Research On Garment Grasping Point Detection Based On Deep Learning

The task of grasping hanging garment has always been a very important topic in robotics research.However,the complex dynamic properties,wrinkles and self-occlusion of the garment always make it not obvious to identify its local features,which brings great difficulties to the grasping task.Therefore,this thesis conducts a research from both data set and detection algorithm aspects to carry out garment grasping point detection.The main work as following:(1)A high-quality simulation data set is made.Since there is much difficulty to collect a real data set,this thesis creatively proposes to generate large amounts of data using a simulation method.And clear depth pictures and accurate coordinate information of every grasping point can be obtained,by using a simulation software called Maya to imitate the natural hanging state of the garment.This simulation data is reliable by comparing with real one,which can provide high-quality data for the research of detection algorithm.(2)A garment grasping point detection algorithm is designed.This thesis designs a convolutional neural network model for grasping point detection research.First,using a garment classification model to determine garment type and then predicting the location and visibility of grab points with grasping points detection model.The former can maintain a high accuracy rate with few network parameters,the latter can predict both 3D coordinates and visibility information of grab points at the same time.Also,this thesis adopts Grad Norm algorithm to optimize the model by dynamically adjusting the loss weight of the two tasks,as there is a large difference in the magnitude of loss between coordinate regression and visible point classification tasks.Under the same network model,experiment results show that the Grad Norm algorithm improves the recognition rate of visibility,which can greatly increase the success rate of crawling tasks.(3)A data augmentation method based on GAN network is proposed.This thesis acquires real data using by Azure Kinect camera.Cycle GAN network is used to translate the simulated images to the real.According to the test results,the data generated by GAN network is closer to the real data than the simulation data.And the effect of data augmentation training using GAN network has been improved both in garment classification and grasping point detection.Using the deep learning method,this thesis predicts 3D coordinates of grasping points directly to perform actual grasping task.The accuracy of this paper reaches 93.97% and 87.07%in garment classification and visibility classification respectively.Also,the average error of grasping point detection is within 10 cm,which all shows the method proposed in this thesis has a practical significance,and can provide theoretical guidance for the task of garment grasping.