基于颜色属性的车辆阴影去除方法研究

发布时间：2018-09-13 15:55

【摘要】：随着科技的发展,时代不断的进步,计算机视觉技术、高速的网络通信技术和电子技术也在不断向前发展。智能交通系统(Intelligent Transportation System,简称ITS)在城市交通的合理规划和管理中发挥着越来越强大的作用。在基于视频帧的序列图像中,运动车辆目标的定位、识别、测速、跟踪是智能交通系统研究的重点之处。智能交通系统要解决的首要问题是进行运动车辆检测,检测的结果会直接影响车辆的后续处理。通常所使用的目标检测算法,很容易将阴影部分当成前景检测出来。所以,能否成功地去除阴影,会直接影响到车辆的检测结果的准确度。本文通过分析阴影的形成原理和颜色属性的特性,提出了一种基于颜色属性的阴影去除方法。由于,光线一般都是平行的直线,离车越近的地方遮挡越大,阴影也就越深,离车远的地方,遮挡相对较小,阴影较轻。颜色属性(即颜色名)是从现实的生活世界中学习得到的。通过Google图像搜索引擎为每个颜色名搜索一定数量的图像数据集,然而,这些数据集中存在着许多错误的正样本。因此,采用概率潜在语义分析模型(PLSA)学习颜色名。该方法利用了颜色属性可以将一张图像映射成边界清晰、颜色分明、不存在渐变像素点的图像,且图像能够很好地保证原有图像的完整性。根据这一特性,可以将较深的阴影部分与车辆映射成不同的颜色块。同时,将映射后的图像进行二值化处理,以达到去除较深的阴影目的。由于,颜色属性是对整幅图像进行处理,会将周围的环境也映射成前景区域。从而很难确定目标所在的位置,以及无法将阴影完全去除。所以,本文结合了基于高斯模型的背景差分。首先,检测出运动车辆,以极大地减小要处理的前景区域。同时,对背景差分的结果进行了形态学上的腐蚀操作,以减少前景边缘部分的噪声。将背景差分与帧问差分结果进行逻辑与操作,在保证车辆目标完整性的前提下,去除周围环境噪声的影响。同时,为了较少内部和边缘孤立的噪声,采用腐蚀操作将其去除,为了减少内部的空洞现象,采用形态学上的膨胀操作。最后,根据阴影区域具有连通性的特性,对车辆区域进行填充,以得到更加精确的车辆目标。
[Abstract]:With the development of science and technology, computer vision technology, high speed network communication technology and electronic technology are developing. Intelligent Transportation system (Intelligent Transportation System,) is playing a more and more powerful role in the rational planning and management of urban traffic. In the sequence image based on video frame, the location, recognition, speed measurement and tracking of moving vehicle targets are the key points of the research of intelligent transportation system (its). The most important problem of Intelligent Transportation system (its) is to detect moving vehicles, and the result of detection will directly affect the subsequent processing of vehicles. Usually the target detection algorithm is used, it is easy to detect the shadow as foreground. Therefore, whether the shadow can be successfully removed will directly affect the accuracy of vehicle detection results. By analyzing the principle of shadow formation and the characteristics of color attributes, a shadow removal method based on color attributes is proposed in this paper. Because the light is usually a parallel straight line, the closer it is to the vehicle, the bigger the shadow is, and the deeper the shadow is, the smaller the shade is and the lighter the shadow is. Color attributes, or color names, are learned from the real world of life. A certain number of image data sets are searched for each color name by Google image search engine. However, there are many wrong positive samples in these datasets. Therefore, the probabilistic latent semantic analysis model (PLSA) is used to learn color names. The method uses color attributes to map an image into an image with clear boundary, clear color and no gradual pixels, and the image can ensure the integrity of the original image. Based on this feature, darker shaded parts can be mapped to different color blocks from vehicles. At the same time, the mapped images are binarized to remove deeper shadows. Because the color attribute processes the whole image, the surrounding environment is mapped to the foreground area. It is difficult to locate the target and to remove the shadow completely. Therefore, this paper combines the background difference based on Gao Si model. First, the moving vehicle is detected to greatly reduce the foreground area to be processed. At the same time, the result of background difference is corroded morphologically to reduce the noise of foreground edge. The background differential and frame differential results are logically and operationally operated to remove the influence of ambient noise on the premise of ensuring the integrity of the vehicle target. At the same time, in order to reduce the internal and edge isolated noise, the corrosion operation is used to remove it, and the morphological expansion operation is used to reduce the internal cavity phenomenon. Finally, according to the connectivity of the shadow region, the vehicle area is filled to obtain more accurate vehicle targets.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【参考文献】