VSLAM系统中的快速特征点匹配算法

A Fast Feature Point Matching Algorithm in VSLAM System

摘要:

连续图像帧之间特征点匹配是视觉同步定位与建图（VSLAM）技术的关键技术。针对VSLAM系统中连续图像帧之间特征点匹配耗时且精度低的问题，一种基于局部像素运动模型的快速特征点匹配算法（LPMM）被提出。该算法基于运动平滑约束假设，利用连续图像帧中局部区域内的像素点的运动具有一致性这一特性，将图像划分成局部网格区域后利用每个网格中的部分特征点估算该网格区域的运动向量，并以此为基础，计算下一帧匹配特征点的搜索范围的中心点；最终在计算出下一帧的中心点的局部邻域内搜索与当前帧的特征点相匹配的特征点。实验结果表明，所提出的匹配算法与广泛使用的ORB-SLAM2中的投影匹配算法相比，平均匹配速度提高50%以上，匹配精度提升了约4%。

Abstract:

Feature point matching between consecutive image frames is a key technology in Visual Simultaneous Localization and Mapping (VSLAM). To address the issues of time-consuming and low accuracy in feature point matching between consecutive image frames in VSLAM systems, a fast feature point matching algorithm based on local pixel motion model (LPMM) is proposed. This algorithm is based on the assumption of motion smoothness constraint, utilizing the consistency of pixel motion within local regions in consecutive image frames. By dividing the image into local grid regions and estimating the motion vector of each grid using some feature points within the grid, the algorithm calculates the center point of the search range for matching feature points in the next frame. Finally, it searches for feature points that match the feature points in the current frame within the local neighborhood of the calculated center point in the next frame. Experimental results show that the proposed matching algorithm, compared with the widely used projection matching algorithm in ORB-SLAM2, has an average matching speed increase of over 50% and an accuracy improvement of approximately 4%.