Research On Unmanned System Localization And Mapping Method Based On RGB-D Image

Simultaneous Localization and Mapping(SLAM)technology has developed rapidly and has been widely used in many fields such as medical,smart home,and military.Among them,the combination of SLAM technology and unmanned vehicle applications has attracted extensive attention from scholars at home and abroad.Indoor scenes usually contain material items such as plastic,fiber,wood and bamboo,stone,ceramic,metal and glass.Among them,glass panels are the blind area for sensors such as monocular camera,binocular camera,RGB-D camera and LIDAR to detect because of their transparent and reflective characteristics,which leads to the decrease of the accuracy of SLAM technology in localization and map building in such scenes,and the difficulty of accurate localization and map building increases.To address this problem,this paper innovatively proposes a method to detect the position of glass panels based on vision sensors of RGB-D images,fuses the detection data with the SLAM building results to obtain a scene map containing glass panels,and applies the map to the autonomous localization and navigation of unmanned vehicle systems.The main research contents of this paper are as follows:1.The kinematic model and dynamics model of the unmanned vehicle system are analyzed,and the laws between the displacement and heading angle of the unmanned vehicle and the velocity and angular velocity with time are obtained.On this basis,the hardware and software control system of the unmanned vehicle is built,the PID-based velocity control and position control of the unmanned vehicle are realized,and the common LAN is built and the remote control of the unmanned vehicle is realized.2.The RGB camera was calibrated using the Zhengyou Zhang calibration method to obtain the RGB camera internal reference.We designed and used a skeleton checkerboard to calibrate the depth camera,and verified the method of using the constraint of parallel and equidistant checkerboard lines to optimize the position of checkerboard corner points,reduce the reprojection error of checkerboard corner points,and obtain the depth camera internal reference.3.The acquisition of RGB images of unmanned vehicles in indoor multi-glass scenes was completed and a dataset was built,and the pixel-level classification of glass panels and ground was performed using PSPNet semantic segmentation network.Based on the semantic segmentation results,we designed the glass panel and ground position detection method to fit the model of glass panel.The initial map points of the experimental scene were obtained based on ORB-SLAM2,and the initial map and glass panel detection data were fused to build a global map for unmanned vehicle navigation.The experimental results show that the calculated angle error between the real glass panel and the detected glass panel is ≤3°,and the distance error from the midpoint of the detected glass panel to the real glass panel is ≤0.1m,which verifies the effectiveness of the map construction method.4.The global path planning is completed based on the A* algorithm,the local path planning is completed based on the DWA algorithm,and the unmanned vehicle localization method with multi-sensor(wheeled odometry,IMU and visual localization)fusion is designed based on the Kalman filter method.Simulation results show that the global path planning algorithm can plan the optimal path based on the global known raster map,and the local path planning algorithm has better obstacle avoidance capability,which can avoid both stationary obstacles and obstacles in motion.The experimental results verify the effectiveness and accuracy of the localization algorithm and navigation algorithm of the unmanned vehicle.