Research On Visual SLAM Of Complex Indoor Scenes Based On RGB-D

Simultaneous Localization and Mapping(SLAM)is one of the key technologies for intelligent mobile robots to achieve intelligent tasks.Although there are many excellent visual SLAM solutions available,many algorithms have certain shortcomings when facing indoor complex scenes.In static environments,there are problems of poor real-time performance and low accuracy;as well as in dynamic environments,where the presence of moving objects leads to motion estimation errors and extremely poor localization accuracy.In this paper,the visual SLAM algorithm is improved to solve the above two problems,and the improved algorithm is experimented to verify the effect of the improvement.The details are as follows:Aiming at the characteristics of complex indoor scenes,real-time performance and localization accuracy of visual SLAM are improved in static environment.In feature extraction,the ORB feature extraction algorithm is improved to eliminate a certain number of edge feature points,and the calculation of r BRIEF descriptors is improved to reduce the computation time,which improving the real-time performance;In feature matching,an improved Random Sample Consensus(RANSAC)algorithm is used to eliminate false matches and reduce the number of iterations,which improving localization accuracy and slightly improving real-time performance;The minimization reprojection error method is used in the positional estimation to reduce the depth value error at the feature points,which improving the localization accuracy.In the experiments,the TUM data set is used as the data input,and the root mean square error of Absolute Trajectory Error(ATE)of the improved algorithm is reduced by 6.54% on average and the time consumed is reduced by 7.63% on average in the nine data sets tested,which effectively improves the localization accuracy and real-time performance.A dynamic visual SLAM algorithm based on semantic segmentation and motion consistency detection is proposed for the problem that moving objects seriously affect the localization accuracy of visual SLAM,and dynamic feature points are filtered three times.Firstly,the Seg Net network is used to segment RGB images,and the segmented images with semantic labels are used to generate semantic information to reject the high dynamic object feature points for the first filtering of dynamic feature points;Secondly,the fundamental matrix is solved with the improved RANSAC algorithm,and the abnormal feature points are output by using the epipolar geometry method,and the feature points of low dynamic objects are eliminated by combining with the semantic information again for the second filtering of dynamic feature points;Thirdly,all the abnormal feature points are considered as dynamic feature points and eliminated,and the third filtering of dynamic feature points is performed;Finally,static feature points are used for tracking and localization.In the experimental part,dynamic sequences from the TUM dataset are used as the image input for testing.The improved algorithm reduces the root mean square error of ATE by 93.99% on average in the high dynamic sequences and 14.6% on average in the low dynamic sequences.The experimental results show that the algorithm can solve the problem of poor localization accuracy of the visual SLAM system in the presence of moving objects.