Research On Point Cloud Registration Algorithm Based On Feature Fusion

In recent years,with the continuous development of the 3D data acquisition equipment,computers have begun to perceive and understand the world in a new way.Since the point cloud acquisition device cannot obtain complete object information through only one scan due to problems such as perspective and occlusion when collecting the same object,it is necessary to scan the object from different perspectives.Due to the different viewing angles,the data of different parts of the same object are in different spatial coordinate systems,point cloud registration technology is needed to map the point clouds in different reference systems to the same coordinate system to achieve stitching between point clouds.With the great success of deep learning in the field of 2D vision,researchers began to try to apply related technologies in the field of 2D to the field of 3D vision.However,due to the disordered and discrete data structure of point cloud,deep learning methods in 2D images cannot be directly applied to point cloud.Deep learning hadn’t begun to be widely used in the field of point cloud processing until the PointNet was proposed.In recent years,a lot of research work on point cloud registration technology has emerged,but most of the work hasn’t fully captured the structural information of the point cloud.Therefore,this paper proposes a point cloud registration algorithm based on feature fusion.On this basis,a hybrid attention mechanism is added to enhance the feature bias of the network to the high salience within the point cloud,and the relevant information between the point clouds is introduced to improve the registration performance of the network.The content of this thesis is as follows:(1)In order to avoid the problem of insufficient use of point cloud structure information,this thesis proposes a point cloud registration network model based on fusing global and local feature.The model mainly includes the global feature extraction network,the local feature extraction network.In the global feature extraction network,the transformation network T-Net is removed based on the PointNet network,and the skip connection is introduced to enhance the network’s perception of low-level and middle-level features.In the local feature extraction network,KNN is firstly used to determine the nearest neighbor nodes of the center point,so as to construct the local graph structure of the center point,and the edge convolution operation is used to extract the edge of the obtained local graph.Adaptive fusion is used instead of the traditional vector splicing for feature fusion to enrich feature information as much as possible.(2)In order to improve the saliency of features in point clouds and solve the problem of mutual isolation between point clouds,this thesis designs a point cloud registration algorithm based on hybrid attention and feature fusion.The hybrid attention module is mainly composed of self-attention mechanism and cross-attention mechanism,in which the introduction of the self-attention mechanism is to consider that in a single point cloud,a high weight value is assigned to the point feature with high saliency,and a low weight value is assigned to the insignificant feature,In order to reduce the influence of insignificant features on the final result as much as possible in the subsequent processing flow.At the same time,the introduction of the cross-attention mechanism can mine the correlation information between point clouds,which is convenient to guide the subsequent search for correct matching point pairs.The corresponding experiments are designed to prove that the proposed algorithm can effectively improve the registration accuracy of point clouds.