Research On 3D Object Detection Algorithm For Autonomous Driving

3D object detection can accurately predict the spatial position,shape and type of objects,which plays an important role in the development of autonomous driving technology.Due to the advantages of simple acquisition,easy storage,and ability to provide accurate 3D spatial information,point cloud has become the most important data format in the field of 3D object detection.However,due to the sparsity and disorder of point clouds,there are still many deficiencies in the current 3D object detection algorithms based on point clouds.Therefore,in order to improve the detection accuracy of the algorithm,thesis studies the 3D object detection algorithm based on point cloud based on deep learning.The main research contents of thesis include:(1)The 3D object detection algorithm based on multi-scale voxels was studied.To solve the problem of disorder and irregularity in point cloud data,the multi-scale voxels were used to divide the point cloud first,and then the voxel feature extraction network was used Feature extraction is performed on multi-scale voxels to obtain more accurate point cloud features;for the problem of low network detection accuracy,sparse convolution is introduced to construct a convolutional intermediate network,and sparse convolution and submanifold convolution are used to aggregate voxel features,so as to obtain point cloud features with a larger receptive field and richer object shape information.(2)The research proposes a feature fusion algorithm based on attention mechanism.In the field of target detection,in order to accurately predict the spatial position and classification confidence of objects,it is necessary to consider both spatial features and semantic features.Therefore,thesis proposes a feature fusion algorithm based on attention mechanism.The algorithm first uses the attention mechanism to adaptively fuse highlevel spatial features and low-level semantic features,and then uses the spatial attention mechanism to enhance the fusion features so that it pays more attention to the features of key areas,thereby improving the positioning accuracy of the network.(3)The study proposes a 3D object detection algorithm based on point cloud and voxel fusion.In the first stage of the algorithm,the efficient voxel detection method is used to extract multi-scale voxel features,and an initial suggestion frame is generated,and then the key point feature encoding network is used to Encode the entire point cloud scene into a certain number of keypoints.In the feature encoding stage,first use the farthest point sampling algorithm based on semantic features to down-sample the point cloud,and then use the feature aggregation algorithm based on self-attention mechanism to aggregate multi-scale voxel features to key points,so as to obtain Keypoint features with precise location information and rich context information.In the second stage of the algorithm,first use the key point weight adjustment module to reweight the weight of the foreground points,and then use the suggestion box optimization network to aggregate the key point features into the region of interest features,and use the region of interest features to predict the category of the object Confidence and bounding box regression parameters,enabling optimization of initial proposal boxes.Finally,experiments and result analysis are carried out on the public KITTI dataset,and the results show that the3 D object detection algorithm proposed in thesis has high detection accuracy.