大曲率运动下基于自适应关键帧选取的动态SLAM算法

Dynamic SLAM algorithm based on adaptive keyframe extraction under large curvature motion

针对传统同步定位与建图(SLAM)算法在大曲率运动下易发生漏选关键帧、在动态场景下无法准确判断物体运动状态等问题,提出一种大曲率运动下基于自适应关键帧选取的动态SLAM算法。为避免弯道运动漏选关键帧,创建局部逆向索引窗口并依据窗口内关键帧、当前帧、参考帧三者中的特征点数量、匹配点数量、区域变化空间点数量计算自适应阈值,增补在弯道运动中遗漏的关键帧,提升了算法定位精度。同时,为避免动态点造成关键帧选取阈值不准确,设计基于视差角模型的动态物体判断策略来估计潜在动态物体的运动状态。在公开数据集和真实场景进行验证,结果表明:与DynaSLAM算法相比,所提算法在TUM数据集的平均绝对轨迹误差减少了20%,在室内和室外大曲率动态场景下的定位精度分别提升了12.1%和15.3%,展现出良好的建图能力。

An adaptive keyframe election-based dynamic simultaneous localization and mapping(SLAM)algorithm under large curvature motion is proposed to address the problems of traditional SLAM algorithms,such as missed keyframe election in large curvature motion and failing to accurately judge the motion state of objects in dynamic scenes.In order to avoid the missing keyframes in the curve movement,a local reverse index window is created and the adaptive threshold is calculated based on the number of feature points,matching points and spatial points of regional changes in the keyframe,the current frame and the reference frame in the window.The missed key frames in the curve motion are added to improve the positioning accuracy of the algorithm.Meanwhile,in order to avoid inaccurate thresholds for keyframe selection caused by dynamic points,a dynamic object judgment strategy based on parallax angle model is designed to estimate the motion state of potential dynamic objects.Tested on public datasets and real scenes,the results show that compared with DynaSLAM algorithm,the average absolute trajectory error of the proposed algorithm is reduced by 20%in the TUM dataset,and the positioning accuracy is improved by 12.1%and 15.3%respectively in indoor and outdoor dynamic scenes with large curvature,which demonstrates a good mapping ability.