Research On Target Semantic Segmentation And Classification Based On Point Cloud Information

With the full bloom of deep learning in the field of computer vision and the popularization of visual sensors,deep learning of 3D point clouds has become a hot issue nowadays.Point cloud deep learning includes many tasks such as point cloud target classification,semantic segmentation,instance segmentation,etc.,which promotes the rapid development of related fields.At present,the robot can use the depth camera to collect 3D point cloud information.The collection process is combined with the robot’s running trajectory to advance from the scene to the individual.Inspired by the robot vision acquisition task,this paper mainly studies the classification and segmentation of point cloud learning tasks.Starting from the classical research methods,the overall point cloud feature learning process is explored from the big to the small.The shortcomings of the classic methods are explored in detail,the network module improvement is implemented in depth,and the effect of the improved network is evaluated by using a wealth of ablation experiments.The specific research contents of this paper are as follows:First,this paper provides an overview of the properties of common point cloud datasets and applications including depth camera acquisition and the Blensor simulation platform.According to the theoretical basis of depth camera imaging and measurement,the debugging and image acquisition process of the depth camera is introduced,and the simulation process is introduced in detail after studying the usage rules of the Blensor platform.Second,this paper introduces the classic point cloud feature learning algorithm,and uses public data sets to complete experiments such as object classification,component segmentation and indoor scene segmentation.The experimental results need to be measured in three aspects:accuracy,time complexity and model scale.This paper draws lessons from the classical feature learning network and absorbs the network structure settings,and at the same time evaluates the problems that still need to be solved in the network model from the experimental results.Based on this,this paper regards the classification of residual and defect clouds as the primary task of network improvement.Third,this paper proposes an improved method for the classification problem of the incomplete point cloud,selects PointNet++as the backbone network,and adds a data augmentation module on the basis of this network to expand the data richness,and adds a similarity measurement module to introduce auxiliary loss to guide network training and add the label smoothing strategy enhances the network’s robustness to outliers.This paper uses ModelNet40 to verify that the improved network model is still effective in the general fullangle point cloud classification problem.At the same time,the ModelNet40 data set is specially processed to generate a incomplete point cloud test set to evaluate the classification effect of the improved network model.Finally,this paper introduces the complete ablation experiment process.Use open source data sets to design experiments,determine the values of hyperparameters in the data enhancement module and the similarity measurement module,and add improvement modules one by one to determine the independent effectiveness and synergy between the modules.In order to comprehensively evaluate the effectiveness of the improvement,this paper first uses a complete 3D model to obtain surface points to make a full-angle point cloud data set for training;uses depth camera shooting combined with Blendor simulation to build a singleangle point cloud data set for testing.Secondly,the design experiment explored the feasibility of simulation instead of real shots to complete network training in practical applications.This paper also designs robust exploration experiments for the problem of sparse sampling points,and comprehensively evaluates the improvement effects.