基于支持向量机的车型识别

【作者】 陈爱斌；

【导师】 夏利民；

【作者基本信息】 中南大学 ， 计算机应用技术， 2004， 硕士

【摘要】 车辆自动识别分类技术是智能运输系统的重要组成部分，它对特定地点和时间的车辆进行识别和分类，并以之作为交通管理、收费、调度、统计的依据。车型的自动分类在部分发达国家已经是比较成熟的技术，但由于诸多原因这些系统在我国其识别率等指标难以满足使用要求。要实现我国公路收费自动化、管理规范科学化，车型自动识别方法的研究势在必行。 本文研究基于车型图像代数特征的车型识别方法。该方法首先利用背景差分法从背景图像中提取出运动车辆，并对车型图像进行预处理，然后采用特征并行融合的方法将用PCA方法和加权LDA方法提取出来的两类车型特征融合成一种待识别特征，最后将决策支持向量机分类器与最近邻法以及动态Boosting算法相结合构造决策支持向量机车型识别分类器进行车型识别。 本文所做的主要工作为： 1、为了获取更丰富的特征信息，提出将PCA特征与加权LDA特征进行特征级融合的一种车型代数特征提取方法。该方法保留了参与融合的两类特征的有效鉴别信息，又很大程度上消除了信息的冗余，提高了识别率。 2、提出将最近邻法与决策支持向量机相结合进行快速分类的新方法，该算法极大地减少了SVM的训练样本数，训练速度较快，同时减少SVM分类器测试样本的规模，提高分类速度。 3、提出将动态Boosting算法与支持向量机相结合，用更优的弱分类器构造强分类器，得到一个增强了的决策支持向量机分类器，该方法能减少分类错误率。 基于以上方法，构建了一个基于与最近邻法相结合的决策支持向量机的车型识别系统。实验表明，该系统具有速度快、识别率高等特点。更多还原

【Abstract】 Automatic recognition and classification of vehicles is an important part of intelligent transportation system. It focuses on the recognition of vehicle at special place and in special time for traffic management, charging, regulation and statistics. Vehicle classification is a mature technology in developed countries, but due to a series of problems, this technology is unsuitable for application in our country. In order to realize the automatic charging and standardization of highway management, the research of vehicle recognition is imperative.In this paper, we focus on method of vehicle recognition based on algebra features of vehicle images. Firstly, the moving vehicles are picked out from background images using Background-Subtraction, and preprocessed into standardized images. Secondly, two kinds of vehicle features are extracted respectively by PCA and Weighted-LDA. Then the features are turned into a fusing feature by means of feature-fusion in complex space. Finally, by combining decision-tree support vector machine with nearest neighbor method and float-boosting algorithm, a new decision-tree support vector machine is built for vehicle recognition.The main work of this paper is:1. To get abundant information of feature, we put forward a feature fusion method, by which the two kinds of vehicle features extracted by PCA and Weighted-LDA are fused. The valid information of the two features is kept, and the redundant information is eliminated. The recognition rate is improved.2. Combining nearest neighbors method with decision-tree support vector machine, a new faster algorithm of classification is bring forward, which can not only reduce the train samples of support vector machine and speed up the train process, but also reduce the test samples and speed up the classification.3. We also combine the float-boosting algorithm with support vector machine and assemble some weak classifiers to build a boosted decision-tree support vector machine classifier. This classifier can decrease the error rate of vehicle recognition.With these methods, We construct a simple vehicle recognition system based on decision-tree support vector machine. The experiment results show that this system can work in higher speed and with higher accuracy.更多还原