动态遮挡场景下基于改进Transformer实例分割的VSLAM算法

Improved Transformer Instance Segmentation Under Dynamic Occlusion Based VSLAM Algorithm

针对传统SLAM(Simultaneous Localization And Mapping)算法在动态遮挡场景下难以标记被遮挡物体,无法准确判断潜在物体运动状态以及剔除动态物体后特征点数量较少等问题,提出一种动态遮挡场景下基于改进Transformer实例分割的VSLAM算法(Improved Transformer instance segmentation under Dynamic occlusion VSLAM algorithm, ITD-SLAM).本算法通过设计一种多注意力模块,引导模型关注被遮挡区域,同时改进相对位置编码优化被遮挡物体边界语义性,精确标记出潜在动态物体.为减少动态物体对SLAM系统定位精度的影响,通过相机位姿估计、物体运动估计与物体运动判断三个步骤估计潜在动态物体运动状态,并剔除其中的动态物体.根据网格流运动模型补全剔除区域的静态背景,并利用信息熵与交叉熵筛选修复区域特征点,补充高质量特征点用于相机位姿估计.在公开数据集TUM和真实场景中进行验证,结果表明本文算法均方根误差与DynaSLAM相比减少22.94%,表现出了较好的构图能力.

For traditional SLAM(Simultaneous Localization And Mapping)algorithms,it is difficult to mark occlud⁃ed objects in dynamic scenes with occlusion,and is impossible to accurately judge the motion state of potential objects as well as the number of feature points after culling dynamic objects is small.This paper proposes a VSLAM algorithm based on improved transformer instance segmentation under dynamic occlusion(ITD-SLAM)in dynamic occlusion scenarios.By designing a multi-attention module,this algorithm guides the model to pay attention to the occluded area,and at the same time improves the relative position encoding to optimize the boundary semantics of occluded objects,and accurately mark potential dynamic objects.In order to reduce the influence of dynamic objects on the positioning accuracy of the SLAM sys⁃tem,the motion state of potential dynamic objects is estimated through three steps of camera pose estimation,object motion estimation and object motion judgment,and dynamic objects are eliminated.According to the grid flow motion model,the static background of the culled area is completed,and the feature points of the repair area are screened and repaired by infor⁃mation entropy,and the high-quality feature points are supplemented for camera pose estimation.Experimental results on the public datasets show that this algorithm has better composition ability with its root mean square error reduced by 22.94%when compared with DynaSLAM.