Autonomous Mapping Via Multi-Machine Cooperation

With the improvement of Microelectronics and Micro-Electro-Mechanical systems(MEMS,Micro-Electro-Mechanical Systems),innovations in robotics structures and applications are developing rapidly.The UAVs,as one of the beneficiary platforms of technological development,are gradually becoming miniaturized and intelligent.Because of unpredictable factor,UAVs have a small field of view and are prone to danger and injury to people or objects when human operated,so it is necessary to provide autonomous navigation skills for UAVs.The existing GNSS-based UAV autonomous navigation scheme has a single function,high power consumption of individual UAVs,low payload,and the intelligence of information acquisition and processing needs to be improved.To overcome the above limitations,this thesis proposes a group of UAVs equipped with vision and inertial sensors,which are networked and communicated to quickly collect environmental information and solve the problem of collaborative autonomous map building in unknown environments with GNSS rejection.At present,collaborative autonomous mapping systems are mostly limited by specific constraint scenarios and do not have full autonomy.In order to build a real-time collaborative mapping system with fully autonomous capability,the following work is in this thesis as fellow:(1)Decentralized relative state estimation.The VINS-Fusion algorithm is used as the basic state estimator in the basic autonomous mapping system to solve the UAV relative positioning problem.Specifically,the calculation of relative position is accomplished based on the idea of matching the key frame library with the Bag-Of-Words of localmapping points between neighboring UAVs,we optimally accelerate the algorithm and then deploy it,thus completing the calculation of relative positions.(2)Multi-machine real-time mapping.This task is achieved by fusing the transformation matrix of relative coordinates and local maps.It should be emphasized that,considering the contradiction between the real-time demand of map construction and data densities,we propose two compromise strategies,on the one hand,using octree color maps to meet the real-time demand,and on the other hand,storing locally dense point cloud data in.bin format for global recovery at the back-end,aiming to improve data processing efficiency and reduce the memory demand.(3)Autonomous multi-machine path planning.By building a configuration space with depth information and representing obstacles as grid nodes,our main work is to build a collaborative autonomous mapping system without human participation,in which all stages of perception-planning-control are performed by machines autonomously.The path planning module involved is divided into discrete path point search at the front-end,which uses the D*Lite algorithm to generate geometric trajectories,and continuous trajectory optimization at the back-end,which combines constraints to generate continuous feasible trajectories.The front-end in collaborative decision making avoids collisions between neighboring machines in advance by suppression heuristics to increase map resources,while the back-end optimizes the collision constraints between newly added machines.The whole collaborative autonomous mapping system ensures that all machines work together and are able to implement the autonomous map building function safely and robustly.(4)Design and implementation of the communication framework.Obviously,communication is necessary for collaborative mapping,and in this thesis,we use a centralized communication framework,which contains both an on-board subsystem and a ground subsystem.The equipment carried by the UAV undertakes the task of data transmission,while the ground equipment mainly runs each relevant algorithm and displays the computational results.The data used in the system are packaged in the form of topics and transmitted in the ROS(Robot Operating System)operating system to meet the basic requirements of the algorithms in terms of time delay.