Research On Visual SLAM Algorithm Based On Dynamic Information Filtering

Simultaneous localization and map construction technology(SLAM)makes it possible for intelligent mobile robots to autonomously complete related tasks in unknown scenes,which is one of the current research hotspots.However,the common SLAM system usually assumes that its application environment is a static environment,which makes the SLAM system sensitive to moving objects.When running in a dynamic environment,there will be problems such as large error in pose estimation and low accuracy,and at the same time,the constructed environment map has ghosting,which is difficult to reuse.In order to eliminate the influence of dynamic information,this paper designs a visual SLAM method applied to dynamic environments based on the ORB-SLAM2 algorithm.Experiments show that the algorithm in this paper has better processing effect on dynamic objects.Compared with the ORB-SLAM2 algorithm,the RMSE value of the absolute trajectory error of the system is reduced by on average of 90.52%in a high dynamic environment,and the RMSE value of the absolute trajectory error in a low dynamic environment is reduced by an average of 27.03%.The pose estimation accuracy is high,and at the same time,the construction of the environment static map is well realized.The main research contents are as follows:Firstly,aiming at the problem of accurate identification of dynamic points,a method for identifying and filtering out dynamic feature points based on target detection is designed.In this paper,the inter-class variance is used to re-define the inliers,and then the random sampling agreement with the most inliers(RANSAC)model is used to obtain the accurate transformation matrix,and the multi-view geometric algorithm is used to pre-determine the dynamic points.Then,the YOLOv3-Tiny algorithm is used for the pre-selection of dynamic regions,and the improved re-discrimination fusion algorithm is used to process only the real dynamic points to achieve accurate filtering of dynamic points.Secondly,in view of the huge error of pose estimation in dynamic scenes,a dynamic SLAM algorithm is designed,which combines the loop closure detection algorithm based on visual dictionary and the BA algorithm for system optimization.At the same time,in view of the problem that there is more redundant information between frames,the key frame screening method is improved based on redundancy detection,and the secondary judgment of key frames is carried out by using common viewpoints between frames,which ensures low redundancy and high precision of system operation.The qualitative and quantitative analysis of the comparative experiments on the TUM data set verifies the improvement of the performance of the algorithm.Finally,aiming at the problem of ghosting in the construction of dense maps in a dynamic environment,combined with the YOLOv3-Tiny algorithm,the object category information is used for processing,and a three-dimensional dense map of the static environment is constructed according to the filtered drawing keyframes.Experiments show that the algorithm in this paper effectively removes the dynamic objects in the environment and realizes the construction of a static environment map,which increases the reusability of the map.Then,experiments are carried out in real indoor dynamic scenes to verify the effectiveness and robustness of the proposed algorithm.