Research On 3D Point Cloud Recognition Method Based On Deep Learning

In the field of computer vision,3D point cloud recognition technology is an extremely important research direction.In recent years,the method based on deep learning is outstanding in the task of 3D point cloud recognition,which has great advantages over the traditional point cloud recognition method.However,the current 3D point cloud recognition method based on deep learning still has the problems of weak local structure feature expression ability and high time and space complexity of the network model.Therefore,in order to further improve the recognition ability of the network model and reduce the time and space complexity,this paper studies the 3D point cloud recognition method based on deep learning.The main research work is summarized as follows:(1)A three-dimensional point cloud recognition method based on Bi-directional convolutional long short-term memory network is proposed to obtain the correlation of multi-scale features in the local a 3D point cloud.Bi-directional convolutional long shortterm memory network is used to capture the correlation of multi-scale features of local regions to enhance the expression ability of local structural features.Bi-directional convolutional long short-term memory network is composed of forward branch,backward branch and fusion unit.The forward branch and the backward branch learn the correlation of the multi-scale features of the local area according to the scale from small to large and from large to small,and the features of the forward and backward branches are fused through the fusion unit.In addition,the dense connection method is introduced into the convolution layer to reuse the local structure features of the Bi-directional convolutional long short-term memory network output to obtain more abundant local structure features.The final experiment shows that compared with the current deep learning-based 3D point cloud recognition method,the proposed method can obtain higher recognition accuracy,indicating that the local structural features extracted by the method have stronger expression ability.(2)In order to reduce the time and space complexity of 3D point cloud recognition method,a lightweight 3D point cloud recognition method is designed.This method uses an improved K-nearest neighbor graph to represent the local structure of the point cloud.Compared with using each point in the original input point cloud as the node of the graph,the sampling points obtained by using the farthest sampling method as the node of the graph can be greatly Reduce redundant local structures and reduce the computational complexity of the network model.1 * 1 convolution is used to extract the edge features of the graph,and the attention mechanisms is used to aggregate the edge features of the graph to enhance the local structure features of the point cloud.In addition,considering that using multi-layer 1 *1 convolution to further extract high-level local structural features will increase the amount of network model parameters,a grouping convolution module is proposed to reduce the parameter amount of the convolution layer under the premise of ensuring effective feature extraction.The experimental results show that the recognition accuracy of this method reaches the same or better level of current mainstream methods,and the time and space complexity is greatly reduced,while maintaining high robustness.