Research On Simultaneous Localization And Mapping Based On Depth Camera In Dynamic Environments

Vision-based Simultaneous Localization and Mapping(SLAM)is an important method for robots to perceive the environment.At present,most visual SLAM systems are designed based on the assumption of a static environment,and their performance is poor in dynamic environments,which hinders the wider application of robot in actual factories and life.In view of this,an RGB-D SLAM system based on ORB-SLAM2 for dynamic environments is designed.The main research work is as follows:First of all,an accelerated processing is performed on an instance segmentation model Mask R-CNN.The precision of Mask R-CNN is high,but the processing speed to image is very slow.When it combined with the visual SLAM,it will damage the real-time performance of the system.In our methods,by the adjusted preprocessing stage of Mask R-CNN and the improved backbone network,the processing speed of visual SLAM on input images has been improved greatly.Secondly,a novel motion segmentation method is proposed.Segmentation of moving objects and removal of motion information are the keys to ensure that visual SLAM is not interfered by moving objects.At present,the accuracy of motion segmentation methods used in visual SLAM is low,which affects the performance of visual SLAM in dynamic environments.The motion segmentation method in this thesis can be divided into the segmentation of active moving objects and passive moving objects.For active moving objects(pedestrians)which are always in motion,a motion segmentation method based on the instance segmentation model Mask R-CNN and the mask inpainting method which can be used to inpaint the under-segmentation results of Mask R-CNN is used,which can ensure the complete segmentation of the active moving objects.For the passive moving objects,the motion detection based on the pyramid LK(Lucas-Kanade)optical flow and the motion segmentation based on the region growing method are used,they are able to improve the segmentation accuracy of passive moving objects.The accurate segmentation results produced by the motion segmentation method in this thesis can improve the localization accuracy and the quality of 3D map of visual SLAM in dynamic environments.Subsequently,the feature points processing method in ORB-SLAM2 is improved.In ORB-SLAM2,there is no method for the feature points of moving objects,and also,many low-response value feature points in low-texture areas will be retained.So these problems lead to chaos in data association and affect the accuracy of pose estimation.Based on the results of motion segmentation,the feature points of the foreground(moving objects)are removed.Besides,the homogenization algorithm based on the quad tree and the response value for background feature points is proposed for dealing with the background feature points with low-response value.The new methods to feature points effectively solves the problem of chaotic data association in visual SLAM under dynamic environments.Finally,combining with the above work,a visual SLAM for dynamic environments is constructed.The improved work is tested in dynamic environments for pose estimation and mapping.The experimental results prove that the localization of the improved work is accurate and it can produce quality maps in dynamic environments.