Research On Key Technologies For 3D Point Cloud Tasks Based On Encoder-Decoder Network Architecture

In recent years,robotics,AR/VR,and intelligent driving have significantly benefited from the widespread use of point cloud data acquisition devices.Classification and segmentation of point cloud data are essential to achieve system environment awareness,which provide the necessary technical support for robot operation,vehicle path planning,and other activities.Typ-ical techniques of representation and network designs,however,are unable to strike a balance between reducing system complexity and improving accuracy due to the unique properties of point cloud data.Therefore,this paper proposes a lightweight,high-performance neural net-work PCRT for classification and segmentation tasks of point cloud data.The corresponding Encoder-Decoder networks are created to better fit the functions of varying complexity for the classification and segmentation of point clouds.Position embedding,downsampling,one-dimensional convolution,and relative attention modules are all included in the classification Encoder.The feature integration module for smoothing high-dimensional fea-tures is an addition made by the segmentation Encoder to the classification Encoder.The clas-sification Decoder uses some straightforward feature integration modules to decode features.In contrast,the segmentation Decoder restores the feature scale using a complicated upsampling module.It utilizes skip connections to add corresponding scale features in the Encoder for feature decoding.While building the encoder-decoder network,several feature extraction modules are used to produce lightweight network architecture based on various scale features.This is done to reduce the conflict between more accuracy and decreased complexity.To increase the accura-cy of extracting point cloud features,this paper’s network architecture focuses on avoiding the problem of point cloud geometric features losing information.The position embedding module explicitly describes the spatial position relationship between points using Manhattan distance.The module output is concatenated and mapped with the raw point cloud data.In the relative attention mechanism,the coefficientsρ~Kandρ~Vare obtained by calculating the relative in-formation between Query and Key or Value,updating the corresponding Key and Value,and finally getting the output of the close attention mechanism.In the classification task,this paper evaluated the performance of the PCRT network on the Model Net40 dataset.The experimental results showed that the complexity of the PCRT network is much lower than that of other networks,while achieving an overall classification accuracy of 93.4%,which is better than several existing network models.In the segmentation task,this paper evaluated the performance of the PCRT network on the Shape Net and S3DIS datasets and compared it with existing point cloud-based deep learning models.The experimen-tal results revealed that the PCRT network can achieve mean intersection over union of 86.3%on the Shape Net dataset,which is comparable to the best existing network models,and mean intersection over union of 68.2%on the S3DIS dataset,which is better than several existing network models.In summary,the PCRT performs well in reducing system complexity and improving sys-tem accuracy.It can support utilizing point cloud data in robot operation and vehicle path planning fields.