【作者】 宋锐；

【导师】 陈辉；

【作者基本信息】 山东大学 ， 电子与通信工程， 2018， 硕士

【摘要】 随着人工智能在全球范围迅速的蓬勃发展,其相关的应用分支也随之进入风口浪尖的地位。智能汽车和辅助驾驶作为人工智能的重要应用分支,等成为人们关注和研究的热点。相较于人为驾驶汽车,辅助驾驶可以在遇到紧急道路情况时做出更及时更准确的判断和解决方案,在一定程度上降低城市道路交通事故的发生率。而车道检测作为车道偏离预警系统的核心技术之一,有重要的研究意义。针对现有车道检测算法复杂度较高,计算量大,适用模型单一和难以应用于低功耗便携设备的不足,本文提出了一种新的基于几何矩采样的最佳去噪车道检测算法,能够快速准确地检测出车道线。首先,通过跳跃式处理连续帧图像,对单帧图进行线性预测分段采样车道信息,获取动态感兴趣区域(ROI);然后,对采样图像进行预处理,用最佳去噪阈值分割出车道候选区域并依次标注二值图像中的连通分量,利用面积阈值初步筛除干扰信息;最后,分析连通分量的各阶几何矩,通过后置去噪处理进一步去除非车道干扰线段,分别计算出各段车道线的质心和主轴偏转角,结合二者分段拟合出车道线,实现车道检测。该方法主要优势有:分段采样车道信息,既克服现有检测算法感兴趣区域固定的缺点,又大幅度压缩图像预处理所需数据量。相比较大部分算法中的整幅图像预处理,本算法耗时少且受环境和光照的影响较小。利用分割出的连通域的各阶几何矩来计算车道中心线,与现有检测算法的区别在于:省略了现有算法对车道线骨架化的处理步骤及曲线拟合过程,是一种简单高效的方法,节省了存储空间并可满足对检测算法的实时性要求。通过对大量合成图以及真实道路图像进行验证测试,本文算法具有较高的检测效率和较好的抗噪能力,能够快速准确地检测出不同形状的车道线,且对不同的光照条件、路面污渍、路面破损,阴影反光等复杂车道环境具有良好的适应性,在多数情况下,该算法都能准确实现直线车道和曲线车道的检测跟踪。与传统车道检测方法相比,文中算法简单高效,利用分段采样有效保留了车道信息,消除了大量干扰噪声,减少了传统方法在图像预处理阶段的数据计算量;与现有前沿检测算法相比,几何矩计算降低了算法的复杂度,提高了车道检测跟踪的处理速度,能够适应对检测算法的实时性要求;且文中实验影像部分采用运动相机和手机拍摄,可应用于低功耗便携设备,具有很高实用性价值。更多还原

【Abstract】 The rapid advances of artificial intelligence(AI)over the world has raised a global boom in the research of its associated applications.Smart cars and Advanced Driver Assistance Systems(ADAS)as one of its significant branch,is concerned and researched by relative experts.Compared to man-driving vehicles,ADAS can make more timely and accurate judgments and reaction to hazardous road conditions,and reduce the incidence of city traffic accidents to a certain extent.As one of the core technology for Lane Departure Warning System(LDWS),lane detection has fatal research significance.Existing lane detection algorithms come with costs regarding the required memory and the complexity of the algorithms limiting ways of generating versions adaptive to challenging road environments and portable low-power consumption equipments.We propose a new solution for lane detection based on optimal denoising geometric moment sampling.First,the current frame of a recorded video is processed by selecting potential regions of interest(i.e.potentially containing lane markings)based on piecewise sampling.Then,the visible road surface is segmented by labeling connected components of equivalent pixel values using an image binarization procedure.By analyzing the different order of geometric moments of connected components in the lane region,the centroid and the direction angle of detected parts of lane marking segments are calculated.We combine these calculated parameters;a lane marking segment is finally detected via piecewise representing to the lane marking segment.The key advantages of the proposed method lie in the following parts:1.Piecewise sampling lane line not only solved the limit of fixed region of interest for various multi-lane models,but reduced the required memory of image pre-processing.2.By using geometric moments computation,lane segments are detected without edge extraction and lane fitting.Hence,it’s a simple,efficient and memory-saving approach,while satisfying the requirement of real-time detecting.The proposed method is evaluated in simulate and real-word situations.Experimental results under various lane shape,weather and ligiting conditions(including straight,curve,nighttime,shadow or overlap)show that our method outperforms traditional lane detection algorithms and some frontier approach in both detection rate and real time consumption.The proposed method not only has real-time performance and a high accuracy in detecting lanes of varying appearance,but also offers convenient adaptation of detection both to light-reflective road surface and further types of lighting disturbances.更多还原