Research And Implementation Of Semantic SLAM Method Based On Visual Features

Simultaneous localization and mapping(SLAM)is a technique to build a map of an unknown environment and localize the sensor in the map with a strong focus on realtime operation.Semantic SLAM applies semantic information to SLAM processes and enhances the performance of SLAM systems by providing high-level semantic information.Combining the deep learning and other artificial intelligence methods with SLAM,we can simultaneously realize the calculation of sence geometry map construction and semantic information extraction,which is of great significance for autonomous robots and intelligent systems.The main research content of this topic is semantic SLAM based on visual features.By combining object detection based on deep learning and semantic segmentation network,dynamic objects in the scene are filtered out,semantic constraints are added to the tracking and localization to improve the accuracy,and semantic information is mapped to a three-dimensional map to build a three-dimensional semantic map.The specific research methods in this paper are as follows:Firstly,to handle the problem of scene with mobile or movable targets,a SLAM method which introduces deep learning based object detection into classic ORB?SLAM2 is proposed to make it more suitable for dynamic scene.The features extracted are divided into dynamic features and potential dynamic features.We calculate the motion model with potential dynamic features,and then use the model to select the static features in potential dynamic features for global map building.Compared with ORB?SLAM2 on KITTI and TUM datasets,The tracking accuracy as well as the application performance of the map are improved.Secondly,the real-time semantic segmentation network is used to carry out semantic segmentation of images,semantic categories are added to the tracking and localization module of SLAM as constraints,and combined with SURF features to effectively solve the tracking loss problem of ORB?SLAM2 system in feature absence scenarios.A comparative test was conducted on the KITTI data set,and a good performance improvement effect was obtained.Finally,based on the result of semantic segmentation,the semantic category information combined with camera pose and environment geometry information was mapped to the 3d space to realize the construction of 3d semantic map of the environment,and the 3d semantic map was constructed and stored with the color octree map structure.A complete semantic SLAM system is realized based on ROS platform.Experiments show that the semantic SLAM method proposed in this paper effectively improves the performance of tracking and mapping,and improves the shortcomings of the traditional SLAM method based on visual features in dynamic scenes and scenes without visual features.