Transparent Object Detection And Grasping Based Machine Vision

Transparent objects are widely used in life because of their beauty,cheapness and practicality.The research on recognition and grasping of transparent objects can extend the application scope of mobile robots,home service robots or industrial robots.Transparent itself has the characteristics of no texture and high light transmission,and its appearance is determined by the background pattern.And common commercial RGBD cameras cannot capture its accurate depth information.Therefore machine vision based detection and grasping is a challenging task.In this paper,machine vision algorithms are utilized to do the following work for the detection and grasping of transparent objects in home life:For the scenes with little change in illumination and relatively fixed target object poses,the Sum of Absolute Difference(SAD)and Normalized Cross Correlation(NCC)algorithms are used,and a homemade beverage bottle dataset is used for template matching algorithm comparison in the beverage bottle dataset.The accuracy of SAD and NCC algorithms on the beverage bottle data set is 15.60% and 27.30%,respectively.Then for the scenes with relatively large lighting changes and complex backgrounds,the deep learning algorithm Faster RCNN was used,while the convolutional attention module and the feature pyramid network module were added to the detection model in order to reduce the missing detection leakage.The experimental results of the algorithm comparison are as follows.The detection accuracy of the improved Faster RCNN model on the non-captive dataset is 86.80%,which is 4.85% better than that before the improvement;the precision detection accuracy on the beverage bottle dataset reaches97.70%,which is 4.30% better than that before the improvement.For the problem of missing depth of transparent objects,from the aspect of speed,we first tried to use the ip＿basic algorithm based on morphological operations to complete the original depth map.ip＿basic algorithm is faster,but it cannot complete the large area missing in the depth map of transparent objects.For this purpose,the depth-completion algorithm Cleargrasp based on deep learning is used,and the data preprocessing method in the model is optimized.Finally,experiments are conducted on the publicly available dataset CLEARGRASP,and the experimental results demonstrate that the optimized model achieves optimal performance on various evaluation metrics.Finally,GR＿Conv Net is selected as the grasping detection network for the problem of grasping transparent objects,and the grasping detection experiments prove that the model trained with the complemented depth map has higher accuracy.And the grasping experimental platform was built with Jaco2 robotic arm and Real Sense D435 i camera in the laboratory environment,and four types of transparent beverage bottles were grasped under the single-object scenario,and 30 grasping experiments were conducted for each object,and the grasping success rate was 63.33%.There are 71 figures,11 tables and 85 references in this thesis.