Research On 3D Target Recognition Method Based On Feature Layer Fusion

3D target recognition has always been a challenging research topic in the fields of artificial intelligence,computer image processing and pattern recognition.Compared to two-dimensional images,three-dimensional descriptions are more accurate in understanding the real world because of rich visual and geometric information about objects or scenes.The point cloud data obtained by LiDAR provides spatial and geometric information of three-dimensional objects,but suffers from occlusion and incompletion due to data sparseness.It is not enough to analyze and understand three-dimensional objects only from projected images.The convolutional neural network(CNN)and restricted Boltzmann machine(RBM)are taken as the theoretical basis of this paper.This paper proposes a 3D target recognition method based on fusion of features extracted from multi-view image and 2.5D point cloud data,on which the high-level nonlinear relationship between between different representations of the same model to improve the recognition accuracy,then the shortcomings of several three-dimensional target recognition methods are analyzedthis paper studies a three-dimensional target recognition method based on feature layer fusion.The basic principles of deep learning.In this paper,the 3D CAD model library ModelNetlO is uesd as the model set,the model data is generated by model rendering and lidar simulation imaging,and then the convolutional neural network is designed from the model's multi-view 2D image and 2.5D point cloud,point cloud image is used for feature extraction,and the dual-channel CNN-RBM model is designed to fuse in the feature layer to obtain more representative three-dimensional object features.The experimental results show that the network fusion in the feature layer has a higher accurary in the recognition and classification process compared to single modal recognition results,indicating fusion of the visual features can effectively explore the nonlinear relationship between the two features and better represent the three-dimensional objects' characteristics.