基于单目相机的VSLAM视觉里程计改进方法

VSLAM Visual Odometry Improvement Method Based on Monocular Camera

经典ORB-SLAM2里程计中的特征提取和剔除误匹配过程中都采用了固定阈值的方法,导致位姿估计的结果不准确,致使后续的建图产生严重偏差。针对此问题,在ORB-SLAM2的基础上,在特征提取环节提出自适应阈值的FAST算法——Adaptive-FAST(Ad-FAST)算法,将设定的最大检测阈值呈梯度式下降至最小检测阈值,每下降一个梯度,进行一次特征点检测,若不满足条件则继续下降阈值,直至满足条件;在剔除误匹配环节提出自适应阈值的RANSAC——Adaptive-RANSAC(Ad-RANSAC)算法,根据输入图像的信息,自适应确定该图像剔除误匹配的阈值。仿真和实验结果表明,Ad-FAST算法解决了固定阈值检测到的特征点中夹杂大量特征不明显的点的问题;Ad-RANSAC算法解决了固定阈值既可能会剔除掉匹配性好的点对,也可能保留了匹配性不好的点对这一问题。

The method of fixed threshold is used in the process of feature extraction and false match elimination in the classic ORB-SLAM2 odometer,which will lead to the inaccurate result of pose estimation and serious deviation in the subsequent mapping.To deal with this problem,on the basis of ORB-SLAM2,a FAST algorithm of adaptive threshold,or the Adaptive FAST(Ad-FAST)algorithm is proposed for feature extraction.The set maximum detection threshold is decreased to the minimum detection threshold in a gradient manner,and the feature points are detected once for each gradient.If the conditions are not met,the threshold continues to decrease until the conditions are met.At the same time,an adaptive threshold algorithm called Adaptive-RANSAC(Ad-RANSAC)is proposed for removing false matches.According to the information of the input image,the threshold of removing false matches of the image is determined adaptively.Simulation and experimental results show that the Ad-FAST algorithm solves the problem that the feature points detected by a fixed threshold are mixed with a large number of non-obvious features.The Ad-RANSAC algorithm solves the problem that the fixed threshold may not only eliminate the good matching point pairs,but also retain the bad matching point pairs.