Deep Learning Based Topological Structure Recognition Method For Deformable Linear Object

Unlike rigid bodies,deformable linear objects(DLO)have the characteristics of soft texture,easy deformation and entanglement.It is difficult for robots to perform autonomous operations on such objects.DLO such as power wires,cords,and industrial hoses,however,are often faced in real industrial production.Works on such materials can only rely on labor at present.In the context of rapid development of automation and increased labor costs,people are eager to replace humans with robots on such works.Facing DLO,robots may confront problems of entanglement of such objects inevitably.Although DLO's geometry is very simple,they can present complicated and variable winding states.Different states may have the same topology,thus it's critical to identificate DLO's topology for the further manipulation of robots.Based on this,this dissertation propose a method to identify DLO's topology,combining the application of deep learning in image processing.In this dissertation,the topological structure of DLO under winding state is characterized based on feature points traversal.The selected feature points are intersection and starting endpoint.The feature points are obtained by deep learning target detection algorithm.When detecting the target,the feature points are marked with a detection box,which gives also the information of position,range and category of the targets.The traversal of the feature points is based on the direction in which the linear flexible body extends.This dissertation uses a gradient color from one end of the linear body to the other to characterize this direction.Specifically,the semantic segmentation algorithm is used to complete the segmentation.The segmentation result can distinguish the linear flexible body from the background and apply this gradient color to the linear part of the pixels at the same time.The intersect of the line,the segmentation can also reflect the real coverage relationship well.When the feature points are traversed,the target detection box of the feature is projected onto the segmentation map,and the local area of intersection is divided according to the different pixel values.The local area distinguish the upper and lower segments of the intersection,and the traversal of feature points is completed by comparing the average pixel values in these segments.At the end of the dissertation,the algorithm is integrated into a linear flexible body topology state detection system and the effectiveness of the method is demonstrated experimentally.The system can identify the linear flexible body topology state in real time.