Research On Semantic SLAM Algorithm Based On Object Plane Constraint And Image Segmentation

Simultaneous Localization and Mapping(SLAM)is a basic problem in Virtual Reality and Augmented Reality.Traditional SLAM systems mainly deal with sparse features from high gradient image areas.While being robust to light changing,camera rotation and scale variation to some extent,they are easy to accumulate drift and cannot provide high-level environment information.To solve this problem,we propose to combine semantic information with SLAM progress.We make use of semantic information to constrain the SLAM optimization as well as to build high-level semantic maps to describe the environment.Firstly,we introduce common visual SLAM system’s theories,and build our own SLAM system instance on top of ORB-SLAM2.We introduce a common operation to accelerate dictionary loading speed and lowering data redundancy: We transform text format dictionary into binary format dictionary thus saving loading time;we also train specific dictionary for different scenes to improve system’s relocation accuracy and final trajectory estimation accuracy.The experiment shows that the dictionary loading is speeding up 20 times and the SLAM accuracy is proved 9.6%.In the end,the paper briefly analyzes the current problems of visual SLAM and the research direction of this paper,which lays the foundation for the follow-up research of semantic SLAM.For indoor scenes with plenty of objects and planes,we use the SSD detector and ICP method to build a map made up of objects,so as to provide semantic information and additional constraints for the SLAM system based on point features.In addition,dynamic objects are determined according to object labels.In addition,the homogeneous coordinates and the corresponding minimum representation coordinates are used to represent the plane parameters,so that the planar constraint are introduced into the nonlinear optimization of SLAM.Experimental results show that the map composed of objects and planes greatly increases the semantic information contained in the scene,and improves the accuracy of SLAM to a certain extent.For indoor and outdoor general scenes,an improved image segmentation algorithm is used for semantic segmentation,and the semantic cost function based on the observation probability model is defined.Semantic cost function and re-projection error function are optimized together to improve the accuracy of SLAM system.The camera poses and semantic segmentation results calculated by SLAM system are also used to construct a 3D occupancy map with semantic labels.Experiments show that the semantic segmentation can not only be used for semantic mapping,but also provide middle distance constraints for the SLAM system to improve the accuracy of the SLAM system.