车路协同系统中基于部件的车辆目标检测方法研究

发布时间：2018-03-04 16:18

本文选题：车辆目标检测　切入点：部件　出处：《长安大学》2016年硕士论文　论文类型：学位论文

【摘要】：车辆目标检测是车路协同系统中的关键技术之一,作为智能交通领域的重要研究方向,受到很多学者的关注,基于机器视觉的检测方法目前已经取得显著效果,但是由于交通环境复杂、天气变化等的影响,依然存在一些问题。本文以交通场景中的车辆目标为研究对象,通过车辆部件进行目标检测。由于摄像机透视投影成像原理,二维图像中的目标会发生尺度、形状的变化,通过对交通场景标定,将二维图像逆投影到三维空间,建立三维逆投影面,重建逆投影数据,可以有效避免二维图像中出现的问题。车辆目标可以通过多个部件进行表征,并通过部件的组合完成目标的检测。在夜间和白天场景中,车辆目标呈现不同特征,夜间车头灯光亮度比较显著,而白天车辆尾灯和牌照部件颜色很突出。前者通过背景差分法分割目标,对目标进行预处理提取车灯部件,结合车灯目标的几何特征,进行目标车灯对的粗筛选,将候选目标与尺度因子相结合,利用已经建立的高斯混合模型,完成目标车灯对的精确检测;后者通过颜色转化模型分别定位牌照和尾灯部件,通过部件之间的空间几何位置关系,引入马尔可夫随机场,实现车辆目标的检测。本文算法在车路协同系统中得到了有效应用,通过现场测试,结果显示能够对车辆目标能够达到较好的检测效果,在白天场景下检测率达到93%,夜间环境能够达到94%,而且算法可以进行实时检测。
[Abstract]:Vehicle target detection is one of the key technologies of cooperative vehicle infrastructure system, as an important research direction in the field of intelligent transportation, attracting the attention of many scholars, detection method based on machine vision has achieved remarkable results, but because of the traffic environment is complex, the impact of weather changes, there are still some problems. The vehicle objects in the scene as the research object, detect target vehicle components. Due to the camera perspective projection principle, the two-dimensional image of the target will scale, shape changes, through the calibration of the traffic scene, the two-dimensional image inverse projection to the three-dimensional space, the establishment of three-dimensional inverse projection reconstruction, inverse projection data, can effectively prevent the occurrence of the two-dimensional image of the problem. The vehicle can be characterized by a number of components, and complete the target detection through the combination of components. In the night and day In the scene, the target vehicle shows different characteristics, the night light brightness front is significant, and the day vehicle taillights and license plate color components is very prominent. The former through background subtraction segmentation, the target extraction pretreatment of lamp components, combined with the characteristics of geometric light target, coarse screening of target lights, candidate targets with the combination of scale factor, using the Gauss mixture model has been established, complete the accurate detection of target lights; the latter through the color transformation model respectively positioning license plate and taillight parts, components through the space between the geometric relation, Markov random field is used to realize the vehicle detection, the target. This algorithm in collaborative vehicle road system in the application of the field test, the results showed that it can on vehicle target can achieve better detection and measurement rate reached 93% in the daytime scene detection, night The environment can reach 94%, and the algorithm can be detected in real time.

【学位授予单位】：长安大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;U495

【相似文献】