Research On Point Cloud Processing Method Based On Deep Learning

With the rapid development of 3D scanning technology,the application of 3D point clouds is becoming increasingly widespread in various fields,such as face recognition,robotics,and autonomous driving.Feature extraction of point clouds is one of the key research topics in point cloud processing and is also the basis for subsequent point cloud applications.However,3D point clouds are unstructured,disordered,and sparse,which makes point cloud feature extraction challenging.In addition,how to maintain the displacement invariance and rotation invariance of the point cloud during processing also needs to be considered.This paper aims to improve the accuracy and robustness of point cloud feature extraction by incorporating a self-attentive mechanism with local feature extraction as a starting point,and investigates the effectiveness of the model in point cloud classification tasks and segmentation tasks on a public dataset.The main contributions of the thesis are as follows:(1)In response to the sparsity and disorder of point clouds,we propose a Local Neighborhood Transformer(LNT)model for point cloud,which can effectively extract point cloud local structure information and achieve linear complexity by dividing the point cloud into local neighborhoods and learning features in the local geometric structure using the Transformer self-attention mechanism.To utilize low-level feature information,we also design a point cloud multi-feature fusion module to aggregate information from different layers.The experiments show that the model achieves an overall accuracy of 93.3% and an average class accuracy of 92.0% on the ModelNet40 point cloud classification task.The average intersection-over-union(Io U)score in the ShapeNet point cloud part segmentation task is 85.2%.Compared to similar self-attention models,this model significantly reduces spatial complexity,with only 6M parameters.(2)In response to the irregularity of point clouds and the drawbacks of the Local Neighborhood Transformer model,we propose a Dynamic Adaptive Position Feedback(DAPF)model for point cloud feature extraction,which consists of an adaptive position feedback module and a feature extraction module.The adaptive position feedback module learns the displacement of each point based on its feature and position information to achieve adaptive position feedback,moving the position of the point in the direction of regularization and enabling the model to learn more effective local features.The feature extraction module reduces the parameters of the Local Neighborhood Transformer model while maintaining performance.The two modules work together,alternating between each other to learn the semantic features of the point cloud.Experimental results show that the model achieves an overall accuracy of 93.4%on the ModelNet40 shape classification task,an average Io U of 85.8% on the ShapeNet part segmentation task,and an overall accuracy of 86.5% on the S3 DIS scene segmentation experiment.The model has only 4M parameters.