Semantic SLAM Based On Topic Model

With the rapid development of artificial intelligence,simultaneous localization and mapping of robots(SLAM)has attracted more and more attention as a research direction in the field of artificial intelligence.Simultaneous localization and mapping of robots is the core technology of robot applications,covering a wide range of fields,from smart home robots to autonomous driving technologies,from augmented reality to virtual reality technology.At present,traditional visual SLAM technology can no longer meet the needs of robot applications.Semantic SLAM with scene understanding is one of the future development directions of SLAM technology.Traditional visual SLAM is prone to failure in environments with repeated patterns or less textures,and it is not tolerant of dynamic objects in the scene.Often,tracking failures occur due to the presence of dynamic objects.This paper proposes a new semantic vision SLAM.Compared with traditional visual SLAM,it is more robust to dynamic objects,and can build semantic maps with object levels,which is easier to be extend to other robot applications.The problem of object association is an inevitable problem when introducing semantic information into traditional SLAM.We use deep learning methods to combine object detection and object pose estimation into visual SLAM based on feature points.Semantic information provides constraints for SLAM motion estimation,SLAM pose estimation promotes object association in return,so the two promote each other.Correct object association is the key factor for successful object SLAM system,because object association and SLAM are inherently coupled.The main work and results of this article are as follows:1.Aiming at the problem of object association,a novel formulation of object association based on hierarchical Dirichlet process(HDP)was proposed.Through the HDP,we can hierarchically partition the environment into separated subspace and only associate objects in the same subspace.This can improve the accuracy and computational efficiency of object associations.Thanks to the novel formulation,the wrong object association can be corrected.2.The object pose which is 6DOF is introduced into the back-end optimization process.The problems of object association and pose optimization are handled in a tightly coupled way,so that these two aspects can promote each other.The proposed method is evaluated on indoor and outdoor datasets.Experimental results show that compared with traditional visual SLAM,the proposed method has a 15%-19% improvement in motion estimation accuracy,and with semantic information our system's robustness has also been greatly improved.3.Implement a complete semantic SLAM system.Combining the object semantic information acquisition module,hierarchical topic model-based object association algorithm and semantic information constraints optimization into the ORB-SLAM2 open source framework,the semantic vision SLAM system in this paper can estimate camera poses in real time and reconstruct sparse semantic maps.