Research On SLAM Algorithm Based On Semantic Information

SLAM technology is one of the key technologies in robotics,autopilot,augmented reality and other fields.It is also the basic technology for intelligent mobile platform to perceive the surrounding environment.The robustness of existing systems and methods is not high.With the development of artificial intelligence technology,the trend of combining deep learning with traditional methods based on geometric model is taking shape,which will promote visual SLAM technology towards long-term and large-scale real-time semantic applications.This paper is devoted to the combination of visual SLAM and semantic information extracted by instance segmentation algorithm.The main work is as follows:Firstly,the hardware framework of visual SLAM system is designed and the key modules are selected.The software framework of the system is designed and the flow chart based on various core software is given.Secondly,a vector format is designed to store semantic information extracted by Mask RCNN.Using the obtained semantic vectors,a machine learning algorithm is proposed to measure image similarity.It includes combining the semantic vectors to generate feature vectors,giving criteria to determine whether the scene is similar,marking samples and automatically adjusting the parameters of the model.The whole process is tested with open data sets,and more than 99% accuracy and 98% recall rate are obtained.Thirdly,aiming at the problems of judging closed-loop based on similarity of feature points in SLAM system,image similarity fusing semantic information is used to identify closed-loop.The probabilistic value that the machine learning model output is used as the similarity of image on the semantic level.First,the outliers in the semantic similarity are removed by combined filtering.Then,the semantic similarity and the similarity of feature points based on the bag of words model are weighted and fused,and the similarity after fusion is used to judge the closedloop.Experiments show that this method can achieve higher accuracy and recall rate,and is more robust to dynamic environment.Fourthly,a three-dimensional mapping method combining semantic information is proposed,and the masked images containing semantic information and corresponding depth images are fused to generate a three-dimensional point cloud map containing semantic information.In order to make the generated point cloud map more convenient to store and more accurate,removing outliers,sparsity and segmentation operations are used.Finally,threedimensional mapping tests are carried out for real experimental scenes and open data sets.