基于ORB-SLAM3的改进型特征匹配与稠密建图算法

针对传统ORB算法存在提取的特征点极易堆积在纹理丰富的区域及误匹配率高等而导致无法满足高精度定位要求，以及ORB-SLAM3系统无法构建稠密地图的问题，提出一种基于ORB-SLAM3的改进型ORB-GMS特征匹配方法，并增加稠密建图线程来实现稠密地图的构建。首先，在特征点提取过程中引入四叉树策略，将图像帧分为若干个网格，并在每个网格中提取最优特征点；然后，在特征匹配过程中将运动平滑约束转换为剔除错误匹配的统计量，通过比较匹配对邻域内的匹配对数量和阈值来快速筛选正确匹配；最后，完成位姿估计并根据关键帧与相应位姿完成稠密点云地图的构建。采用TUM的RGB-D数据集进行实验，改进算法提取的特征点相较传统ORB算法分布更加均匀，匹配数比ORB-SLAM3增加64.5%,比GMS算法增加4.7%,匹配耗时比ORB-SLAM3减少20.4%,比GMS算法减少94.6%,从而验证了改进算法在特征点提取与匹配方面的优越性，并且相较于ORB-SLAM3,改进算法的定位精度提高了3.75%,从而验证了其在总体上提高定位精度，进而实现稠密建图的可行性和有效性。

Improved feature matching and dense-mapping algorithm based on ORB-SLAM3

In view of the problems that the extracted feature points of the traditional ORB algorithm tend to accumulate in the texture-rich area and the high false matching rate, which cannot meet the requirements of high-precision positioning, and that the ORB-SLAM3 system cannot build a dense map, this paper proposed an improved ORB-GMS feature matching method based on ORB-SLAM3, and added dense-mapping thread to realize the construction of dense maps. Firstly, the feature points extraction process adopted the quadtree strategy to divide the image frame into several meshes, and extracted the best feature points in each mesh. Then, it replaced the motion smoothing constraint with a statistic that rejected incorrect matches during feature matching, and used a comparison of the number of matching pairs and thresholds within the neighborhood of matching pairs to quickly filter correct matches. Finally, it completed the positional estimation and constructed a dense point cloud map by keyframes and corresponding poses. Testing by RGB-D dataset from TUM, the improved algorithm can extract uniform feature points, the number of matches increases by 64.5% than ORB-SLAM3, increases by 4.7% than the GMS algorithm, and the matching elapsed time decreases by 20.4% than ORB-SLAM3, and decreases by 94.6% than the GMS algorithm, which proves that the improved algorithm is superior in feature point extraction and matching. And compared to the ORB-SLAM3, the accuracy of the improved algorithm increases by 3.75%, thus, demonstrating its feasibility and effectiveness in improving the localization accuracy and building dense maps.