一种基于GMS的ORB-SLAM3特征匹配改进算法

An improved ORB-SLAM3 feature matching algorithm based on GMS

针对ORB-SLAM3中ORB特征点匹配准确率低的问题,本文提出了一种改进的特征点匹配策略。首先,考虑到特征点提取与匹配会受到场景昏暗与对比度太低的影响,对昏暗的场景数据集进行对比度增强和去噪。其次,为了提高特征点匹配的数量及速度,将运动平滑性约束作为去除特征点错误匹配的依据,舍弃旋转不变性和尺度不变性并将图片转换为3×3网格来加速运算。最后,为了提高特征匹配的精度,考虑距离与置信度之间的关系,通过计算特征点邻域内的匹配数量,与设定的阈值进行对比,筛选出正确的特征匹配,之后再进行相机的位姿估计等视觉里程计算,估计出相机的移动路径。通过实验分析,该方法能够提高约72.8%的正确特征点匹配数量,并比原本的匹配时间减少约9%。在对RGB-D数据集和Euroc数据集进行实验后,面对昏暗的数据集,定位精度分别提升约21.20%和63.67%。与其他对比方法相比较,该方法不仅增强了系统的处理速度和鲁棒性,也使平均定位精度有所提升。

Aiming to solve the problem of low accuracy of ORB feature point matching in ORB-SLAM3,this paper proposes an improved feature point matching strategy.First,considering that feature point extraction and matching are affected by dim scenes and low contrast,this paper enhances the contrast and denoises the datasets of dim scenes.Second,in order to increase the number and speed of feature point matching,motion smoothness constraints are used as the basis for removing incorrect feature point matches,abandoning rotation invariance and scale invariance,and converting the image into a 3x3 grid to accelerate computation.Finally,to improve the accuracy of feature matching,this paper considers the relationship between distance and confidence by calculating the number of matches within the neighborhood of feature points and comparing it with a set threshold to filter out correct matches.Subsequent,camera pose estimation and visual odometry calculations are performed to estimate the camera's movement path.Through experimental analysis,this method is shown to increase the number of correct feature point matches by approximately 72.8%and reduce the matching time by about 9%.After conducting experiments on RGB-D and Euroc datasets,the localization accuracy for dim datasets increased by approximately 21.20%and 63.67%,respectively.Compared to other methods,this approach not only enhances the system's processing speed and robustness but also improves average localization accuracy.