Semantic Mapping Optimization Of Visual SLAM

For mobile robots,the traditional simultaneous localization and map-ping(SLAM)has provided a good basis for positioning and mapping.Mobile robots can estimate their posture through visual odometers,perform back-end optimiza-tion through filtering and nonlinear optimization algorithms,estimate the state of the entire system from noisy data,and perform sparse or dense map construction as needed.However,the performance factors of all aspects of visual SLAM currently lead to very limited application scenarios.This paper selects ORB-SLAM2 as the basis of visual SLAM,improves the positioning accuracy of ORB-SLAM2 by fus-ing semantic information,and proposes a method for constructing topological maps to purposefully construct objects in the environment.In this paper,the target detection algorithm is applied to the RGB image cor-responding to the key frame selected by ORB-SLAM2 through the target detection algorithm to obtain the object-level semantic information corresponding to each key frame.Build a bag-of-words model for key frames and object-level objects,filter the key frames by judging the dynamic influencing factors of each key frame,and remove the dynamic signpost points corresponding to the reserved key frames to assist ORB-SLAM2 to obtain more Stable key frames and signpost points.For the loopback detection of ORB-SLAM2,this paper proposes a closed-loop detection method based on object-level semantic information for some specific environments.This method is based on the bag-of-words model,using uniquely recognizable ob-jects as words,matching each newly observed word with existing words,and com-pleting closed-loop detection.In this paper,the topology map is used to construct the SLAM map.The object-level objects are used as landmarks,and the posture of the corresponding objects of each landmark is obtained through the 6D target pose estimation algorithm,which provides the basis for the reconstruction of the built map in other environments.In the end,this paper conducted verification experiments and performance tests on the improved SLAM scheme,and proved that the improved scheme improved the positioning accuracy of ORB-SLAM2 in a dynamic environment,and repeated the built topology map in Unreal Engine 4,and the expected construction effect has been achieved.