| The semantic perception of the environment by mobile robots while self-location in an unknown environment is the frontier of the field of robotics research,Semantic Simultaneous Localization and Mapping(Semantic SLAM)as a universal solution for this direction,which combine the SLAM algorithm and the scene semantic under-standing to help robots describe the environment from both geometric and semantic,providing preconditions for advanced tasks such as robotic autonomous movement and behavioral decision making.However,the majority of research about Semantic SLAM to date require a priori known 3D object models or re-volve around mapping with a few object categories but neglect separating object individual.Recent years,the Semantic SLAM based on object detection has attracted wide attention because its ability to dis-tinguish object individual.In these methods of Semantic SLAM,some use the point cloud segmentation to extract object that ignore the tracking of inter-frame objects,others are matching the objects with the samples in the database to track the objects,which have large search scope and high algorithm complexity.To deal with these problems,a new way of Semantic SLAM is proposed in this pa-per.Firstly,in order to realize the SLAM system of instance-level environment seman-tic perception,visual SLAM and instance segmentation algorithm are effectively fused.First,the objects in each keyframe are obtained by instance segmentation.Then,the inter-frame object is tracked and data-associated based on the online object database,and the keyframe is instance-level semantic annotated.Finally,the keyframe carrying semantic annotation is mapped to construct semantic maps.Secondly,aiming at the problem of excessive search range when inter-frame objects are long-term tracking,this paper proposed an inter-frame object matching and track-ing method based on common-view of keyframe.For the objects detected in each keyframe,the corresponding candidate instance set is extracted according to the com-mon-view to match the target instance set.It not only considers the historical data in the Semantic SLAM process,but also narrows the scope of search and improves the time efficiency of object trackingThirdly,according to the influence of dynamic objects on the positioning accuracy of Semantic SLAM,the dynamic objects in SLAM process are detected based on se-mantic information and the corresponding dynamic feature points are removed ac-cording to the detection results to improve the geometric consistency of feature points matching.Meanwhile,the inter-frame object tracking results are incorporated into the pose optimization.The pose estimation is optimized by fusion object matching and feature point matching,which solve the problem of large estimation error caused by sparse feature points in texture smoothing scene,improving the accuracy and robust-ness of the pose estimation of the Semantic SLAM.Finally,in order to verify the effectiveness of this Semantic SLAM system,the functional modules of Semantic SLAM such as location,inter-frame object tracking and semantic mapping are experimentally tested and verified respectively in real scene and dataset.Meanwhile,the pose optimization method based on semantic information is experimentally evaluated in the dataset,The experimental results show that the Se-mantic SLAM system can self-locate in the unknown environment and construct the instance-level semantic map of the environment,which verifies the feasibility and ac-curacy of the Semantic SLAM system. |
PDF Full Text Request