近景远景多维度融合的三维目标检测算法  

Three Dimensional Object Detection Algorithm Based Close Range And Distant View Fusion

作　　者：薛俊 赵铎涵 陶重犇[1,3] 王琛 XUE Jun;ZHAO Duo-han;TAO Chong-ben;WANG Chen(School of Electronics and Information Engineering,Suzhou University of Science and Technology,Suzhou Jiangsu 215009,China;College of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing Jiangsu 210003,China;Tsinghua University Suzhou Automotive Research Institute,Suzhou Jiangsu 215134,China)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215009 [2]南京邮电大学通信与信息工程学院,江苏南京210003 [3]清华大学苏州汽车研究院,江苏苏州215134

出　　处：《计算机仿真》2024年第12期234-239,249,共7页Computer Simulation

基　　金：国家自然科学基金(62201375);中国博士后自然科学基金(2021M691848);江苏省自然科学基金(BK20220635,BK20201405);苏州市科技项目基金(SYG202142)。

摘　　要：针对用于自动驾驶的三维目标检测算法特征层融合深度低,从而影响检测精准度的问题,提出了一种近景远景多维度融合的三维目标检测算法。算法分为两个阶段,第一阶段利用CenterNet检测得到粗粒度的目标信息,结合鸟瞰特征对ROI区域和检测任务进行划分。为了加强对全局的感知能力,并且更充分利用速度信息,第二阶段分为近景检测和远景检测,分别利用点级特征和径向速度作为先验信息,设计不同的特征融合模块,生成检测结果。在NuScenes和KITTI数据集上进行了实验,实验结果表明,上述方法较其它先进算法精确度更高、鲁棒性更好、泛化能力更强。Aiming at the problem of low fusion depth of feature layers in 3D object detection algorithms for autonomous driving,which affects detection accuracy,a 3D object detection algorithm based on close range and distant view multi-dimensional fusion is proposed.The algorithm is divided into two stages.The first stage uses CenterNet to obtain coarse-grained target information,combined with bird's-eye view features to delineate the ROI region and detection task.In order to enhance the global perception capability and to make better use of velocity information,the second stage is divided into near-view detection and distant-view detection,using point-wise features and radial velocity as a priori information respectively,and designing different feature fusion modules to generate detection results.The experiments were conducted on the NuScenes and KITTI datasets,and the results show that the above method has higher accuracy,better robustness,and stronger generalization ability compared to other advanced algorithms.

关 键 词：自动驾驶 三维目标检测 多模态数据融合 速度信息回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]