Research On 3D Object Detection Based On Point Cloud Scene In Complex Traffic Environment

The outbreak of COVID-19 has brought unprecedented opportunities for the development of self-driving cars,and self-driving technology has also developed rapidly.Autonomous driving technology is not only needed in autonomous vehicles,but also widely applied in many intelligent technology products,such as disinfection robots and food delivery robots in the front line of fighting against the epidemic.In this paper,3D point cloud collected by lidar is mainly used as data.In view of the important position of convolutional neural network and Transformer mechanism in object detection,a key link in unmanned driving technology is studied,which is to perceive and identify objects around through target detection algorithm.Nowadays many kinds of 3D object detection algorithm based on point cloud data is not like a two-dimensional detection algorithm can in actual application,the reason is because the three-dimensional point cloud data has irregularity and sparse,so in view of the complex traffic environment based on point cloud scene 3D object detection,this paper puts forward 3D detection algorithm is of great significance.Aiming at the difficulty of feature extraction in 3D object detection,this paper focuses on a single-stage 3D target detection method that integrates spatial semantic features.The sparse features of point cloud are extracted by manifold sparse convolution network,and then the spatial semantic features of detected objects are extracted respectively by spatial semantic convolution layer.A multi-channel fusion module based on attention mechanism is designed to fuse the two features,and the fused features were output for feature prediction.Then in order to effectively deal with the empty voxels sparse feature,paper also draw lessons from the sparse convolutional coding idea,according to dimensional sparse convolution and submanifolds convolution structure,studied the sparse voxel module and the flow form module,input and output is not empty voxel position is the same,in expanding the receiving domain while maintaining the original 3D structure.At the same time,multi-head attention mechanism in Transformer structure is used to capture different contexts in different scopes in sparse voxel module and sub-fluid voxel module,and the search process of non-empty voxel is accelerated through fast voxel query to obtain the final feature.By combining the idea of multi-scale feature hierarchy with Transformer model,this paper studies the multi-scale fusion expression of voxel and point cloud,and solves the problem that the feature representation of Bird’s Eye View(BEV)is not synchronized.And can effectively cross represent with different levels of semantic features.With point-by-point correlation Transformer,the fusion multi-scale 3D detection framework learns to capture deeper local-global structures and richer geometric relationships between point clouds.The paper proposed a single-stage 3D target detection method,TFAF-SSD,which adaptively integrates high-level abstract semantic features and low-level spatial features.And The paper proposed VFTNet,a voxel feature extraction network based on Transformer mechanism,and 3D object detection based on multi-scale Transformer.The validity of the proposed method is verified in KITTI dataset and Waymo Open dataset,and compared with other 3d target detection methods.