Research On Visual Slam Map Fusion Technologies For Multi-Uavs System Based On Mutually Loop Closing

The UAV(unmanned aerial vehicle)positions and maps depending on vision-based SLAM(simultaneous localization and mapping)in unknown conditions,which has gained increasing focus these years.When GPS fails to meet the requirements,SLAM is one of the key algorithms to realize autonomous flight and navigation.The vision-based positioning drift will be gradually accumulated with the motion of UAV platform.Besides,the UAV cannot be equipped with multiple sensors to help reduce accumulated deviates as the robot does due to its limited load capability and fast speed.Though the loop closing detection can manage to reduce deviates,the increase in quantity of such closings amounts to extra turns for a UAV,which not only puts huge pressure on the flight control system,but also decreases the coverage per unit time.Therefore,it is of great significance in current research to figure out how to diminish errors caused by positioning and mapping and at the same time hoist the speed for three-dimensional reconstruction of environment.Based on above,this paper aims to propose a framework to help multi-UAV achieve SLAM based on MLC(Mutually Loop Closing)detection,that is to improve the amalgamation method of the point cloud map and conduct experiments through simulation and UAV platform to prove this method can be used to solve deviations and improve the accuracy as well as efficiency for map construction.On account of the SLAM framework for multi-UAV cooperation based on MLC,the UAV transmits key frame information to central server and hereof complete further MLC detection,amalgamation of point cloud map,optimization of global map,and feedback on position and condition to achieve such cooperation.The key frame information consists of feature descriptor of each frame of picture constructed by bag of word,and corresponding point cloud information,which avoids bandwidth as well as delay caused by picture transmission.At the same time,we supplement popular SIFT feature in the BOW algorithm with ORB feature in MLC to optimize the bag of word.The improvement of global map through graph optimization on the basis of the least square method will raise the speed.We conduct systematic analysis and research on the amalgamation algorithm of the map,and thus propose the improvement version based on the traditional ICP(Iterative Closest Point)and utilize the revised ICP algorithm of k-d tree to improve the map's amalgamation performance.This algorithm uses k-d tree to search within the closest area to increase the convergence rate of the algorithm.Moreover,with the assistance of euclidean distance threshold and point cloud direction threshold to eliminate the wrong pairing of point cloud,the integration algorithm performance of point cloud will be enhanced.The comparison between ICP and improved ICP algorithm in the simulation will prove the advantages of the latter one.The research of the project will help a UAV team to achieve separate flight at the junction,the later convergence(witch is to mean that the picture shows the same place),the sharing of each map and the final integration.Another merit the UAV equips is that with the help of other UAVs,the redundant data of an overlapped area in detection can be used to locate when certain UAV loses the position.This will improve the stability and robustness of UAV SLAM system.