基于视频图像的车型识别算法研究与实现

发布时间：2018-10-13 19:55

【摘要】：车辆自动识别技术是智能交通系统（ITS）的重要组成部分，通过对车辆进行自动识别，可以为交通管理、收费、调度、统计提供数据。车型识别是智能交通领域的研究热点和难点之一，目前我国的车型识别率还难以满足使用要求，对车型识别率提高算法的研究势在必行,本文研究的就是基于车脸图像特征的车型识别率提高算法。本文首先对车辆进行检测与中值滤波，建立车辆样本库。然后截取了车脸图像，并分别采用灰度特征、Canny边缘特征、Sobel边缘特征和HOG特征来表示图像，，通过支持向量机训练并检测样本，获得各自的识别率，比较并分析其结果。接着通过一种投票提升算法，将24种识别率不高的Gabor特征组合起来进行车型识别，以提高识别率。本文使用华硕A43EI235SD-SL笔记本电脑，在Windows7操作系统中，用OpenCV和VS2008搭建实验平台。实验共采集了100类车型，每类车型各1张训练和测试样本，车脸图像尺寸为192*64。实验中利用灰度特征的识别率为53%，平均识别时间为39.47ms/张；采用Canny边缘特征，识别率为55%，平均识别时间为47.85ms/张；采用Sobel边缘特征，识别率为69%，平均识别时间为46.37ms/张；采用HOG特征，识别率为78%，平均识别时间为59.82ms/张；将24种gabor特征识别结果投票提升，识别率可以达到81%，识别平均时间85.24ms/张。实验结果表明，本文使用的投票提升算法，识别率不仅优于单个gabor特征，也优于前面几种特征表示法，代价是增加了一定的时间消耗。
[Abstract]:Automatic vehicle recognition is an important part of Intelligent Transportation system (ITS). It can provide data for traffic management, charge, scheduling and statistics through automatic identification of vehicles. Vehicle recognition is one of the research hotspots and difficulties in the field of intelligent transportation. At present, the vehicle recognition rate in our country is still difficult to meet the requirements of application, so it is imperative to study the algorithm of improving vehicle recognition rate. In this paper, the vehicle recognition rate improvement algorithm based on vehicle face image features is studied. First of all, vehicle detection and median filtering are carried out to establish the vehicle sample bank. Then, the images are captured and represented by gray level feature, Canny edge feature, Sobel edge feature and HOG feature, respectively. The recognition rate is obtained by training and detecting samples by support vector machine (SVM), and the results are compared and analyzed. Then 24 Gabor features with low recognition rate are combined to improve the recognition rate by a voting lifting algorithm. This paper uses Asus A43EI235SD-SL notebook computer, in Windows7 operating system, using OpenCV and VS2008 to build the experimental platform. A total of 100 kinds of vehicle models were collected, with one training and testing sample for each type of vehicle. The image size of the face of the vehicle was 1922 ~ (64). In the experiment, the recognition rate of gray feature is 53 and the average recognition time is 39.47ms/ sheet; using Canny edge feature, the recognition rate is 55 and the average recognition time is 47.85ms/ sheet; using Sobel edge feature, the recognition rate is 69 and the average recognition time is 46.37ms/ sheet; using HOG feature, The recognition rate is 78 and the average recognition time is 59.82ms/, and the result of 24 gabor features can be improved by voting, the recognition rate can reach 81 and the average recognition time is 85.24ms/. The experimental results show that the recognition rate of the proposed voting lifting algorithm is better than that of the single gabor feature and several previous feature representations at the cost of a certain amount of time consumption.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】