Research On Simultaneous Localization And Mapping Based On Instance-level Semantic Information In Dynamic Environment

The traditional vision-based Simultaneous Localization and Mapping(SLAM)technology cannot obtain the semantic information of the surrounding environment,and cannot meet the needs of robots for the perception,understanding and modeling of its own surrounding scenes.Most existing semantic SLAM methods either assume that the surrounding environment is static,or only obtain pixel-level semantic information,and cannot identify each object in the environment at the instance-level.This will cause robots to face many challenges,such as the inability to complete human-computer interaction,intelligent grasping,collision detection and other high-level semantic decision-making tasks;the interference of dynamic objects in the real environment will not only reduce the accuracy of the camera's pose estimation,but also lead to the map cannot be reused.In order to solve the above problems,we design a SLAM system based on instancelevel semantic information to improve the robot's camera positioning accuracy and system robustness in real and complex dynamic scenes.At the same time,a 3D semantic map of instance-level after removing dynamic objects was established,which makes a foundation for the robot to complete advanced decision-making tasks such as intelligent grab,human-computer interaction,path planning and so on.The specific research work of this dissertation includes:(1)Firstly,this dissertation clarifies the research background and significance of the subject,and describes the development of SLAM,as well as the research status of visual SLAM and semantic map.Secondly,it introduces the theoretical basis of visual SLAM based on semantic information,including the classic framework and camera model of visual SLAM,and the design principle of visual odometry based on feature point method and the direct method.Finally,the relevant theoretical basis of semantic SLAM is explained,including common target detection algorithms used for object semantic information extraction,and related algorithms of semantic SLAM are also introduced.(2)This dissertation proposes a design method of visual odometry that eliminates moving objects in dynamic environment,which aims to solve the problem that the common visual SLAM cannot obtain the environmental semantic information and low positioning accuracy in a complex dynamic environment.Aiming at the problem of low segmentation accuracy in common semantic segmentation tasks,a optimized semantic segmentation algorithm based on the context correlation information is proposed to improve the accuracy of semantic segmentation.At the same time,in order to solve the problem of poor camera pose estimation accuracy of traditional visual SLAM in dynamic environment,a dynamic object feature point detection and elimination algorithm based on instance-level semantic information is proposed,which greatly improves the camera estimation accuracy and system robustness of visual slam in dynamic environment.Related experiments are carried out in the data set and real environment,and the results prove that the algorithm in this dissertation has certain advantages compared with other algorithms in dynamic scenarios.(3)This dissertation proposes an object-oriented instance level algorithm for 3D semantic map construction,which aims to build a global,static and labeled 3D semantic map.Aiming at the existence of black holes in depth images due to object occlusion and other reasons,an adaptive joint mean and bilateral filtering algorithm based on semantic information to build a more complete semantic map of the target.At the same time,a construction and update algorithm of instance level 3D object semantic label library is proposed,and on this basis,a 3D semantic map oriented to the object instance level is constructed.Related experiments are carried out in dynamic data sets and real dynamic environment,and the results verify the effect and value of the algorithm in this dissertation.