Research On 3D Point Clouds Semantic Segmentation Method Based On Deep Learning

The task of semantic segmentation of 3D point clouds based on deep learning is the focus of research in the fields of autonomous driving,target detection and tracking,industrial manufacturing,robotics,and so forth.With the development of hardware devices for collecting 3D point clouds,the acquired scene-level 3D point clouds are becoming increasingly complex.The current 3D point clouds semantic segmentation methods based on deep learning face the problem of insufficient feature utilization of point clouds when dealing with scene-level 3D point clouds,resulting in reduced semantic segmentation performance.Moreover,many valid values of manual annotations are required as training sets,making the labor cost higher.In contrast,reducing the number of accurate value annotations will lead to degradation in the training of network models.This thesis focuses on the semantic segmentation of 3D point clouds of different scenarios based on deep learning methods.In order to improve the performance of the 3D point clouds semantic segmentation network model,a 3D point clouds semantic segmentation method fusing multi-dimensional spatial information is designed,and the whole network structure is structured with six coding layers and ten decoding layers.A hybrid pooling is constructed to enhance the perceptual capability of the local region,using a short-hop multi-dimensional feature fusion module to increase the perceptual field of the network model in the local region.Local spatial coding with local standard vector information is constructed using standard vector estimation in the local region to enhance spatial features and improve the network model’s understanding of spatial information.A weakly supervised 3D point clouds based semantic segmentation method is designed for the degradation problem in training semantic segmentation network models due to the reduced amount of accurate value annotations.A consistent regularization training method for the network model after data enhancement is designed to reduce the under-fitting that occurs in training.A pseudo-label generation strategy is proposed,as well as the construction of a loss function for the influence of pseudo-labels on the network model in the process of participating in training to improve the smoothness of the network model training.This thesis conducts ablation and comparison experiments for the above algorithms on the S3 DIS dataset,Semantic3 D dataset,and Semantic KITTI dataset.The experimental results show that the fused multi-dimensional spatial information 3D point clouds semantic segmentation method can effectively improve the performance of 3D point clouds semantic segmentation in different scenarios,and the weakly supervised 3D point clouds semantic segmentation method can achieve the semantic segmentation results close to those of the fully supervised method with 0.1% truth annotation training.