基于高分辨率遥感图像的车辆分类识别研究

发布时间：2018-08-17 19:20

【摘要】：随着遥感技术的不断发展和卫星空间分辨率的提高,基于高分辨率遥感影像的车辆识别研究已取得一定的成果,并且建立了较为完整的车辆识别流程。本文以高分辨率遥感图像的车辆识别流程为基础,改善车辆识别过程中所采用的方法,以获取更高的车辆识别精度。论文主要研究内容如下:(1)在图像预处理方面,本文采用sobel边缘滤波方法处理图像,并与已有的车辆识别研究中采用的线性变换等图像处理方法对比分析,sobel边缘滤波能更好地突出车辆边缘特征,有利于后续车辆的识别。(2)研究高分辨率遥感图像中关于车辆分割的最优尺度问题。本文通过分析车辆在不同分割尺度下的特点提出了基于车辆面积的最优尺度确定方法——全局车辆面积最大法,该方法以车辆的面积作为判定车辆最优分割尺度的依据,计算不同分割尺度下的车辆总面积,当总面积达到最大值时所对应的分割尺度即为车辆的最优分割尺度。该方法比已有研究中常用的基于光谱异质性的RMAS法具有更好的适用性。(3)对高分辨率遥感图像中车辆的特征进行了详细分析并建立了相应的特征类描述,简要介绍了阈值分类和模糊规则两种分类方法,结合两种分类方法的特点,构建了多特征阈值-模糊规则的双分类器方法,用以提高不同亮度车辆的识别精度。(4)采用多种不同类型的高分辨率遥感图像进行实验,结合相应的评价因子对车辆的识别结果进行评价分析,结果正确性达到96%、完整性达到89%和总体质量达到86%。对比于同一分辨率下亮色车辆的识别结果,本文识别方法比以往研究具有更高的识别精度,极少出现错误识别和遗漏识别。结果表明本文所提出的车辆最优分割尺度确定方法与分类方法具有很好的普适性和可行性。
[Abstract]:With the development of remote sensing technology and the improvement of satellite spatial resolution, the research of vehicle recognition based on high-resolution remote sensing image has made some achievements, and a relatively complete vehicle recognition process has been established. Based on the vehicle recognition process of high resolution remote sensing image, the method used in vehicle recognition is improved in order to obtain higher vehicle recognition accuracy. The main contents of this paper are as follows: (1) in the aspect of image preprocessing, sobel edge filtering method is used to process images. Compared with other image processing methods, such as linear transformation, which have been used in vehicle recognition research, Sobel edge filtering can better highlight vehicle edge features. It is beneficial to the subsequent vehicle recognition. (2) the optimal scale of vehicle segmentation in high-resolution remote sensing images is studied. By analyzing the characteristics of vehicles at different segmentation scales, this paper puts forward the method of determining the optimal scale based on vehicle area, which is based on the area of vehicle as the basis for determining the optimal scale of vehicle segmentation. When the total area of vehicle reaches the maximum value, the corresponding segmentation scale is the optimal scale of vehicle segmentation. This method is more suitable than the RMAS method based on spectral heterogeneity. (3) the characteristics of vehicles in high-resolution remote sensing images are analyzed in detail and the corresponding description of feature classes is established. In this paper, two classification methods, threshold classification and fuzzy rule, are introduced briefly. Combined with the characteristics of the two classification methods, a multi-feature threshold-fuzzy rule double classifier method is constructed. It is used to improve the recognition accuracy of vehicles with different luminance. (4) the vehicle recognition results are evaluated and analyzed by using various kinds of high-resolution remote sensing images and the corresponding evaluation factors. The results were 96% accuracy, 89% integrity and 86% overall quality. Compared with the recognition results of brightly colored vehicles at the same resolution, the recognition method in this paper has higher recognition accuracy than the previous research, and there are few errors and omissions. The results show that the method and classification method of optimal vehicle segmentation scale proposed in this paper have good universality and feasibility.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】