A Research Of Semantic Descriptor Based Simultaneous Localization And Mapping In Indoor Dynamic Scenes

Simultaneous Localization and Mapping(SLAM)means that a robot uses its own sensors in an unknown environment to locate its position and posture through the observed environmental features during movement,and at the same time incrementally build the map of the surroundings.SLAM has a wide range of applications such as autonomous driving,obstacle avoidance navigation,service robots,etc.It is the key technology of intelligent robot applications.At present,most researches on visual SLAM assume that the environment is static,which makes the SLAM algorithm unable to handle complex and changeable dynamic scenes,that limits the application of SLAM in real environments.This dissertation proposes a novel SLAM algorithm based on semantic descriptor,combined with knowledge graph,which can effectively deal with SLAM problems in dynamic scenarios.Additionally,without prior information,this dissertation constructs a cube model for objects in the environment,and uses the cube model to optimize the camera pose as well as establish an object-level semantic map.The main contributions of this dissertation are as follows:(1)A SLAM algorithm based on semantic descriptor applied in dynamic scenes is proposed.First,according to the segmentation results of the image semantic segmentation network,a new type of semantic descriptor is constructed to describe the semantic information around the key points;then,the knowledge graph is used to construct the relationship between the entities in the environment,combined with the semantic descriptor to obtain high-level semantic information in order to accurately detect dynamic objects;finally,remove the detected dynamic objects to improve the accuracy of pose estimation.(2)Construct a cube model of the object.Combined with the semantic descriptor,we jointly optimize the accuracy of the cube model and the camera pose,and an object-level semantic map is established finally.First,use the image information of a frame and the initial camera pose,combined with the semantic segmentation network,to estimate the rough cube model of all the detected objects in the frame;then by minimizing the reprojection error and the semantic distance,the cube model and the camera pose are jointly optimized,and the model information is continuously updated as the camera moves;finally,an object-level semantic map is established based on the optimized cube model.(3)Extensive experiments are carried out on the TUM dataset.Compared with ORBSLAM2,the accuracy of pose estimation in dynamic scenes is greatly improved.We also compare our method with several SLAM algorithms applied in dynamic scenes and the experiment results show that our algorithm is competitive both in terms of accuracy and real-time performance.