| Automated Guided Vehicles(AGV),as a branch of mobile robots,are widely used in medical,service,industrial,intelligent three-dimensional warehouses,and the rapid development of intelligent logistics in recent years.The introduction of AGV has reduced labor demand,saved labor costs,and improved production efficiency.It has huge demand potential and development prospects in the market.How mobile robots perceive environmental information through sensors and perform autonomous positioning is the key to autonomous mobile robots to complete navigation,guidance,and control tasks.Simultaneous localization and mapping(SLAM)technology is a promising mobile robot.Robot intelligent perception technology has gradually become a hot spot in the field of mobile robot research today.In recent years,with the advantages of low hardware cost and rich environmental information,visual sensors have attracted more and more researchers to participate in the research of visual navigation technology.Visual SLAM has stood out among many navigation methods and has become a current research hotspot.In this paper,under the conditions of illumination changes,scale changes,image rotation,noise,etc.,artificial feature extraction algorithms have low feature matching accuracy,reduced matching accuracy,and key points overlap.The research based on the(Geometric Correspondence Network,GCN)deep learning network,a lightweight GCN-L deep learning network suitable for low-power embedded systems is used for feature point extraction to generate key points and descriptors in the same format as ORB features.Improved in the classic visual SLAM framework,built a GCN-SLAM system and visual navigation AGV experimental platform,and verified the proposed algorithm.The main research work of this paper is as follows:(1)The imaging principle of binocular camera is studied.Use the Zhang Youzheng calibration method in the Camera Calibrator toolbox under the ROS system to calibrate the binocular camera,obtain the internal parameters of the camera and correct the image obtained by the camera.The motion model and observation model of the visual navigation AGV are established.(2)The algorithm of front-end visual odometer is studied.The GCN-L feature extraction algorithm,a lightweight deep learning network structure suitable for embedded low-power development board,is used to compare with the traditional artificial feature point extraction algorithms ORB,SIFT and SURF.The experimental results show that the spatial distribution and speed of the GCN-L feature extraction algorithm are more suitable for visual SLAM system.At the same time,Flann feature matching algorithm and Random Sample Consensus(RANSAC)elimination algorithm were used to match feature points.The experimental results show that the proposed feature matching method significantly eliminates a large number of mismatched points and has a better matching effect.Finally,the iterative Closest Point(Iterrative Closest Point(ICP)algorithm is used to solve the motion estimation of images collected by binocular camera,and the G2 O optimization library is used to optimize the estimated motion trajectory.(3)The nonlinear optimization method based on graph optimization in back-end optimization is studied.Combine with pose optimization BA(Bundle Adjustment)for optimization.The loop detection algorithm is analyzed,and a DBOW bag-of-words model is used to judge the similarity of two images,and the loop detection judgment is performed,which effectively improves the accuracy of the loop detection.(4)Combine the GCN-L feature extraction algorithm with the traditional ORB-SALM to build a visual SLAM system.The Euroc data set was used for simulation experiments,and GCN-SLAM was used for actual scene map construction experiments through the selfbuilt visual navigation AGV experimental platform,which verified the uniformity of feature extraction and fast rotating image tracking of the algorithm proposed in this paper,and explained the algorithm feasibility. |
