Research On Pose Estimation For Mobile Robots In Indoor Scenes

3D reconstruction is an important research hotspot in computer vision to obtain 3D models of objects and environments and it is applied in various fields such as robotics,autonomous driving and unmanned aerial vehicle(UAV).In the 3D reconstruction,high-precision pose estimation is the key to accurately reconstruct the 3D models of objects and environment.Due to the rapid development of mobile devices and information technology,real-time localization of indoor scenes has gradually received widespread attention.The complexity and changeability of indoor scenes also bring challenges to the pose estimation.To address the shortcomings of existing pose estimation methods,which based on multi-view and 3D point cloud and widely used in mobile robots,this paper conducts research in term of improving the accuracy,robustness and calculation speed of the pose estimation system.The main works are:1.An optimized ORB-SLAM2 system,which can be used in mobile robots,based on blur image elimination and weighted pose optimization is proposed.In order to improve the robustness of ORB-SLAM2 when dealing with blur images,a blur image detection and elimination based on Harr wavelet is integrated to the visual odometry of the original system.Additionally,in order to enhance the accuracy of pose estimation,the correlation between matching feature is used as the weighting coefficient to construct a correlation weighted pose optimization.The test results in the open source datasets and actual indoor environment show the effectiveness of the optimization method.2.A dynamic ORB-SLAM2 based on Yolov6 has been constructed in this paper.The Yolov6 helps system to detect and remove dynamic features,and only static features can be used in subsequent operations.That improve the robustness and accuracy of system in the dynamic environment.Furthermore,a dynamic parameter feature extractor is constructed in this paper,which helps the system obtain enough features to avoid system tracking fail.The test results in open source benchmark and actual indoor environment show the effectiveness and robustness of these optimization methods.3.Two optimized REGTR method are constructed: Fast-REGTR with faster pose estimation speed,and GMS-REGTR with higher pose estimation accuracy.Fast-REGTR filters out 3D features with low overlap coefficients through a threshold.That reduces the amount of data and improves the speed of pose estimation.A multi-scale network module,which is similar to Res2 Net,is integrated to the backbone network.That improve the feature learning ability and receptive filed of the backbone network,which is conducive to improving the accuracy of the final pose estimation results.The optimized method is named as GMS-REGTR.The test results in 3DMatch and Model Net40 dataset demonstrate the effectiveness of both Fast-REGTR in faster pose estimation and GMS-REGTR in improving the accuracy of registration.