Research On Indoor Dynamic Visual Slam And Loop Closure Detection Based On Semantic Segmentation

Intelligent mobile robots provide automation and intelligent solutions for the traditional machinery manufacturing industry,profoundly affecting and promoting the development of the manufacturing industry.SLAM is a key technology for robot intelligence,which utilizes the rich sensor data acquired by robots for simultaneous localization and map construction.The dynamic objects in the scene affect the localization accuracy of SLAM system,and the traditional SLAM map construction lacks semantic information and cannot have higher level understanding of the surrounding environment.An indoor dynamic vision SLAM algorithm combining semantic information was studied to introduce semantic information into ORBSLAM2 loopback detection and map construction tasks,which mainly includes the following parts,namely dynamic point filtering,loopback frame similarity detection combining semantic information and semantic map construction.A dynamic point detection filtering algorithm based on semantic information and geometric methods was studied,which combined Seg-T-Mask network to obtain semantic information,filters out feature points on a priori dynamic targets,and performs dynamic point fine filtering by geometric constraints to improve SLAM algorithm accuracy in dynamic scenes.Comparative tests were conducted on the TUM dataset,and the results showed that the improved algorithm can effectively filter out dynamic feature points in the scene and improve the SLAM localization accuracy,and the dynamic object moving faster scene outperforms the low dynamic scene.In order to verify the performance of the algorithm in real scenes,the dynamic point filtering algorithm was tested in classroom,office and building scenes,and the experiments showed that the dynamic point filtering algorithm can effectively filter out the dynamic feature points in the scenes and reduce the SLAM localization error.A similarity detection method based on semantic information was studied,which combined semantic segmentation to process image frames to obtain their semantic information,constructed image feature vectors and performed matching to solve the similarity,and improved the accuracy of loopback detection to detect the similarity of candidate frames when the scene changes.The algorithm comparison test was conducted on TUM dataset,and the experiments showed that the improved algorithm improves the discrimination of similarity detection and detection accuracy under different scene conditions compared with Dbow algorithm.When the algorithm comparison tests were performed on real scenes,the semantic information-based similarity detection method had better differentiation for similar scenes and improved detection accuracy in exposure scenes.A point cloud map construction thread was added to build a semantic point cloud map with a complete structure.A key frame selection strategy was studied to eliminate redundant key frames and filter out the dynamic overshadowing in the scene by combining semantic information,which reduced the redundancy of the constructed map and increased the expressiveness of the map.The test results of the TUM dataset showed that the constructed maps have a clear and complete structure,do not contain dynamic objects,have less redundancy and are rich in semantic information.Static point cloud maps were constructed for actual scenes such as classrooms,offices and hallways,and the constructed maps have less dynamic ghosting and intuitively reflect the real shape of the experimental scenes.