融合图像与激光雷达的车道偏离预警算法研究

发布时间：2017-12-28 17:22

本文关键词：融合图像与激光雷达的车道偏离预警算法研究　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：本文研究了融合图像与激光雷达的车道偏离预警系统。车道偏离预警系统是高级辅助驾驶系统中的一项关键技术,主要用于防止车辆偏离当前车道而产生交通事故。本文首先分析了车道偏离预警系统的各组成部分,总结了当前研究的三大难点,并针对每个难点进行了相关研究。主要研究的问题与取得的成果概括如下:1.复杂道路条件下的车道线检测问题。目前的车道线检测算法在稳定光照、良好道路条件下已取得较好的效果。难点在于,在光照变化、阴影遮挡、障碍物遮挡等复杂路面条件下检测车道线。以及在道路磨损,车道线为虚线的条件下检测车道线。本文提出一种适应复杂路面条件的车道线检测算法。通过灰度空间与HSV(Hue Saturation Value)空间的双阈值分割进行特征提取,减少特征的丢失。针对障碍物遮挡、阴影遮挡、路面其他道路标识的问题,利用同一条车道线两边缘相互平行的性质,使用梯度方向判决的方法,减少遮挡物对车道线检测的影响。针对车道线磨损,车道线为虚线的问题,利用改进的河流算法进行特征连接。算法经过高速公路采集的视频数据的测试,能在多种复杂道路条件下稳定检测车道线。2.车道偏离预警模型虚警率高,预警信息不稳定的问题。车道线偏离预警的虚警会干扰驾驶员的注意力,降低系统可用性,甚至错误的改变车辆运动状态,造成危险。预警信息不稳定会导致驾驶员无法判断当前行驶状态是否有危险性,是否需要修正车辆轨迹。本文分析了现有车道偏离预警模型的优缺点,并提出一种阶段式偏离预警模型。在保持检测率与预警时间的基础上,降低了虚警率,提高了预警信息的稳定性。3.单一传感器车道感知可靠性低的问题。目前大多数车道偏离预警研究都使用视觉传感器作为唯一的道路数据来源,这在实际场景下可靠性不足。一方面当传感器出现误差时,没有其他测量结果用于比较和修正。另外,无法适应结构化道路与非结构化道路的切换。本文提出了一种在特征级融合激光雷达与图像数据的方法。首先根据实际道路结构提出了一种应用于两种传感器测量数据的平行直线模型。提出了一种数据融合的方法,根据先验信息与历史数据减少单传感器随机误差对系统检测结果的影响。
[Abstract]:In this paper, the lane departure warning system of fusion image and laser radar is studied. Lane departure warning system is a key technology in the advanced auxiliary driving system, which is mainly used to prevent vehicles from deviating from the current lane and causing traffic accidents. This paper first analyzes the components of the lane departure warning system, summarizes the three difficulties of the current research, and carries out some related research on each difficult point. The main problems and achievements are summarized as follows: 1. lane detection in complex road conditions. The current lane detection algorithm has achieved good results under the condition of stable illumination and good road conditions. The difficulty is to detect lane lines in complex pavement conditions such as illumination change, shadow occlusion, obstructing, and so on. And the lane line is detected under the condition of road wear and lane line as the dotted line. A lane detection algorithm adapted to complex pavement conditions is proposed in this paper. Feature extraction is carried out by double threshold segmentation of gray space and HSV (Hue Saturation Value) space to reduce the loss of features. Aiming at the problem of obstacle occlusion, shadow shelter and other road signs, the gradient direction decision method is applied to reduce the impact of occlusion on Lane Detection Based on the parallel nature of the two edges of the same lane line. In view of the problem of lane line wear and lane line as the dotted line, the improved river algorithm is used to carry out the feature connection. Through the test of the video data collected from the expressway, the algorithm can detect the lane line steadily in a variety of complex road conditions. The 2. lane departure warning model has high false alarm rate and unstable warning information. The lane line deviation from the warning of false alarm will disturb the driver's attention, reduce the availability of the system, or even erroneous to change the state of vehicle movement, causing danger. The instability of early warning information will cause the driver to be unable to judge whether the current driving condition is dangerous or not to correct the vehicle trajectory. In this paper, the advantages and disadvantages of the existing lane departure warning model are analyzed, and a phased deviation early warning model is proposed. On the basis of maintaining the detection rate and the early warning time, the false alarm rate is reduced and the stability of the early warning information is improved. 3. single sensor lane detection reliability is low. Most lane departure early warning studies use visual sensors as the only source of road data, which is less reliable in actual scenarios. On the one hand, when the sensor error occurs, no other measurement results are used for comparison and correction. In addition, it is impossible to adapt to the switching between structured and unstructured roads. In this paper, a method for fusion of laser radar and image data at feature level is proposed. Based on the actual road structure, a parallel straight line model applied to two kinds of sensor data is proposed. A method of data fusion is proposed. According to the prior information and historical data, the effect of the single sensor random error on the detection results of the system is reduced.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN958.98;TP391.41

【相似文献】