为实现视觉里程计在大规模复杂环境下的准确导航, 消除载体位姿估计的累计误差, 提出了一种基于几何约束的视觉闭环检测和位姿优化方法。首先建立基于LATCH二进制描述子的视觉词袋(BoVW), 并采用视觉单词向量描述图像进而归一化相似度的闭环检测方法, 然后对闭环候选关键帧和当前关键帧进行基于RANSAC-HORN运动估计的闭环验证, 最后将闭环关键帧的局部地图点投影到当前帧并最小化重投影误差以使得位姿得到优化。数据集实验表明, 提出的闭环检测和位姿优化算法能够有效地对闭环进行准确检测和验证, 并对视觉里程计的误差积累进行闭环位姿优化, 提高了视觉导航精度。

In order to realize the accurate navigation of visual odometers in the large-scale complex environment and eliminate the cumulative error of pose estimation, this paper presents a method of visual loop closure detection and pose optimization based on geometric constraints.Firstly, a Bag of Visual Words(BoVW) based on LATCH descriptor is set up, and the loop closure detection method is adopted with the images described by the visual word vector and the similarity degree normalized.Then, the loop closure between the candidate keyframe and the current keyframe is checked based on RANSAC-HORN motion estimation.Finally, the local map points of the loop closure keyframe are projected to the current frame, and the re-projection error is minimized to optimize the pose.The experimental results show that the loop closure detection and pose optimization algorithm proposed in this paper can effectively and accurately detect and verify the loop closure, carry out the loop closure pose optimization of the error accumulation of the visual odometer and improve the accuracy of visual navigation.