3D Target Detection Of Point Cloud Based On Frustum

The point cloud is three-dimensional data acquired by a three-dimensional laser scanner.The number of points is dense and has high density and high accuracy.Therefore,in order to achieve more complex target detection of real three-dimensional scenes,it contains rich semantic information Point clouds are important research data for 3D target detection.Object detection is a research hotspot in the field of computer vision.It is mainly divided into two tasks: frame the bounding box of the target object in the input data,and determine the category of the object in the bounding box.2D object detection technology has achieved good results on mainstream public data sets,but 2D object detection performed on images lacks spatial information and it is difficult to extend it to 3D object detection.For complex scenes and smaller objects,The robustness is poor,and it cannot meet the needs of high-level detection tasks such as autonomous driving.Therefore,an efficient and intelligent 3D object detection model that directly uses 3D data such as point clouds has high research value.This article will use an end-to-end Frustum-Pointnets point cloud deep learning model to achieve 3D object detection,and improve and optimize the model to further improve the detection accuracy of the Frustum-Pointnets model.The main content of this article is as follows:(1)Summary comparison of mainstream target detection algorithms.Traditional target detection algorithms rely heavily on the accuracy of manual feature extraction.To address this problem,a comparative analysis of the current target detection algorithms implemented using deep learning networks can be divided into: target detection algorithms using images,The perspective detection target detection algorithm,the voxel-based target detection algorithm,and the original point cloud target detection algorithm,meanwhile briefly describe the implementation process of a representative algorithm.(2)A summary of the relevant theories of object detection algorithms based on deep learning networks.The internal components of the convolutional neural network and their functions are analyzed.The convolutional neural network is mainly composed of a convolutional layer,a pooling layer,a fully connected layer,and a classifier.(3)Construction of Frustum-Pointnets model and its improvement and optimization.Most of the current target detection algorithms implement 2D target detection on the image,and there are also other data forms such as point cloud conversion into multi-view images or voxels to achieve 3D target detection,however,there are significant limitations.In order to directly use the original point cloud data and avoid the loss of features caused by the transformation of the data form,this article will use the Frustum-Pointnets model to achieve 3D target detection using the original point cloud,while changing different activation functions and parameter initialization methods to further improve Detection accuracy of the model.Finally,training and verification are performed on a public large-scale real-life traffic KITTI dataset.The final experimental results show that the detection effect of the improved Frustum-Pointnets model has been improved.