融合点线特征的双目视觉-惯导SLAM算法

Stereo visual-inertial SLAM algorithm based on point and line features

针对视觉与惯导信息融合的同时定位与地图构建中,因使用场景缺少纹理和相机图像模糊导致的点特征缺失,以及线特征容易错误关联的问题,提出了一种融合点线特征的双目视觉-惯导SLAM算法。该算法首先使用惯导信息辅助系统前端进行点线特征数据关联,跟踪相机位姿;然后,在系统后端根据点线特征的丰富程度对点线特征的重投影残差进行加权,采用非线性优化的形式进行状态估计,并为系统前端提供局部地图信息;最后,使用点线综合的视觉词典进行回环检测,当检测到回环时,对系统中各状态量进行全局优化。基于EuRoC数据集以及KITTI数据集的实验结果表明:该双目视觉-惯导SLAM算法可以有效剔除线特征误匹配,提高系统前端的跟踪精度,而点线综合的视觉词典可以提升系统回环检测的准确率。

To solve the problem of point features missing and incorrect data association of line features,as well as the lack of textures of scene and image blurring in simultaneous localization and mapping(SLAM)based on visual-inertial information fusion,a stereo visual-inertial SLAM algorithm based on point and line features was proposed.Firstly,in the front-end of the algorithm,the inertial navigation information was used to associate the point and line features to track the camera pose.Then,in the back-end,the re-projection residuals of the point and line features were weighted accor-ding to the richness of the point and line features.As a result,the state estimation was carried out in the form of nonlinear optimization,and the local map information was provided for the front end of the system as well.Finally,loop detection was carried out using the visual dictionary of point and line features.When the loop was detected,the state variables in the system were optimized globally.The experiments on the open dataset EuRoC and KITTI show that the VI-SLAM algorithm can effectively eliminate the mismatching of line features,so as to improve the tracking accuracy of the front-end,whereas the visual dictionary of point and line features can improve the accuracy of loop closure detection.