Research And Implementation Of Improved Transformer-based VSLAM Algorithm In Dynamic Scenes

Simultaneous Localization And Mapping(SLAM)is a mobile robot that operates in an unknown environment,obtains environmental information through a variety of sensors carried by itself,and realizes its own positioning and environment map construction.SLAM algorithms are mainly divided into two categories:laser SLAM and visual SLAM,according to the sensors they carry.Due to the disadvantages of laser SLAM in obtaining the color,texture and semantic information of the surrounding environment,visual SLAM has gradually become the mainstream direction of research.However,most of the visual SLAM algorithms at this stage rely on the assumption that the environment is static and unchanged.The presence of moving objects in the environment can easily cause the failure of the visual SLAM algorithm.Therefore,how to eliminate the impact of dynamic objects on the positioning and composition of the visual SLAM system is an urgent problem to be solved.Aiming at the problem that the traditional visual SLAM algorithm is difficult to accurately judge the motion state of feature points in dynamic blurred scenes,this paper improves a dynamic VSLAM method based on motion differential entropy.In the feature extraction process,the method obtains features by using images of different scales by improving the multiscale neural network,combining multi-scale features to gradually clear the blurred image,and extracting ORB feature points at the same time.In the feature point motion judgment link,the potential dynamic object area is determined through deep clustering,the motion state of the feature point is judged according to the differential entropy value in the nonlinear attitude optimization stage,and the dynamic feature point is eliminated to reduce the positioning error of the system.Aiming at the problems that the traditional visual SLAM algorithm is difficult to accurately mark occluded objects in dynamic scenes,and there is less information for positioning and composition after removing dynamic objects,this paper improves a VSLAM method based on instance segmentation and background restoration.In the object segmentation process,the method integrates the improved multi-attention mechanism and position coding into the Transformer segmentation network to optimize the segmentation of occluded objects,determines the area where the potential dynamic object is located according to the instance segmentation result,and judges and eliminates the area where the dynamic object is located by the motion differential entropy.In the background restoration process,by improving the mesh flow network model,the previous frame information is used to repair the static background of the culled area,and the feature points are optimized to increase the information used for positioning and composition.In order to verify the effectiveness of the algorithm proposed in this paper,the dynamic VSLAM method based on motion differential entropy and the dynamic VSLAM method based on instance segmentation and background restoration were tested on the public TUM dataset,and compared with mainstream ORB-SLAM2,DS-SLAM,DynaSLAM algorithms Comparative analysis verifies the positioning accuracy and composition effect of the proposed method in dynamic scenes.It is used on the built hardware platform to prove the feasibility of this algorithm and provide a reference for indoor dynamic visual SLAM research.