基于大气散射原理构建模型检测夜间交通视频多景深车灯

发布时间：2018-03-25 16:08

本文选题：夜间交通视频　切入点：中远景车灯　出处：《天津工业大学》2017年硕士论文

【摘要】：近几年来,由于夜间事故的频发和其环境的复杂性,夜间车辆检测作为系统中的一部分,加上智能交通系统车辆检测技术越发娴熟,对该方面的研究越来越受国内外学者以及企业的关注和重视。夜间交通环境下,照明条件不足,导致车辆如轮廓、颜色等信息的丢失,这些因素极大的限制了夜间车辆的检测。车辆在运动过程中,车灯的相对于车辆其它特征信息而言较为稳定、可靠,因此目前关于夜间车辆的检测大多数以车灯作为研究的主要特征选择。但是由于夜间场景中环境光及车灯强反射光的干扰,以及视野远处的车辆车灯容易粘连。针对以上的这些问题,本文提出了基于大气散射原理构建模型检测夜间交通视频多景深车灯的方法,以解决提高检测率的基础上延长车辆车灯的检测范围。我们主要是通过对夜间场景下所有光源及其对图像成像过程中的影响进行了分析,并根据大气散射原理构建了车灯复原模型,以实现夜间车辆的检测。通过所构建的模型对夜间交通视频的车灯复原时,需要先对模型中透射率、环境光以及场景景深等参量的估计:透射率的估计是通过对原始图像取反归一化得到.,本文中所定义的环境光不仅是指夜间场景中如路灯、广告牌等光源形成的环境光,还有车辆行驶过程中车灯的本身散射以及车灯在路面的反射光引起场景光源变化的环境光,即背景环境光和车灯区域环境光,然后对图像经相应的处理之后分别通过暗原色通道原理估计得到;在同一场景中不同景深的车灯在传输过程中经大气散射的程度不同,本文在模型中引入了场景景深参量,并根据透射率公式估计该参量。最后通过复原模型得到了车灯复原结果。由于场景中强反射光与车灯的亮度值相近,会同车灯一起被复原,因此需要进一步筛选。本文考虑到不同景深对车灯的影响,车灯的特征信息也会因不同景深而不同,所以划分区域分别对相应区域的光斑进行筛选,使车灯与路面强反射光分离,从而达到车灯检测的目的。通过对9段视频,共14492帧进行测试,结果表明,在延长了检测距离的同时,本文所提模型与同类先进算法相比,平均检测率提高了 31.39%,平均漏检率降低了 20.93%,平均误检率降低了 10.46%。
[Abstract]:In recent years, due to the frequent occurrence of night accidents and the complexity of its environment, nighttime vehicle detection as a part of the system, plus the intelligent transportation system vehicle detection technology, has become more and more skillful. The research on this aspect has been paid more and more attention by domestic and foreign scholars and enterprises. Under the traffic environment at night, the lighting conditions are insufficient, resulting in the loss of information such as the outline and color of the vehicle, etc. These factors greatly limit the detection of vehicles at night. In the course of vehicle movement, the lights of vehicles are relatively stable and reliable in comparison with other characteristic information of vehicles. Therefore, at present, most of the detection of nighttime vehicles is based on vehicle lights. However, due to the interference of the ambient light and the strong reflected light of the vehicle lights in the night scene, In view of these problems, this paper presents a model based on atmospheric scattering theory to detect multi-field vehicle lights in night traffic video. In order to solve the problem of increasing the detection rate, we extend the detection range of vehicle lights. We mainly analyze all the light sources in the night scene and their influence on the image imaging process. According to the principle of atmospheric scattering, the vehicle light restoration model is constructed to detect the night vehicle. The transmissivity of the model is needed to restore the vehicle lights of the night traffic video through the built model. Estimation of environmental light and scene depth of field: the estimation of transmittance is obtained by reverse-normalization of the original image. The ambient light defined in this paper not only refers to the ambient light formed by light sources such as street lights and billboards in the night scene. There is also the background ambient light and the regional ambient light of the vehicle lamp, which is caused by the scattering of the vehicle lights and the reflected light of the vehicle lights on the road. Then the image is estimated by the principle of dark primary channel after the corresponding processing. In the same scene, the different depth of field lights in the transmission of different degrees of atmospheric scattering, this paper introduced the scene depth parameters in the model. The parameters are estimated according to the transmittance formula. Finally, the restoration results of the vehicle lamp are obtained by the restoration model. Because the strong reflected light in the scene is close to the luminance value of the vehicle lamp, it is restored together with the vehicle lamp. Therefore, further screening is needed. Considering the influence of different depth of field on vehicle lamp, the characteristic information of vehicle lamp will also be different according to different depth of field. The vehicle lamp is separated from the road surface strongly reflected light, and the purpose of vehicle lamp detection is achieved. Through the test of 9 video segments, 14492 frames are tested. The results show that the model proposed in this paper is compared with the similar advanced algorithm while prolonging the detection distance. The average detection rate increased 31.39 percent, the average missed detection rate decreased 20.93 percent and the average false detection rate decreased 10.46 percent.
【学位授予单位】：天津工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【相似文献】