| As grows industrial automation and intelligence by leaps and bounds,the application of robot technology and machine vision is becoming more and more extensive.In the logistics and transportation industry,container transportation is an important way.In the traditional way,the handling of container materials mainly relies on labor,which is inefficient and expensive.Therefore,it is of great significance to develop a container material handling robot to improve the efficiency of the production line and reduce labor costs.This thesis takes the loading and unloading robot as the hardware platform and the container cargo loading and unloading as the research background,and uses machine vision,deep learning and laser radar positioning technology to study the key technologies of robot perception.The following are the main contents of this thesis:First of all,in response to the low accuracy of current object detection algorithms in identifying and locating goods in containers,an image dataset was constructed,and YOLOv5 was selected as the algorithm for object recognition and positioning,and the deficiencies of YOLOv5 were improved in three aspects,and proposed the YOLOv5-CG(YOLOv5-CBAM-Ghost Net)network model.Specifically,by embedding the attention mechanism CBAM module in the feature extraction stage of the backbone network,the feature extraction ability of the model for the target object is improved,and the ordinary convolution in YOLOv5 is replaced by Ghost convolution,which reduces the number of parameters and complexity of the model,and improve the bounding box loss function to increase the convergence speed and the training speed of the model.The experimental results show that the improved model witnesses an accuracy of 93.87%,an increase of 2.55%,a reduction of 13.13% in model parameters,and a drop of 11.18% in computational complexity.Secondly,the robot needs to operate alternately in the open space and in the container.In order to improve the flexibility of robot positioning and navigation,combined with the two positioning principles of reflectors and laser SLAM,a combined positioning scheme of them is proposed,which achieves global positioning of the robot in the work scene and and the positioning error is verified to be within 2-5CM through experiments.In addition,in order to obtain the navigation path of the robot,the position and posture of the container are detected,and the position and posture of the container in the global coordinate system are obtained.Then,point cloud clustering is used to fit the edge of the container,calculate the centerline of the container,and use the centerline of the container as the guiding path for the robot to carry out loading and unloading operations,guiding the robot to the work site for material identification and positioning.Subsequently,on the basis of target recognition and positioning,the measurement principle of the depth camera was delved into deeply.And a 3D structured light camera was selected to realize the three-dimensional positioning of the woven bag in the container,and the internal and external parameters of the depth camera were calibrated.The three-dimensional coordinates of the target were calculated by combining the color map and the depth map,which achieves the spatial positioning of the woven bags on the robot-based coordinate system.The final experimental results prove that the accuracy of the improved target detection algorithm for the spatial positioning of the woven bag meets the loading and unloading requirements of the robot.Finally,the target recognition and positioning system studied in this thesis is verified through practical application.The perception system studied in this thesis is deployed in the Ubuntu 18.04+ROS(Robot Operating System)environment,a robot experiment platform is built,and the process of perception data interaction is designed.The experimental results verify the feasibility of the loading and unloading robot identification and positioning system studied in this thesis. |
PDF Full Text Request