| For a mobile robot,realization of real-time localization of its own position and the construction of environment cognitive maps have been two key problems in the research of Simultaneous Localization and Mapping(SLAM).The research of SLAM based on traditional mathematical probability method cannot be widely applied due to the limitation of algorithm as well as sensor precision.Visual SLAM(VSLAM)takes camera as an external visual sensor,realizes calculation of robot position information through feature extraction,description and matching of environmental image information acquired by camera,and completes the construction of environment cognitive map on this basis.Compared with traditional SLAM methods,VSLAM has become the focus of SLAM research in recent years due to its better real-time performance and closed-loop detection.On the basis of bionic SLAM based on navigation mechanism of mouse brain cells,an improved method was proposed in this thesis in order to solve the problems of light angle change and sudden obstacles in the process of mobile robot learning and cognizing the environment.In traditional probability algorithm of SLAM problem,disadvantages like large amount of calculation,high complexity,easily trapped at local optimal solution and etc.Some mouse brain cells that have functions of navigation such as view cells,grid cells and pose cells were applied to SLAM,also,the relationship between each cell and their influence to SLAM were studied respectively in this thesis.With models of view cells,grid cells and pose cells,GVP-SLAM(Grid cells+View cells+Pose cells+SLAM)was proposed base on multi-cell navigation mechanism,and cognitive map was constructed.Then,a characteristic and direction parameters based dynamic growing self-organizing feature map(DGSOM)neural network model was studied and this neural network model was applied to the proposed GVP-SLAM model.A variety of sensors were selected in order to build the experimental platform to verify the feasibility and effectiveness of GVP-SLAM.The GVP-SLAM model with multi-cell navigation mechanism learns unique scenes in the environment through view cells,forms pose cells to represent current position through head direction cells and place cells as well as correlative competed artificial neural network,also completes topological cognitive map with the cooperation of view cells and pose cells.Based on SURF feature matching algorithm,rotation,scale transformation and brightness remain unchanged,which reduces the error cognitive points during cognitive map construction and improves the matching rate of feature points in the environment.The relocation mechanism and closed-loop detection algorithm of grid cells are referred in order to avoid the impact of light angle changing on SLAM and improve the localization accuracy.The proposed DGSOM neural network model avoids confusion of maps by adding feature points and reduces learning score by adding motion direction.The accelerometer and gyroscope are integrated to more accurately measure the velocity and angular velocity of the mobile robot in the process of movement,so as to avoid the failure of view cells under the influence of sudden obstacles.The fusion of laser rangefinder can realize the rapid detection and real-time obstacle avoidance of the mobile robot.Biology concept was introduced to traditional SLAM model in this thesis,and a step by step research system was formed with the integration of mathematical modeling,software simulation and experimental validation.Multiple mouse cells and the mathematical model derived from neural network were applied to analysis the robutness and real-time performance of the system,to provide important theoretical reference for mobile robot diversified SLAM research. |
PDF Full Text Request