Research On Detection And Localization For Clustered Industrial Objects Based On Point Cloud Processing

Highly automated manufacturing is one of the most important goals of modern industry,and most critical challenges of that is enable robots identify and locate industrial parts intelligently.In the domain of industrial production,industrial robots need to grab the object workpiece instances one by one from randomly piled up industrial parts to the next working position,which could replace the repetitive operation of removing the workpiece from the material box manually.In order to achieve this goal,the robot needs to estimate the 6 degrees of freedom(6Do F)pose of the object instance quickly and accurately.The 6Do F pose consists of 3 degrees of freedom spatial position information and the pose information represented by 3rotation angles.This task of automatically picking objects is referred as bin picking problem.However,in real industrial scenes,lots of identical industrial parts are scattered and stacked together,and there are serious occlusions,which poses a huge challenge to the detection and localization for industrial objects.Moreover,as the appearance of industrial products becomes more and more complex,existing algorithms can no longer perform the detection and localization of complex objects,especially small objects.In the research field of bin picking,there are few researches for small-sized industrial parts.Since industrial parts have complex shapes,small volumes,textureless surfaces and mutual occlusion,which cause industrial parts are difficult to detected.This article starts from practical applications and combines point pair feature and deep learning to study bin picking in industrial scenarios.the main research work and innovations are as follows:(1)Due to the serious occlusion of scattered workpieces and the interference with lots of similar features,the mainstream methods of obtaining the objects 6D pose based on RGB images which directly resort to regression or classification is not ideal and feasible in the above-mentioned complex and changeable industrial scenes.In order to solve this problem,this paper resort to the 3D laser point cloud and propose a point cloud segmentation method suitable for scattered industrial parts based on Euclidean clustering.The algorithm first designs a detection method for contact edge points to solve the problem of close contact between workpieces.The contact edge points are filtered out by the mean value of the normal angle and the point density of the neighborhood of the point cloud,so as to solve the under-segmentation problem caused by the contact between the workpiece.Secondly,a distance threshold adaptive Euclidean clustering is proposed to deal with the uneven point cloud density more accurately.The experimental results on our own workpiece instance segmentation dataset show that the method proposed has better segmentation performance.(2)Based on the Hough voting and point pair features,an improved 6Do F object pose estimation method for the workpiece is proposed and an industrial scattered multi objects detection and localization software system is designed.The algorithm first introduces an adaptive point cloud down sampling method based on the normals angle,which preserves the rich geometrical features while ensuring the compression rate,and provides a basis for the balance of speed and accuracy.In addition,in order to prevent a large number of redundant features on the plane region from affecting the voting results,constraints were imposed according to the normal features of point pairs during feature extraction to avoid the pose estimation results falling into local optimum.At the same time,in order to eliminate the influence of sensor noise,a new voting method is designed to enhance the robustness of feature matching.Then,a pose optimization method based on hierarchical complete linkage clustering and iterative 3D-NDT was introduced to further improve the speed and accuracy.The comparative experiments on our own DB9 connector dataset and public dataset LINEMOD show that the algorithm proposed can effectively detect and locate industrial objects in clustered scenes.(3)Although the pose estimation method based on point pair feature and Hough vote has good time performance,it is easy to make the Hough vote deviation and fall into the local optimal result due to the large number of repeatable geometrical structures of scattered industrial objects.To solve this problem,a 3D local geometry descriptor matching network based deep learning is proposed in this paper,which extracting the 3D local geometry features to build the matching relationship between the object and the model,and improve the accuracy of pose estimation and robustness with occlusion and sensor noise.Firstly,point pair features are used as network input to obtain complete rotation invariance and avoid the calculation of local reference frame(LRF).Secondly,a new point cloud autoencoder is designed by combining Point Net and Tearing Net to deal with disordered and sparse point clouds with arrangement invariance.At the same time,this network is able to weak-supervised training,and the dataset are not required for densely labeling.Finally,in order to further improve the discrimination of the network,the augmented Chamfer distance(CD)is introduced as the loss function of the autoencoder.Experiments show that the performance of the proposed method are higher than existing algorithms on the DB9 dataset and the public dataset 3DMatch.