Research On The Algorithm Of Semantic SLAM

In recent years,computer technology,sensor technology,deep learning technology,etc.have developed rapidly,and the application of robots has become more and more extensive,but the robot's autonomous positioning and environment perception is still a key issue.At present,most of the visual SLAM algorithms rely on traditional manual feature points and use them to represent the surrounding environment.The understanding of the environment is lacking.For example,information such as what kind of objects are contained in the image is ignored.Part of the method of combining semantic information needs to model the object in advance,or just construct a semantic map,and does not use it to help the SLAM process.Based on the traditional visual SLAM technique,this paper extracts the semantic information while extracting the traditional geometric feature points,optimizes the current robot pose and constructs the semantic map,and realizes the mutual promotion of SLAM technology and semantic information acquisition.At the same time,the camera is equipped with a robot to realize positioning and mapping in an indoor environment.The main work done in this paper is as follows:(1)The classical visual SLAM algorithm framework and the probabilistic representation of the SLAM problem and the least square representation are studied.The geometric model of the visual sensor is introduced,and the concept of Lie group and Lie algebra is introduced.(2)The methods of acquiring geometric features and semantic information are studied.Firstly,the method of geometric feature point extraction and matching is studied,and the relative motion estimation of the camera between adjacent frame images is realized.The three-dimensional map points corresponding to the two-dimensional feature points are restored,and the visual odometer based on geometric features is realized.At the same time,the method of extracting semantic information is studied.According to the speed,accuracy and information richness of semantic information extraction,Mask RCNN is selected for semantic feature extraction,and it is associated with geometric information to give geometric feature point semantic features.(3)The optimization problem of robot pose estimation is studied.Firstly,the EKF and graph optimization methods for robot pose estimation are analyzed.Then,the local optimization of the local features of the BA optimization,geometric and semantic information optimization,and the global pose optimization based on loop detection are mathematically expressed.It is then transformed into a problem that can be solved by graph optimization.(4)Set up the experimental platform to verify the algorithm of this paper.First,the improved SLAM algorithm is tested on the TUM public data set,and a good positioning accuracy is obtained.Then build a hardware platform for testing in real scenes,experiment with robots in indoor venues,and construct a semantic map of indoor scenes.