基于L-junction特征的路面箭头标记检测和识别

发布时间：2018-05-25 20:44

本文选题：箭头检测 + L-junction　；参考：《华中师范大学》2015年硕士论文

【摘要】：随着智能汽车的发展,自动驾驶技术受到广泛的关注,其中感知周围环境和路面信息是自动驾驶技术关键的部分。路面箭头标记作为自动驾驶系统感知信息的重要部分,实现对它的智能检测和识别具有重要的意义。由于箭头检测过程中经常受到箭头标记自身磨损、外界多种非箭头标记等因素的干扰,对它的检测和识别往往不能达到满意的效果。本文针对箭头路面标记的检测和识别进行了有效的研究。本文基于L-junction结构将路面上的箭头标记用多个L-junction构成的L-junction链描述,使干扰因素对检测的影响转化为L-junction偏差。然后提出了三种编码方法将L-junction链编码为字符串,将一定偏差范围内的L-junction编码为同一个字符码,提高检测模型对磨损箭头标记的包容性；同时运用箭头标记上L-junction的角度和拓扑关系对检测标记设置条件,排除非箭头标记的干扰。为了判断检测标记是否为箭头标记,本文提出了权重编辑距离算法。该算法对检测标记不同的偏差赋予不同的权重,实现检测标记编码字符串与标准箭头标记编码字符串之间的相似性的计算,提高检测模型检测准确率。为了测试该模型的检测效果,在北京五环高速公路上收集了三个路面图像集：清晰箭头标记图像集、磨损箭头标记图像集、视频图像集。采用本文的检测模型分别检测三个图像集中的图像,并统计检测的准确率和召回率。为了更综合地衡量本文模型的检测效果,本文还运用了深度学习模型:Boosting Convolutional Deep Neural Network (CDNN)来对这三个图像集进行检测。大量的检测结果显示,本文的检测模型性能较稳定,取得较高的准确率和召回率。
[Abstract]:With the development of intelligent vehicles, self-driving technology has been paid more and more attention, in which sensing the surrounding environment and road information is the key part of autopilot technology. As an important part of the sensing information of autopilot system, road arrow marking is of great significance to its intelligent detection and recognition. The detection and recognition of arrowhead are often not satisfactory due to the wear and tear of arrowhead markers and many outside factors such as non-arrowhead markers in the process of arrow detection. In this paper, the detection and recognition of arrowhead road markings are studied effectively. Based on the L-junction structure, the arrowhead marks on the road surface are described with L-junction chains composed of multiple L-junction, so that the influence of the interference factors on the detection is transformed into the L-junction deviation. Then three coding methods are proposed to encode the L-junction chain as a string and the L-junction in a certain range of deviations as the same character code to improve the tolerance of the detection model to the frayed arrow marks. At the same time, the angle and topological relation of L-junction on arrow mark are used to set the condition of detection mark, and the interference of non-arrow mark is eliminated. In order to determine whether the detection marks are arrow marks, this paper proposes a weighted edit distance algorithm. The algorithm gives different weights to the different deviations of the detection marks, realizes the calculation of the similarity between the detection tag encoding strings and the standard arrow tag coding strings, and improves the detection accuracy of the detection model. In order to test the effectiveness of the model, three road image sets were collected on Beijing Wuhuan Expressway: clear arrow marking image set, wear arrow marking image set and video image set. The detection model of this paper is used to detect the images of three image sets, and the accuracy and recall rate of the detection are calculated. In order to measure the effectiveness of our model more synthetically, we also use the depth learning model: boosting Convolutional Deep Neural Network to detect these three image sets. A large number of test results show that the performance of the detection model is stable, and high accuracy and recall rate are obtained.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U463.6;U495;TP391.41

【相似文献】