Research On Dynamic Environment Visual SLAM Based On Semantic Information

Vision based simultaneous localization and mapping(SLAM)is the core function of robot autonomous movement,which mainly solves the problem of "where am I?" in robot What's around me It is the key to realize autonomous intelligence of robot.In recent years,with the extensive research and application of object detection and semantic segmentation based on deep learning,it is possible to obtain very accurate semantic information.Integrating the semantic information of objects in the environment into SLAM system is a hot research direction of SLAM.In order to provide high-level semantic information for robots,this paper mainly carries out research work in two aspects: one is to design a lightweight target detection model of practical SLAM system,which can achieve high detection accuracy as well as high detection speed.The other is the combination of multi view geometry and semantic information to eliminate the dynamic objects in the mobile environment,so as to improve the robustness and speed of the system.Therefore,this paper studies from the following aspects:(1)A lightweight target detection algorithm is proposed,which combines YOLO-Tiny with Vary Block module.Aiming at the high real-time requirements of slam and the low detection accuracy of YOLOv3-Tiny in SLAM,this paper designs the YOLO-Tiny algorithm to replace the subsampling layer in the benchmark network,reduce the information loss caused by subsampling through Vary Block module,optimize the initial anchor of the network by Kmeans++,and fuse the target detection network with SLAM system.(2)Aiming at the problem of too long matching time caused by hierarchical clustering in feature matching,a feature point matching algorithm based on target detection is designed to replace hierarchical clustering.The trained model is applied to feature point matching to detect object categories in real time and attach semantic tags to feature points.The feature points are preliminarily divided by tags,so as to improve the feature point matching of SLAM system.(3)In order to eliminate the dynamic objects in the environment,a dynamic object elimination algorithm based on semantic constraints and geometric constraints(DOEA)is proposed.YOLO-Tiny algorithm is integrated into slam system to eliminate the dynamic objects that can be defined in advance.The positioning accuracy of SLAM system is optimized by combining geometric theory.