一种多焦距动态立体视觉 SLAM

现有的双目同步定位与建图(SLAM)都使用标准立体相机,所处环境为静态的假设会影响其在动态环境中的精度。 提 出了一种多焦距动态立体视觉 SLAM 方法,它克服了标准立体相机无法兼顾远距离和宽视场感知场景的缺点,并去除了动态物 体对 SLAM 的影响。 具体来说,对传统的立体校正方法进行了改进,并使用校正参数修正了特征点的位置,而不是整张图像,还 提出了一种自适应特征提取和匹配方法以增加多焦距图像的特征匹配数量。 综合使用多视图几何、区域特征流和相对距离检 测动态对象,剔除动态对象上的特征点。 在公开数据集 KITTI 上,该方法相对 ORB-SLAM3 和 DynaSLAM 的定位精度都提高了 6. 97% ,在自建数据集中,该方法的定位精度比 ORB-SLAM3 提高了 26. 64% ,比 DynaSLAM 提高了 32. 09% 。

A multi-focal length dynamic stereo vision SLAM

The existing stereo simultaneous localization and mapping ( SLAM) methods all use standard stereo cameras, and the assumption of the static environment has influence on their accuracy in dynamic environment. A multi-focal dynamic stereo vision SLAM is proposed. It could overcome the insufficiency of standard stereo cameras that cannot perceive the scene at long distance and wide field of view. The impact of dynamic objects is also removed. To be specific, the stereo calibration method is improved and the calibration parameters are utilized to rectify ORB features instead of rectifying stereo images. For multi-focal stereo images, a feature extraction and matching method is also proposed to increase the number of matched features. Multi-view geometry, regional feature flow and relative distance are used to detect dynamic objects. The feature points on the dynamic objects are eliminated. Compared with ORB-SLAM3 and DynaSLAM, the positioning accuracy of the proposed method on the public data set KITTI is increased by 6. 97% , and the positioning accuracy on the self-made data set is increased by 26. 64% and 32. 09% , respectively.