Visual SLAM For Large-Scale Environment

Simultaneous Localization and Mapping(SLAM)has been a fundamental technol-ogy for autonomous navigation of robots.Visual SLAM which uses cameras as sensor-data input has become an attractive focus in robotics.With visual SLAM method reaching maturity in the past decades for indoor small-scale environments,the issue of exploring large-scale environments attracts increasing attentions.However,the memory consump-tion of visual SLAM systems grow rapidly and the efficiency of single-robot operation reduces rapidly with their operation ranges increase.In this paper,we focus on solving the unlimited memory consumption problem and the low efficiency of single-robot operation problem under the constraint of the limited hardware resources for large-scale environments.Our work mainly includes three parts:(1)We redesign the framework of visual SLAM to solve the low efficiency of single-robot SLAM problem for large-scale environments.We analyze the mathematics of visual SLAM system.Then,we present a centralized multi-robot visual SLAM system based on ORB-SLAM2 which consists of multiple ORB-SLAM2 clients and a server The clients can build the environment map independently and the server fuses all the local maps from multiple clients into a global map.(2)We propose a multi-layer visual-SLAM data management method to solve the unlimited memory consumption problem for large-scale environments.In this paper,we divide the data management module into three layers:short-term memory,working mem-ory,and long-term memory.In the working memory,we propose the SIBDO(Spatial Information Based Dada Organizing)method to schedule the map data between working memory and long-term memory.In the long-term memory,we employ PostGIS database and ODB relation-object mapping tool to store the map data.(3)We propose a map fusing method based on image-feature detection.The method includes two main parts:map overlap detection and local-map fusion.In the map overlap detection,DBoW2 method is used to detect overlap and the PnP method is used to cal-culate the transform matrices.In the local-map fusion,we merge the local map into the global map.Then,we optimize the global map by bundle adjustment.To sum up,this paper propose a multiple-robot visual SLAM framework for large-scale environments.The multi-layer data management method we propose efficiently reduces the memory consumption of visual SLAM.And the map fusing method im-proves the map building efficiency.Experimental results with public datasets demonstrate that our system can perform multi-robot SLAM collaboratively with data management method.