Research Of 3D Instance Segmentation Technology On Deep Learning

In the field of intelligent manufacturing,how to use three-dimensional vision sensors to perform three-dimensional modeling and segmentation of randomly placed workpieces is the basis for the use of robots to accurately locate and grab workpieces to achieve highly automated production.Since deep learning technology was first applied to the field of 3D instance segmentation in 2018,relevant research has made preliminary progress,but there are still problems of low segmentation accuracy and slow speed.Especially in the field of intelligent manufacturing,small-sized workpieces also have technical difficulties such as low texture,small volume,and easy occlusion.Therefore,this article takes DB9 nine-pin serial communication interface parts as a typical application object to study the three-dimensional segmentation of densely placed workpieces.The main contents of this article are as follows:(1)Summarizes the current research status of 3D instance segmentation technology,and expounds related algorithm principles to provide theoretical support for the research content of this article.(2)Aiming at the problem of outliers and redundant planes in the original data collected by the sensor,statistical filtering and Ransac algorithm are used to preprocess the point cloud data to provide high-quality training data for the segmentation network.At the same time,in view of the time-consuming problem of manual labeling in the instance segmentation data set for constructing artifacts,an algorithm for synthesizing the artifact instance segmentation data set from manual labeling data is proposed,and the synthetic data set is compared with the artificial labeling data set.,To verify the effectiveness of the data set generation algorithm.(3)Proposed an instance segmentation network IP-BoNet(Industrial PartsBounding BoxNet)for scattered workpiece scenes.The network is based on 3D-BoNet,and has been improved in four aspects according to the characteristics of the point cloud in the scattered workpiece scene,including: 1)The workpiece point cloud is large in scale,and in the bounding box regression branch of 3D-BoNet,due to Only global features are used as the input of bounding box regression,resulting in a lack of local information.In response to this problem,this paper fuses multiple down-sampling layers of the network by adding the features of each point to increase the local information for the bounding box regression and improve the accuracy of segmentation;2)The workpiece point cloud density is uneven,and 3D-BoNet does not consider the change of point cloud density in the feature extraction process.In response to this problem,this paper designs a density-weighted feature extraction module based on the attention mechanism.This module can give higher weight to point cloud features in low-density areas,allowing the network to learn more expressive features;3)For the problem of using fixed-size neighborhoods in the network downsampling process,part of the point cloud cannot be used.,This paper uses the hole convolution to expand the neighborhood of the sampling center point,so that the feature map can obtain a larger receptive field,and can make full use of the input point cloud data;4)The sampling algorithm in the 3D-BoNet network is the farthest point Sampling,as the number of point clouds increases,the amount of calculations increases sharply,which does not meet the real-time requirements in the context of industrial production.In this paper,the farthest point sampling and random sampling methods are used in the high-level and low-level of the network,respectively,to improve the calculation speed.Finally,a series of experiments were carried out on the workpiece data set,which showed that the network proposed in this paper has higher accuracy and speed than 3D-BoNet,and verified the effectiveness of the improved method.