基于深度学习的3D目标检测算法综述

Overview of 3D Object Detection Algorithms Based on Deep Learning

随着自动驾驶领域对目标检测的精度和速度需求的提高,目标检测的研究从传统检测算法转向深度学习方向。由于2D目标检测算法存在小目标丢失等问题,基于深度学习的3D目标检测算法以能提供物体的位置、尺寸和方向等一些空间结构信息的优势,迅速在自动驾驶领域发展起来。首先,简单陈述了2D目标检测算法,将3D目标检测算法分成5个类别,分析了各类目标检测算法的优缺点;然后,详述了最新被提出的基于图神经网络(graph neural network, GNN)的2种算法;最后,对3D目标检测所应用的领域和其研究意义进行总结,并对3D目标检测今后可能发展的方向做出猜想。

With the increase in the accuracy and speed requirements of object detection in the field of automatic driving,target detection has shifted from traditional detection algorithms to deep learning.Due to the 2D object detection algorithm,there are problems such as the loss of small targets,the 3D object detection algorithm based on deep learning has the advantages of providing some spatial structure information such as the position,size,direction,and external shape of the object,and it has developed in the field of autonomous driving rapidly.Firstly,the 2D object detection algorithm is briefly stated,and the 3D target detection algorithm is divided into five categories,and the advantages and disadvantages of various object detection algorithms are analyzed.Then,two newly proposed algorithms based on graph neural network(GNN)are described in detail.Finally,the application field of 3D object detection and its research significance are summarized,and the future direction of 3D object detection may be guessed.