Research On Simultaneous Localization And Mapping Method Of AGV Based On Laser Visual Information

With the improvement of artificial intelligence,intelligent mobile AGV has been widely used.SLAM technology is the key to realize intelligent navigation of AGV.SLAM technology can be divided into laser SLAM and visual SLAM according to the different sensors.The two methods have advantages and disadvantages.An effective SLAM technology needs to be applied to AGV,in order to overcome their limitations,this paper studies SLAM method based on laser and visual information,and applies it to the research and development project of intelligent mobile AGV.The motion and observation model of AGV is established.Firstly,the SLAM problem based on Bayesian filtering is researched.According to the motion model in SLAM,the two-wheel differential kinematics equation of AGV based on odometer information is derived.According to the observation model in SLAM,the AGV observation equation based on lidar and depth camera sensor is established.On this basis,the laser vision information is collected and the camera internal parameters are calibrated.The SLAM method based on laser information is researched.According to the shortcomings of the traditional FastSLAM method based on Extended Kalman filter and particle filter,an improved RBPF-SLAM method is proposed.Aiming at the problem of "proposal distribution" selection in FastSALM,laser observation information is introduced to concentrate particles in high probability intervals;aiming at the errors of AGV odometer motion model,inter-frame matching method is used to calculate motion transformation;weight optimization method is used to improve original resampling to alleviate particle dissipation problem;aiming at the low degree of visualization of feature map,occupied grid map is used.The simulation results show the effectiveness of the improved RBPF-SLAM method.SLAM method based on visual information is researched.the feature extraction matching,pose estimation and bundle adjustment optimization in visual odometer is researched.The “Brute-Force” matching method is used to make the initial matching,the mismatched points are eliminated by using the random consistency sample consensus,the EPnP method is used to solve the pose according to the matching result.And construct a minimized reprojection error model using bundle method adjustment.Aiming at the shortcomings of visual odometer,a back-end optimization method is proposed,which introduces the global pose grahp optimization model,constructs the minimum error formula and solve it using g2 o.The word bags model is used to establish a visual dictionary,Kmeans++ is used to cluster and establish a k-d tree structure.For the disadvantage that the point cloud map is large and difficult to use,the octree map construction method is used to construct a dense map.The simulation results of the TUM data sets prove that the precision and map effect of the back-optimized visual SLAM method is better than the visual odometer.Based on the above research,the AGV principle prototype was built for the actual AGV demand.ROS was used as the standardized system platform.The SLAM system was built by using lidar and depth camera respectively.The communication mechanism between the systems was analyzed and tested in different environments.The verification results show that the effectiveness of the above SLAM method provides a basis for the selection and rapid development of AGV in SLAM.