基于交叉视图和深度引导变换器的3D目标检测方法  

Cross-view and Depth Guided Transformer for 3D Object Detection

作　　者：夏红霞[1] 薛亚娣 XIA Hong-xia;XUE Ya-di(School of Software Engineering,Anhui College ofElectronic Information Technology,Bengbu 233000,Anhui,China;School of Digital Media,Lanzhou University of Arts and Science,Lanzhou 730000,China)

机构地区：[1]安徽电子信息职业技术学院软件工程学院,安徽蚌埠233000 [2]兰州文理学院数字媒体学院,甘肃兰州730000

出　　处：《兰州文理学院学报（自然科学版）》2024年第2期62-67,122,共7页Journal of Lanzhou University of Arts and Science(Natural Sciences)

基　　金：安徽省高等学校省级质量工程项目(2020xskt035);安徽省省级高校科研项目重点项目(2023AH053339);甘肃省科技计划项目软科学专项(20CX4ZA082);甘肃省高等学校创新基金项目(2020A-157)。

摘　　要：基于多相机拍摄的多视角图像,提出用于3D目标检测的深度引导变换器.首先,设计了轻量级的深度预测器,在无需额外的深度数据集下生成基于目标的稀疏深度图和低维深度特征;其次,开发了一种具有深度引导的跨视图变换器,用于融合深度特征和不同视图的图像特征,进而生成高质量的3D检测结果.实验结果表明,该方法的行人3D检测精度比现有的多相机检测方法高10%AP,多类别平均精度和和NDS指标均高出约3%,且整体运行性能比现有方法快近5倍.A cross view and depth guided transformer is proposed for 3D object detection.Firstly,a lightweight depth predictor is designed to generate object-level sparse depth maps and low dimensional depth features without requiring additional depth datasets.Secondly,a depth guided cross view converter is developed to fuse depth features and image features from different views,thereby generating high-quality 3D detection results.Experimental results show that the 3D pedestrian detection accuracy of this method is 10%AP higher than the existing multi camera detection methods,and the average accuracy and NDS index of multiple categories are about 3%higher,and the overall performance is nearly 5 times faster than the existing methods.

关 键 词：交叉视图 深度预测器 变换器 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]