基于改进RANSAC算法的雾天自动驾驶汽车视觉图像配准方法

Visual image registration method of autonomous vehicle in fog based on improved RANSAC algorithm

针对雾天中依靠视觉SLAM的自动驾驶车辆图像匹配率较低的问题,通过模拟不同雾天中的行车场景,对三种特征点检测和描述算法进行了性能测试,得出采用SURF检测子结合SIFT描述子在雾天中鲁棒性和配准率较强,并通过FLANN算法完成粗匹配。同时对RANSAC算法进行了改进,通过设立测试集和自适应降低阈值的方法改善错误单应矩阵造成的运算效率较低,并采用自适应采样次数的方法解决算法迭代次数的上限问题,完成图像配准。在DAIR-V2X数据集上的实验结果表明:改进的RANSAC算法能够相对最优地得到正确匹配点对,且相较于PROSAC和传统的RANSAC算法,运行速度分别提升了18.5%与42.7%。

Aiming at the low matching rate of automatic driving vehicle image relying on visual SLAM in fog days,three feature point detection and description algorithms are tested by simulating driving scenes in different fog days.It is concluded that SURF detector combined with SIFT descriptor has better robustness and registration rate in fog days,and rough matching is completed by FLANN algorithm.At the same time,it improves the RANSAC algorithm,improves the low operational efficiency caused by error homography matrix by setting test set and adaptive lowering threshold,and uses the method of adaptive sampling times to solve the problem of upper limit of iteration times and complete image registration.The experimental results on DAIR-V2X data set show that the improved RANSAC algorithm can obtain the correct matching point pairs relatively optimally,and the running speed is increased by 18.5%and 42.7%respectively compared with PROSAC and traditional RANSAC algorithm.