Visual-inertial Simultaneous Localization And Mapping In Dynamic Environments

With the lower hardware costs and the rapid development of robotics,some service robots are gradually appearing in our daily life.The development of service robots could not be separated from a key technology: simultaneous localization and mapping(SLAM).Among these approaches,the vision-based SLAM has been paid more and more attention because of its low cost and easy access.At present,all visual SLAM approaches are based on an important assumption: the environment is static.However,the actual environments will contain some moving objects such as pedestrians,vehicles and so on.Unfortunately,these moving objects will reduce the accuracy of localization and mapping.To cope with this problem,a visual SLAM combining with moving object detection in dynamic environments is proposed.An algorithm of moving object detection is presented using frame difference detection method with ego-motion compensation.For a moving camera,the method of direct frame difference using continuous images will introduce much background in a difference image.To address this drawback,we transform previous image into current image coordinate system with homography matrix,which results in a highlighted motion area in the difference image.However,it contains a small amount of background.Then,a particle filtering is used to detect moving object for improving its detection accuracy.Each particle in this filter is considered as part of the motion area,and a weighed clustering method is used to obtain the final detection result.The result shows that the proposed algorithm can detect the moving objects effectively.A SLAM system that can adapt to the dynamic environments is built using visual and IMU's data.The mainstream visual SLAM framework is used,which divides the entire process into three parts: tracking,mapping,and closed-loop detection.In the tracking thread,a moving object detection algorithm is added to reduce the localization error caused by moving objects.For feature matching,we use feature filtering of epi-polar geometry and reprojection error to further reduce the interference of dynamic features.In addition,we also use the feature points of moving object regions to calculate their position in the global map.Using closedloop detection to further reduce the cumulative error and build a globally consistent map.The experimental results show that the localization accuracy of our system is consistent with the best approach in the static environment,and accurately localization and mapping can also be achieved in the dynamic environment.