基于单目视觉的车辆前方障碍物测距方法  

Distance measurement method based on monocular vision for obstacles in front of vehicles

作　　者：高维岳 陈宇拓[1] 刘洋 陈标 GAO Wei-yue;CHEN Yu-tuo+;LIU Yang;CHEN Biao(College of Computer and Information Engineering,Central South University of Forestry and Technology,Changsha 410004,China;Department of Information Engineering,Hunan Automobile Engineering Vocational College,Zhuzhou 412001,China)

机构地区：[1]中南林业科技大学计算机与信息工程学院,湖南长沙410004 [2]湖南汽车工程职业学院信息工程学院,湖南株洲412001

出　　处：《计算机工程与设计》2022年第4期1016-1022,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(11771453);湖南省教育厅科学研究项目优秀青年基金项目(18B566)。

摘　　要：针对行车过程中的防碰撞预警问题,提出一种基于单目视觉车辆前方障碍物检测与测距方法。为解决传统车辆检测泛化性差且人工提取特征不准确问题,通过深度学习目标检测YOLOv4算法对车辆前方多种障碍物进行检测,获取障碍物的类别信息与位置信息。运用改进的边缘检测算法调整检测框的位置,提升检测算法目标定位的准确性。根据摄像机成像原理及几何关系,得到路面三维坐标与像平面二维坐标转换模型从而进行测距,对所得测量数据进行三次曲线拟合、对测距过程和算法进行优化提升测距精度。在50 m范围内平均误差为0.54 m,在80 m范围内平均测距误差为0.78 m。实验分析对比结果表明,所提方法能够实现较精准、高效率的单目视觉测距。To solve the anti-collision warning problem during driving,a method for detecting and ranging obstacles in front of vehicles based on monocular vision was proposed.To solve the problems of poor generalization of traditional vehicles detection and inaccurate manual extraction of features,the YOLOv4 algorithm for deep learning object detection was used to detect multiple obstacles in front of the vehicle and to obtain the category information and location information of the obstacles.To improve the accuracy of target localization of the detection algorithm,the improved edge detection algorithm was applied to adjust the position of the detection frame.According to the camera imaging principle and geometric relationship,the conversion model from the three-dimensional coordinates of the road surface to the two-dimensional coordinates of the image plane was obtained and distance measurement was performed.To improve the distance measurement accuracy,the cubic curve fitting of the obtained mea-surement data was performed,and the distance measurement process and algorithms were optimized.The average error in the range of 50 m is 0.54 m,and the average error in the range of 80 m is 0.78 m.Through experimental analysis and comparison,the results show that the proposed method can achieve accurate and effective monocular vision ranging.

关 键 词：单目视觉 目标检测 YOLOv4算法 障碍物测距 三次Bezier曲线拟合 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]