Research Of Multi-robot Cooperative Slam Based On Vision

Mobile robot is a branch of Robotics, which can be applied in the fields such as exploration in unknown environment, patrol, service, etc. But the mobile robot technology is not mature enough yet at present, several key technologies need to be improved. As the prerequisite to autonomous, intelligent robots, simultaneous localization and mapping(SLAM) this technology in particular needs to be further improved. The problems studied in this paper derive from the National Natural Science Foundation of China “Research of binocular mobile robot cooperative SLAM based on local invariant mapping”, we delve into multi-robot cooperative visual SLAM and related technologies. The contributions are discussed in detail as follows:Firstly, multi-robot system and task allocation algorithm are studied. We proposed a new multi-robot system UMRS(UPn P-based Multi-robot System) based on UPn P(Universal Plug and Play). UMRS enables members discover each other. It avoids the problems such as single point failure and high degree of coupling between cooperation protocol and lower layer communication protocol. Based on this system, multi-robot task allocation algorithm is studied. A task allocation method named CMRTA(CHNN-based Multi-robot Task Allocation) is proposed, which is suitable for MT-SR-TA problem.Secondly, the method of natural landmark extraction and description is studied. To solve the problem that high data association complexity and low accuracy in the process of SLAM because of the huge number of feature points, a method of natural landmark extraction and fast matching based on 3D information of feature points is proposed. The method matches feature points which are extracted from environmental images based on binocular vision. Then the three-dimensional information of space points corresponding to these feature points is reconstructed. Point clusters, which are viewed as landmarks, are obtained by cluster analysis based on the distance between space points. To facilitate the fast matching between landmarks, a landmark descriptor is proposed in this thesis. We discussed the method of generating and matching of landmark descriptors. In order to obtain qualified landmarks, the Mean Shift clustering algorithm is improved, which can generate the appropriate number of different clustering according to the distribution of space point by adjusting the parameters such as minimum points, the initial value of the clustering radius, radius growth rate, the maximum cluster radius etc.Thirdly, the visual SLAM is studied. SLAM algorithm based on EKF(Extended Kalman Filter) is not suitable for applying in large-scale environment because of its high computational complexity, to solve this problem, a novel NL-SLAM(Natural landmark and Local map based SLAM) algorithm based on natural landmark and local map updating is proposed. The error of estimation of robot pose and map is reduced by using natural landmark. Meanwhile, the computational complexity is reduced effectively because of reduction of the number of landmark and the using of local map.Finally, on the basis of the above work, multi-robot cooperative SLAM(CSLAM) based on vision is studied. A CSLAM algorithm based on sharing landmark information is proposed, coined by MR-v SLAM(Multi-robot visual SLAM), which improved the Fast SLAM to make it suitable for multi robot cooperative SLAM. In MR-v SLAM, every member of multi-robot system executes SLAM, viewing other members as its own extended sensors, whose observation information will be merged into its map continuously. MR-v SLAM can increase the speed of mapping process in big scale unknown environment.