Research On 3D Point Cloud Vehicle Recognition Algorithm Based On Deep Learning

With the development of deep learning and the popularization of artificial intelligence technology in the new era,driverless car has become a popular research direction in deep learning.The breakthrough of this technology will be able to solve the increasingly serious traffic congestion problem.The main research of this paper is to use the three-dimensional point cloud collected by lidar as data to complete the task of road vehicle and obstacle recognition.It mainly includes a vehicle detection algorithm based on spatial map convolution,a vehicle detection algorithm based on Point Net++,and a vehicle detection algorithm based on a full convolutional neural network for bird's-eye point cloud images,in order to obtain a three-dimensional detection model that can predict vehicles more accurately.Firstly,When extracting point cloud features,most of the current 3D detection algorithms adopt a method of processing each point separately,which makes it easy to ignore local features.This paper proposes a vehicle detection algorithm based on spatial map convolution,which constructs a local neighborhood map in a three-dimensional point cloud to connectadjacent points,and performs convolution calculations on the opposite edges between adjacent points to use the local points geometry structure.The local neighborhood graph varies with the change of the network layer.Combined with the features of the point cloud image,a novel local feature aggregation module and a local feature learning module are used to gradually increase the receptive field of each 3D point and change the geometry of different layers.The aggregation of feature information,image information and global features enables accurate prediction of vehicles,pedestrians and bicycles.Secondly,Aiming at the sparse problem of 3D point cloud,this paper proposes a vehicle detection algorithm based on Point Net++ network.The improved segmentation network is used to obtain high-quality front scenic spots and background points in the three-dimensional point cloud.The high-quality front scenic spots represent the entire scene information,and the threshold is set to further divide the target points from the previous scenic spots.The average size of the vehicle is defined as the distance threshold to gather the vehicle points,calculate the local spatial characteristics between the gathered points,retain the detailed information of the vehicle,and improve the accuracy of vehicle detection.Thirdly,The objects in the bird's-eye view(BEV)do not overlap each other,and the measurement space is reserved.Therefore,this paper proposes a vehicle detection algorithm for bird's-eye point cloud images based on a fully convolutional neural network,and uses an improved segmentation and regression network to predict the three-dimensional point cloud pixel by pixel.Use a smaller number of channels in the low-level convolutional layer to extract more detailed information.The high-resolution feature map is combined with the low-resolution feature map to improve the problem of high false detection rate caused by too few vehicle pixels after downsampling,and better realize the vehicle detection function.