融入语义信息的VSLAM研究综述

Review of VSLAM Research with Semantic Information

随着实际应用需求的不断提高,传统视觉同步定位与地图构建(visual simultaneous localization and mapping,VSLAM)难以建立符合要求的环境地图。语义VSLAM在传统VSLAM基础框架上引入深度学习获取环境图像的语义信息,并且构建环境的3D点云语义地图,能为更高层次的交互任务提供有效支持。通过分析传统的VSLAM框架中视觉里程计、回环与优化以及建图上的不足,阐述了近年来使用语义信息改善VSLAM系统性能及为环境交互任务提供助力的方法。最后,结合研究现状结论及课题组正在开展的工作对语义VSLAM未来研究方向及应用进行展望。

With the increasing demand of practical applications,it is difficult to build a compliant environment map with traditional visual simultaneous localization and mapping(VSLAM).Semantic VSLAM introduces deep learning to obtain semantic information of environment images based on its traditional VSLAM,and constructs 3D point cloud semantic maps of the environment,which can provide effective support for higher-level interaction tasks.In this paper,by analyzing the shortcomings of the traditional VSLAM framework in visual odometry,loop detection,nonliear optimization and mapping,we describe the recent approaches of improving the performance of VSLAM systems and providing a boost for environmental interaction tasks using semantic information.Finally,the research findings and the ongoing work of our group are combined to give an outlook on the research direction and future application development of semantic VSLAM.