| With a new round of artificial intelligence technology innovation and accelerated development of computer vision technology,mobile robot technology has been widely concerned.Simultaneous localization and Mapping(SLAM)is the key technology of robot intelligence,which has a wide range of application scenarios in people’s living environment.For example,navigation robot,sweeping robot,military robot,unmanned robot and so on.For some heavy work and some complex environment,especially in the threat to human life and safety,mobile robot is more able to replace people to complete special work.Compared with the SLAM algorithm based on lidar,the vision-based SLAM system has the characteristics of low price and rich data acquisition.However,visual SLAM technology still has some problems such as low accuracy,poor map quality and poor robustness in practical complex environment.Aiming at the problems of low positioning accuracy and poor robustness of indoor intelligent mobile robots in indoor dynamic environment,this paper studied and constructed an improved ORB SLAM2 system,and established a ROS-BASED mobile robot platform,which was evaluated in public data sets and actual indoor environment.The main research contents are as follows:1.This paper systematically analyzes the basic knowledge theory involved in visual SLAM algorithm,and analyzes the problem that in the real dynamic environment,when the front-end system of visual SLAM performs data association with image matching,the interference of moving objects will lead to the error matching of feature points,thus affecting the accuracy of camera pose estimation.The performance of ORB-SLAM2 based on feature point method was tested,and the shortcomings of ORB-SLAM2 algorithm in positioning and mapping were analyzed.It did not deal with moving objects in dynamic environment,and the establishment of dense landmark point map could not be further used,the theoretical basis for the subsequent improvement research is elaborated.2.In view of the problems mentioned above,this paper improved the ORBSLAM2 system based on vision sensor and designed a moving target processing algorithm to eliminate the false matching points in indoor dynamic environment: the Mask R-CNN network was used to segment the prior object Mask.On this basis,aiming at the problem that deep learning could not distinguish the motion attributes of the object without prior information,the pyramid optical flow method was used to detect the motion feature points of the object,and then the location region of the moving object was extracted based above information.In the subsequent pose calculation process,the motion feature points within the range of the target object are removed to improve the positioning accuracy.3.Build mobile robot experimental platform based on ROS operating system,and add mapping module based on topic research to generate dense point cloud map that can be used for robots to perform advanced tasks.In this paper,localization and mapping experiments are carried out in TUM data set and real environment respectively.The results show that compared with ORB SLAM2,the improved SLAM system can significantly reduce the localization error in dynamic environment.Compared to the Dyna SLAM and DS-SLAM algorithms for dynamic environments,the improved SLAM system can increase the average preciseness of localization by 20 percent in the same scene,as well as a dense point cloud map of the environment for further robot work.In this paper,an improved ORB-SLAM2 system was designed and implemented in indoor dynamic environment,which combined deep learning and optical flow assisted SLAM to eliminate external points.Compared with other visual SLAM systems,the positioning accuracy and robustness of the system in this paper were higher in indoor dynamic environment. |
PDF Full Text Request