Point Cloud Segmentation With 3D Model For Industrial Scenarios

With the development of the social economy and computer technology,industrial manufacturing requires higher productivity.Based on this background,intelligent industrial robots play an increasingly important role in the field of industrial manufacturing and have a broader scope in the fields of electronics,machinery,logistics,etc.They can be widely used in industrial tasks such as intelligent destacking,workpiece loading,goods grabbing,article positioning,defect detection,size measurement,etc.In these industrial tasks,how to get an accurate perception of objects is the most important part of the entire task.However,with the development of 3D scanning technology point cloud data came out,which can provide rich geometry and scale information.Due to the lower cost and less impact of light,it has become an important research area in computer vision.Based on task demands we mentioned before,how to accurately segment the point cloud of objects is the key of the task,as well as the basis for industrial robot control and other parts.In industrial tasks,most of the objects are rigid objects,and most of the objects are usually produced according to a known designed 3D model,and objects of the same kind stacked on each other make up the scene.The main research of this article is to face this special industrial scene,which mainly focuses on segment objects of the same kind 3D model from the industrial scene.In this article,we first collect a set of data in real industrial scenarios that meet the task requirements and propose a method for labeling point cloud data automatically based on traditional point cloud processing methods such as edge detection,conditional Euclidean clustering,ICP registration,and so on,and construct a point cloud segmentation dataset with the 3D model of objects for training and evaluating neural networks.Based on that condition,a neural network framework for point cloud segmentation using the information of the 3D model of objects is proposed.Because the number of instances cannot be known in advance for the instance segmentation task,the neural network proposes a method of learning the center point of the object instance and the pose information relative to the object model position.Based on the characteristics of knowing the 3D model of the object,a model-based attention mechanism module using the information of the 3D model of the object is proposed.Our neural network is designed based on these two main ideas to learn the relevant semantic information,center point information,and relative pose information.Finally,we filter the center point information and relative pose information predicted by the neural network through a post-processing process based on clustering optimization,and then use the prior bounding box based on the object model to segment and select the point cloud to obtain the final point cloud segmentation result.According to the experiments,the point cloud segmentation dataset with object 3D models constructed in this paper based on real industrial robot application scenarios can provide a data basis for the use of deep learning-based methods to solve specific point cloud segmentation problems in industrial tasks.The proposed point cloud segmentation method based on neural networks using object model information can obtain point cloud segmentation results similar to the state-of-the-art point cloud segmentation methods,and the geometric structure information of the object instances in the segmentation results is more accurate.The center point of the object instance and the pose information relative to the object model position predicted by the neural network can be directly provided to the industrial robot,which can provide the visual information for industrial robots to finish tasks such as sorting,grasping,etc.