Research On Mobile Robot Visual SLAM Algorithm Based On Dynamic Scene

In recent years,with the development of automation technology and computer science,mobile robots have begun to be more used in production and life.One of the basic tasks of mobile robots is to be able to perform autonomous positioning and navigation in unknown scenes,while based on vision.SLAM(Simultaneous Location and Mapping)provides a suitable solution for this task.The binocular camera has the characteristics of restoring scene scale information and can be applied outdoors.The visual SLAM based on binocular camera has become an important direction for robot autonomous navigation and positioning research.This paper focuses on the research of mobile robot SLAM technology based on binocular camera.The specific research contents are as follows:Firstly,it expounds some problems encountered by mobile robots in autonomous positioning and navigation under unknown dynamic scenes,thus clarifying the research significance of the subject,and then proposes a mobile robot visual SLAM algorithm based on dynamic scenes to solve the above problems.Then,based on the problems to be studied,the current research status of mobile robot visual SLAM and the main research contents of visual SLAM are discussed,and the specific research methods are determined.The overall framework of mobile robot visual SLAM algorithm based on dynamic scene is designed.Then,for the problem to be studied,the whole system is divided into two parts based on binocular camera simultaneous location and mapping and dynamic scene segmentation,and detailed research is carried out separately.Based on the simultaneous location and mapping algorithm of the binocular camera,stereo matching is performed by combining sparse matching and dense matching,thereby avoiding global matching and improving the efficiency and accuracy of stereo matching.Then based on the results of stereo matching,Gaussian iteration and bundle adjustment method are used in the RANSAC framework to complete the estimation of robot motion and complete the construction of point cloud map.In the dynamic scene segmentation part,an accurate method for dynamic scene segmentation is proposed.This method uses the multi-task cascade network to perform instance-aware semantic segmentation of images,and uses optical flow motion and geometric mapping to calculate scene flow.The result of instance-aware semantic segmentation is merged with the scene stream to determine the motion state of the instance,and a foreground and background segmentation is performed on the scene.Use the background part to estimate the motion of the robot to improve the accuracy of the motion estimation of the algorithm in dynamic scenes.In order to verify the effectiveness of the proposed dynamic scene-based mobile robot SLAM algorithm,the algorithm is then verified in the dataset and the actual object scene.The open source visual SLAM algorithm with good performance and the visual SLAM algorithm based on dynamic scene segmentation are presented in experiments under different datasets.The experimental results show that the proposed algorithm can improve the robustness and accuracy of the whole system.Better performance in dynamic scenarios.At the same time,it has better adaptability in the actual physical scene.