Collaborative Mapping Of Multi-robot Based On Visual Scene Identification

Faced with a complex and unknown large scene,multi-robot cooperative mapping provides a navigation map which is used in polling and rescue,and it is basis for robots to fulfil tasks.Faced with a complex and unknown large scene,single-robot has a low efficiency in mapping,and even can't fulfil the task.Compared with single-robot system,multi-robot system has more advantages in the aspects of perception,decision-making and execution.Multi-robot system has high efficiency in fulfilling tasks,and they can fulfil some tasks which single-robot system can't fulfil.Therefore,the multi-robot cooperative mapping system is suitable for a variety of occasions,and worthy of further research.Based on local maps,this thesis designs a system of distributed collaborative multi-robot mapping based on visual scene identification.The main work is as follows:In order to get the overlap between local maps,this thesis proposes a method to identify the same scene.The local maps can be jointed together only if there is overlap between local maps.Therefore,the method uses the algorithm of DBo W2 to analysis visual information that got from the process of mapping.Then,a screening process is designed to screen the similar area.Finally,the similar candidate frames between two robots' historical scene frames will be got.Further more,despite the similar candidate frames have been screened,it also has some frames that is not similar with the current frame.Therefore,in order to solve this problem,this thesis proposes the voting method.After voting,all similar candidate frames represent the same scene.In order to joint the local maps,a method which uses the point cloud registration to get transformation matrix is proposed.Jointing two local maps needs the transformation matrix which is between coordinate system of two local maps.This thesis uses the overlap to figure out the transformation matrix.The method regards the algorithm of SAC-IA as rough registration,and regards the algorithm of NDT as precise registration.The input of the algorithm is point cloud information of overlap.The output of the algorithm is transformation matrix,and using this transformation matrix can joint two local maps.In the actual room,the design of the distributed collaborative multi-robot mapping method is verified.The room includes sofa,bed or other obstacles.The experiment gets a global map of the room.The experimental results are in line with expectations,and show that the proposed method is reasonable and effective.