Research On Object 3D Reconstruction And Location For Robotic Sorting Task

With the rapid development of electronic commerce,the number of orders for small goods is increasing in the logistics warehouse,but sorting step needs lots of manual work.Because there are a large number of new goods enter in warehouse every day and the environment of warehouse is complicated,it is difficult for traditional auto-sorting line to solve the problem of sorting kinds of small goods.And manual sorting has better recognition accuracy and stability.However,with the trend of “robot substitution” and rising labor cost,using robot to complete sorting task has become a new research hotspot.In general,firstly,robot sorting system needs to find the target object in the scene,and calculate the 6D pose of it.Finally,it controls robot arm to grab target item.For rapid circulation of goods in logistical warehouse,logistical robotic sorting system has three main problem: how to capture new object features like shape,color and size in a short time;how to form labelled object's image dataset for training algorithm in a short time and how to realize object recognition and pose estimation in complex environment with lower computation complexity and training time.Therefore,this paper proposes an object localization and grabbing system which is based on 3d reconstruction and combing deep learning network with point pair features.The main content includes the following aspects:1)In order to capture new object's features fast,this paper designs an object fast 3D reconstruction system.This system uses RGBD cameras and determines the initial pose of camera through QR code.It optimizes camera pose by using frame-to-model ICP and fuses different view points' point clouds with TSDF method.At last,it reconstructs object colored mesh model with poisson reconstruction.2)As for forming the labelled object image dataset,this paper designs an object rendering method based on OpenGL.Through setting object model in virtual sampling ball and changing environment light condition,it can capture object model's pictures from different view-points.At last,it merges model's pictures with random background and generates file with object class and bounding box position to form object's image dataset.3)As for object detecting and pose estimation in complex environment,this paper proposes combining deep learning network with traditional point pair feature to calculate target object's 6D pose.It uses auto-formed image dataset to train object recognition network and segment scene point cloud according to recognizing result.In the point pair feature pose estimation,hash table of object point pair features is generated with object's 3d model.And point pair feature exampled from segmented scene point cloud can find matched point pair feature from object model in hash table.By using this way,matched point pair feature,which means a hypothsis of object pose,can vote for the last 6d pose of target object.Finally,this paper designed sorting experiment in warehouse environment and analyzed the object recognition rate,repeated sorting rate and successful sorting rate for verifying the system's sorting effect.It is not only proving the sorting result,but also illustrating the application value...