融合语义信息与物体级几何特征的实时动态激光SLAM算法

Real-Time Dynamic Laser SLAM Algorithm Combining Object-Level Geometric Features and Semantic Information

针对现有的激光SLAM算法在动态场景下鲁棒性差、定位与建图精度易受动态物体干扰的问题,提出了一种融合语义信息与物体级几何特征的实时动态激光SLAM算法Object-SuMa。首先通过地面滤波、物体分割、方向包围盒解算等过程,生成物体级几何特征并表示为纹理,用于并行计算和修正物体内部错误的语义分割结果;然后在配准过程中分解计算方向包围盒间的IOU,并基于包围盒IOU和语义分割结果引入物体级几何加权和语义加权,减少误匹配和动态点匹配;借助图形渲染管线构建并行计算过程,进行地面点配准和非地面点配准两步优化,降低计算复杂度和耗时;最后在KITTI里程计数据集测试表明,Object-SuMa算法相比SuMa++算法的相对位姿估计精度提升15%,ICP平均耗时下降17%,改善了动态场景下的激光SLAM定位精度和鲁棒性。

In view of the problems of the existing laser SLAM algorithm in dynamic scenes,which has poor robustness and the positioning and mapping accuracy is easily disturbed by dynamic objects,a real-time dynamic la-ser SLAM algorithm called Object-SuMa that combines object-level geometric feature and semantic information is proposed.Firstly,through processes such as ground filtering,object segmentation and pose size calculation,object-level geometric features are generated and represented as texture and used to correct semantic segmentation errors within the object.Then,in the odometry stage,the IOU calculation of the oriented bounding box is decomposed,and object-level geometric weighting and semantic weighting are introduced based on the bounding box IOU and se-mantic segmentation results to reduce mismatching and dynamic point matching.In addition,the graphics rendering pipeline is used to build a parallel computing process,and the computational complexity and time consuming are re-duced by two-step optimization of ground point registration and non-ground point registration.Finally,tests on the KITTI odometry data set show that compared with SuMa++,the Object-SuMa algorithm has improved the relative pose accuracy by 15%and reduced the average time of ICP by 17%,which improves the positioning accuracy and robustness of laser SLAM in dynamic scenarios.