Research On AGV Trackless Navigation Technology Based On Visual Perception

The AGV has been in existence for more than 70 years since its invention in the 1950 s.The traditional AGV reciprocates in a single fixed scene by means of orbital navigation.This method of navigation limits their range of motion and unable to adapt to flexible application scenarios.With the improvement of the intelligence and automation of the factory,the trackless navigation method represented by inertial navigation and laser navigation came into being.However,these navigation methods focus on the pose estimation of the robot itself,while ignoring the perception of the external environment and adjustments based on your surroundings.Vision sensors have unique advantages in terms of their ability to acquire environmental image information.The application of machine vision technology to AGV trackless navigation is an inevitable trend.Based on the analysis of the current status of AGV trackless navigation and visual perception methods,the visual trackless navigation and positioning technology of AGV,the visual perception technology based on deep convolutional neural network and the map construction method are discussed.The main research contents are as follows :(1)Under the ROS operating system,the preliminary work of AGV visual trackless navigation was completed,including the internal parameter calibration,image registration and data acquisition of RGB-D camera Kinectv2,which laid the foundation for the subsequent navigation algorithm verification.(2)Exploring the inter-frame image ORB feature matching,RANSAC PnP interframe pose coarse calculation and local pose BA optimization in the feature point method visual trackless navigation frame;in the laboratory environment,by generating a dense point cloud map alignment Preliminary observation of the pose estimation results,using the mobile platform equipped equipment to complete the assessment of the overall positioning accuracy.(3)Integrate intelligent visual perception technology into the trackless navigation architecture.The deep convolutional neural network YOLOv3 is used to detect the object of interest during the operation of the AGV.The perceptual result is used to remove the dynamic feature points on the dynamic class object according to the image pixel motion trend to improve the robustness of the navigation in the dynamic environment.(4)Save the environment map based on the navigation results.The detection result of YOLOv3 uses the GrabCut algorithm combined with depth information to achieve pixellevel segmentation to improve the positional accuracy of the object of interest in the map;convert the three-dimensional dense point cloud map into an octree and generate twodimensional according to the dynamic classification level of the object The perceptual probability map is corrected and displayed in the designed QT 5.8-based map editor.The experimental results show that the positioning accuracy of the designed visual trackless navigation scheme basically meets the practical requirements,and the generated 2d perception probability map can accurately reflect the position and category information of objects in the environment.