挑战性环境下基于双尺度CBAM的毫米波雷达与视觉特征融合目标检测  

Object Detection in Challenging Environments via Dual-scale CBAM Feature Fusion of mmWave Radar and Vision

作　　者：任坤 李盼 韩红桂 REN Kun;LI Pan;HAN Honggui(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Engineering Research Center of Digital Community,Ministry of Education,Beijing 100124,China;Beijing Laboratory for Urban Mass Transit,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]数字社区教育部工程研究中心,北京100124 [3]城市轨道交通北京实验室,北京100124

出　　处：《北京工业大学学报》2025年第3期284-294,共11页Journal of Beijing University of Technology

基　　金：国家重点研发计划资助项目(2023YFC3904605);国家自然科学基金资助项目(62125301,62203022)。

摘　　要：针对恶劣天气和低光照对基于深度学习的视觉目标检测算法带来的挑战,提出一种基于双尺度卷积注意力模块(convolutional block attention module,CBAM)的双模态目标检测算法,旨在通过视觉与毫米波雷达数据的特征融合,提高目标检测算法在挑战性环境下的鲁棒性和准确性。该算法采用双分支的一阶段检测结构,图像分支采用预训练的CSPDarkNet53骨干网络提取图像特征,雷达分支采用基于体素的雷达特征生成网络提取雷达特征。然后,分别在颈部网络前后利用提出的基于双尺度CBAM的特征融合模块进行雷达-视觉特征融合。最后,使用解耦检测头实现目标的分类和定位。在nuScenes数据集上,利用对比实验和消融实验验证了该特征融合检测算法在挑战性环境下的有效性和优越性。A dual-modality object detection algorithm,based on the dual-scale convolutional block attention module(CBAM),is addressed to tackle challenges posed by adverse weather conditions and low lighting for visual object detection algorithms based on deep learning.The algorithm aims to improve the robustness and accuracy of object detection in challenging environments by fusing features from vision and millimeter wave(mmWave)radar.It utilized a dual-branch one-stage architecture,with the image branch using a pre-trained CSPDarkNet53 backbone network to extract image features and the radar branch employing a voxel-based radar feature generation network to extract radar features.The proposed dual-scale CBAM feature fusion module integrated radar and visual features before and after the neck network.Finally,a decoupled detection head was deployed to classify and locate objects.The effectiveness and superiority of the proposed fusion detection algorithm were validated by comparative and ablation experiments conducted on the nuScenes dataset in challenging environments.

关 键 词：深度学习 目标检测 毫米波雷达 特征融合 多模态 注意力机制 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]