of multi angle multi-dimensional characteristics. This paper built a front view, side view, tail angle multidimensional mixed tree structure system, and proposed an adaptive level set based on the contour model for segmentation of different regions from the perspective of system, the model adopts the initial position get the evolution curve based on significant initial contour adaptive localization algorithm of two-dimensional convex hull, while using P-M equation regularization to replace the original Li model of Gauss filter. On this basis, completed the definition and quantitative optimization parameters of different perspective.
2, research on the optimization method of parameters. In this paper a dimensionality reduction method improved feature parameter analysis based on kernel independent components, the image acquisition methods of multidimensional KICA algorithm through independent primitives to construct independent basis subspace, using 2DPCA algorithm to complete the image to two order correlation and further reduction. At the same time, this paper proposes a Amari error and average correlation degree as the evaluation standard evaluation method based on dimension reduction effect. The simulation results show that the optimization method can complete the multidimensional characteristic parameters of Jane dimension reduce.
3, this paper presents an improved support vector machine classification based on model. This paper proposes an adaptive combination of kernel function of support vector machine classification model based on the model of combination of kernel function and parameter combination of super group, on the basis of this, using double angle constraint to improve the efficiency and accuracy of classification, i.e. a design of Mahalanobis distance and a sigma principle realize automatic optimization of sorting based on the sample data, and combined with the weighted algorithm to accelerate the training speed of testing support vector machine; on the other hand, the design of search algorithm for kernel parameters based on adaptive iterative optimal parameters a priori knowledge of the set, in order to improve the classification accuracy classifier.
Simulation results show that the segmentation method of contour model adaptive level set accuracy rate is more than 95% based on improved KICA model; feature parameter optimization method based on the Amari error is less than 6%, the average correlation stable above 97%; adaptive combination of kernel function of support vector machine classification model to identify the different models of the rate 97.926%, based on the training and testing time are respectively 1.9s and 44.7ms. prove that the improved model can meet the vehicle classification requirements, high recognition speed, high accuracy, it has important theoretical and practical significance for the development of intelligent transportation system and vehicle recognition system.

【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】