Design Of AGV Perception System Based On Binocular Vision

With the rapid advancement of Industry 4.0,computer advanced manufacturing,controllable flexible manufacturing and automated warehousing and logistics systems have been widely used.As an important automatic tool,handling robot plays an important role in it.The key technology of AGV autonomous navigation is: planning the driving path based on the prior map,sensing the obstacles on the driving road with sensors,and dynamically adjusting its own trajectory according to the sensing information.In this process,obstacle perception is a very important part of AGV autonomous navigation.In this paper,the key technologies in the sensing task of AGV,including target detection,target tracking and depth estimation,are deeply studied to realize the tasks of detecting obstacle categories,tracking dynamic targets and estimating the position of obstacles,so as to help AGV achieve safe and efficient production.In this paper,the binocular camera is used as the sensing sensor,and the visual sensing system can judge the dynamic obstacles in the external environment by acquiring the external image information.Specifically,perceptual tasks can be divided into two categories of functions:(1)Obstacle type recognition: this task uses the convolutional neural network detector YOLOV4-Tiny to realize obstacle detection,and the recognition range includes: workshop staff and AGVs in other work.The task was trained on its own data set.To ensure data balance,part of the data was extracted from the MS COCO data set to expand the data set.In terms of model structure,an attention mechanism is introduced into Yolov4-Tiny in this design,which improves the accuracy of the detection network and enhances AGV’s ability to identify obstacles on the premise of guaranteeing the real-time detector.(2)Quantification of obstacle distance: this link introduces target tracking and binocular depth estimation,which is used to judge the motion state of an obstacle and its relative position from the vehicle.In the design of this paper,target detection results are used as the state observation quantity,Deepsort is used to track the obstacles,and the tracking results are used as the ROI input depth estimation model.This link will carry out feature point extraction and stereo matching on the ROI of the tracking output obstacle,and use the triangular parallax method in binocular vision to calculate the matching key points,and quantify the task of the actual distance between the obstacle and the three-dimensional space.The whole sensing system can quickly perceive the category and position of obstacles in the scene,provide necessary sensing information for the path planning of the AGV,and assist the AGV to adjust its own motion state in time,so as to achieve the purpose of safe and efficient production.