Design And Implementation Of A Moving-Object Removal System For Visual SLAM Of Mobile Robots

Since mobile robots have potentials to achieve strong adaptability and autonomy,they begin to be used in various areas such as medical service,rescue,home service.Establishing accurate three-dimensional map and accurate positioning is a key to autonomous navigation of mobile robots,which is also a prerequisite for subsequent tasks.At present,most of the visual simultaneous localization and mapping(SLAM)systems are mainly used in static scenes,which limits the applications of SLAM systems.For SLAM in dynamic scenes,if there are moving targets within the field of view of a mobile robot,three-dimensional(3D)models of moving targets may be built in the map of the SLAM system for many times.This will also affect the performance of the loop-closure detection and the precision of the visual odometry in SLAM,leading to inaccuracy in location and mapping.For visual SLAM of mobile robots in a dynamic scene,it is of significance to detect and remove moving targets for the construction of 3D point cloud and 3D grid map of the static scene,as well as improving the performance of the loop-closure detection and visual odometry.In this paper,on the basis of the existing detection method of moving target,a moving target detection method based on RGB images and depth images is proposed.The relationship between the two image frames is solved by using image feature extraction and matching to solve the homography matrix between the two images.Then,image differences are computed in the two images at corresponding image coordinates.The difference image is divided into two frames,which are based on the depth information of the moving target.Finally,a tag image which marks the candidate regions of moving targets is obtained.Furthermore,we propose a moving object removal method based on the previous tag image and point cloud cluster segmentation.The 3D point cloud is calculated according to the camera model.The ground plane is removed by filtering and down-sampling the 3D point cloud.The point cloud is further divided into multiple clusters through clustering algorithm.By considering the tag image,the corresponding clusters of moving targets are found.Then,the corresponding clusters are projected to a plane,so that the moving target can be removed from the original point cloud by using the boundary extraction algorithm and the prism extraction algorithm.Finally,a RGB-D SLAM system based on the proposed moving object removal algorithm is designed based on the existing Real-Time Appearance-Based Mapping(RTABMap)RGB-D SLAM system.After removing moving targets,we do evaluation on the performances of 3D point cloud map construction,3D grid map construction,loop-closure detection and visual odometry,as well as real-time performance of the system.Experiments show that 3D maps of static scenes can be constructed by removing the moving target when constructing the 3D point cloud map and 3D grid map.Removing the moving target can improve the performance of the loop-closure detection and visual odometry to some extent.Moreover,the proposed moving target detection and removal algorithm has good real-time performance.