| With the continuous development of Li DAR 3D sensors and related software,processing and analyzing 3D models has become a popular direction in the field of computer vision.As 3D data can effectively represent shape information and has a wide range of applications in fields such as autonomous driving,remote sensing and robotics,the processing and analysis of 3D models is of great importance.With the booming development of deep learning in the field of 2D data,attempts have been made to apply deep learning techniques to the field of 3D data.For convolutional neural networks used for image classification,3D data can be represented in a variety of ways,and one important representation is the point cloud model.However,due to the sparse,unstructured and disordered nature of point clouds,traditional convolutional neural networks are not suitable for direct application to the point cloud domain.In recent years a number of methods have been proposed for processing and exploiting point cloud information.However,most of them define local regions in a relatively simplified or rough way,without taking into account the fact that there are complex and tight connections between features inside and outside the local regions and at different levels.The main innovative work in this paper to address the above problems is as follows:(1)To address the point cloud classification problem without considering local feature information,this paper proposes a point cloud classification network based on local location information for feature extraction and fusion,which can calculate local features based on the relative position relationship between each point in the point cloud and its neighboring points,and can adaptively adjust the neighborhood range and density to better capture the detailed information of different scales and shapes in the point cloud.Experiments on publicly available datasets demonstrate the effectiveness of the model,with an accuracy of 92.8% on the Model Net40 dataset.(2)For the traditional feature extraction method,this paper proposes an algorithm to obtain distance weights based on nearest neighbor sampling,which can assign different weights according to the distance between each point in the point cloud and its neighboring points,thus weakening the interference of distant points for local feature extraction and relatively strengthening the contribution of near points for local feature extraction,helping the network to better learn meaningful features in the point cloud.The increased perceptual field and the ability to better capture contextual information effectively improve the accuracy of the model.(3)The feature fusion idea is further used to optimize the feature extraction process and realize multi-level feature fusion.Specifically,in this paper,global features and local features are fused in each layer of the network,and cross-layer fusion is also performed between different layers,thus enabling the network to make full use of the information at different layers and enhancing the expressiveness and generalization ability of the network. |
