Research On Point Cloud Recognition Based On Attention Mechanism And Graph Convolution

Point cloud recognition is the intelligent recognition of objects or scenes in space through point cloud data.Point cloud data is a kind of data that can derive geometric characteristics of objects in space,which make point cloud recognition play an important role in autonomous driving,robot space vision,etc.,and have important significance in human real life.However,point cloud data is different from pictures and text.Point cloud data has the characteristics of unstructured and disordered,which makes point cloud recognition a very difficult task.Conventional algorithms achieve point cloud recognition by manually extracting features and then using machine learning algorithms.It is difficult to achieve good recognition results.Meanwhile,the existing point cloud recognition methods based on deep learning mostly have the disadvantages of low scene versatility and low efficiency.Therefore,it is of great significance to design a point cloud recognition method which can directly process point cloud data with high accuracy and strong universality.In this paper,after studying the characteristics of point cloud data and the advantages and disadvantages of existing point cloud recognition methods,a point cloud recognition method based on attention mechanism and graph convolution is proposed,which directly processes point cloud data for recognition,and has the advantages of high accuracy and good universality.The main research work of this paper is as follows:First of all,the task of this paper is to use deep learning to identify the directly input point cloud data.The point cloud data is unstructured data.Different from the image data,which can be directly identified by convolution neural network,it needs to learn a continuous function and a symmetric function through the network to achieve the point cloud data fitting.This method is equivalent to let the network learn the key points in the point cloud data to summarize the entire object or scene.Due to the disorder of point cloud data,it is difficult for network to fit point cloud data and obtain key points of point cloud data.In this paper,we propose an attention mechanism combined with residual learning to improve the network's ability to fit point cloud data,and pay more attention to the features of key points in point cloud data,so that the model has a better recognition effect.Secondly,when using the direct input point cloud for recognition,due to the unstructured and unordered characteristics of the point cloud set,the algorithm also needs to consider the loss of local spatial feature information of the point cloud.In order to obtain local spatial features from unstructured point cloud data,this paper proposes a distance-based neighborhood graph for point cloud data,and then uses graph convolution to obtain local features of input point cloud data.Finally,the whole model of point cloud recognition is obtained by integrating the attention module combined with residual learning and the graph convolution of point cloud.On the basis of this model,this paper conducted experiments such as 2D point cloud recognition,3D point cloud recognition and 3D point cloud semantic segmentation,and achieved good performance.