Research On 3D Object Detection Algorithms Based On Attention Mechanism

3D object detection,which is to detect the location coordinates and categories of objects in3 D space,is an important research direction in computer field.With the development of point cloud acquisition hardware,3D object detection algorithm based on point cloud has been widely used in the fields of automatic driving,smart home and augmented reality.At present,deep learning network is often used to design detection algorithm for 3D object detection,and the feature extraction of point cloud by convolutional neural network is the key of detection.However,because convolutional neural network only uses local information to obtain features,it can not fully describe the object,and its detection results are prone to false detection and missed detection.In addition,the point cloud acquisition hardware cannot collect enough point cloud data to ensure the detection accuracy of such objects.In view of the above problems,based on the attention mechanism,this paper studies the acquisition of global features of feature map and implicit context information of objects.The main contents of this paper are as follows:(1)A single-stage 3D object detection algorithm based on global feature learning is proposed to solve the problrm of false detection and missing detection of the objects.Firstly,the sparse convolutional neural network is used to obtain voxel feature maps with three different scales,which can provide the local feature of objects.Secondly,the global information of feature maps of different scales is extracted by feature extraction module and cascaded and fused with the local feature.Finally,the detection network predicts the 3d detection boxes of the object.In the feature extraction module,the proposed algorithm applies the designed channel attention layer to obtain the global information of the channel dimention of three different scale feature maps and applies the designed self attention layer to obtain the global information of the spatial dimention of the minimum scale feature map.They calculate the corresponding similarity matrix in the channel and spatial dimension of the feature maps to update the channel features and spatial element feature,and then the feature map fusion module detect the objects by fusing the semantic and detail features of objects.Experimental results show that the algorithm has higher detection accuracy,which proves the validity and rationality of the module designed in this paper.(2)A two-stage 3D object detection algorithm based on implicit context learning is proposed to solve the problem that it is difficult to detect distant point cloud object.The algorithm uses sparse convolutional neural network to obtain voxel feature map of two different scales.Secondly,the feature map is input to the RPN network to get the initial proposal.Finally,the RCNN network is used to fine tune the proposals to obtain a more accurate detection box.In the RCNN stage,we design an implicit context learning module to learns the background features around the object area in feature maps of different size.The module uses different expansion factors to expand the size of the proposals,and integrates the object features in the original proposals and the expanded proposals according the cross attention mechanism,so as to avoid the interference of irrelevant information.Then we designed an attention fusion module to fuse multi-scale object features with implicit context.the module adds the object features of different scales,and performs global pooling in the channel dimension and spatial dimension respectively to obtain the global semantic information of the object in these two dimensions,and then calculates the fusion weight matrix corresponding to the semantic features.For the original input features of different scales,the fusion weight matrix of the corresponding grid features in the channel dimension is calculated.The global semantic feature and local detail feature of the fused object can be improved respectively by using two weight matrices,and then the fusion effect of multi-scale feature can be improved.Simulation results show that the algorithm has strong detection performance and better indicators.