Vision Based Simultaneous Localization And Mapping In Dynamic Scenes

SLAM(Simultaneous Localization and mapping)is a basic problem in the field of computer vision and robotics.Compared with the SLAM scheme based on IMU(Inertial measurement unit),Li DAR(Light Detection And Ranging)and ultrasonic sensors,vision based scheme can get richer and more intuitive information,and the hardware cost and the requirement of scene of visual sensor are relatively low.Visual SLAM has become a research hotspots in recent years because of the advantages of visual sensors.The current visual SLAM algorithm is generally based on the static world assumption;When the scene contains dynamic objects,it is easy to cause false matching,which has a great impact on the localization and mapping accuracy of the whole system;So it is necessary to deal with the moving objects in the scene.To solve this problem,this paper analyses the standard approach for dealing with dynamic objects in current visual SLAM algorithm;In this work we adopt a method based on machine learning and visual SLAM algorithm to solve this problem,and the corresponding algorithms are studied and tested.This paper use machine learning algorithms to obtain dynamic and semantic information in the scene,we propose an approach for dealing with dynamic objects in visual SLAM algorithm by fusing machine learning algorithm and traditional visual SLAM algorithm,which realizes visual SLAM algorithm in dynamic scene.The whole system framework of the project is based on ROS operating system;This paper analyzes the vision based object detection and tracking,scene semantic segmentation and scene flow algorithms to obtain the motion and semantic information in the scene,and this paper makes an in-depth study of each algorithm,compares and analyzes the advantages and disadvantages of various algorithms as well as the fusion strategy of these algorithms and visual SLAM algorithm.At the same time,the advantages and disadvantages of current visual SLAM algorithms are analyzed,and we use ORB-SLAM2 to get the geometry information of the scene;The motion and semantic information of the scene is obtained by machine learning algorithm,which is assigned to each feature point in the image;In the process of image feature matching,the dynamic points are eliminated to ensure the correct data association;Through the experiments,the effectiveness of the algorithm is proved,the requirements of the project are finished and the rationality and validity of the whole methods are verified.