华中科技大学学报（自然科学版）2025，Vol.53Issue(11)：15-20,27,7.DOI:10.13245/j.hust.251158
结合深度梯度与对齐融合策略的目标检测方法
Target detection method combining depth gradient and alignment fusion strategy
摘要
Abstract
To address inaccurate depth estimation and feature misalignment caused by depth-mutation in autonomous driving,a target detection method integrating depth gradients and feature alignment was proposed.First,the depth gradient sensing-lift splat shoot(DGS-LSS)module utilized LiDAR gradient information to construct geometric representations and optimized pixel-level depth estimation,improving the fitting precision in depth-mutation scenarios.Then,the Align-Fusion module was proposed to achieve dynamic bird's-eye view(BEV)feature alignment via learnable flow fields,enhancing the effect of information fusion.Research results show that on the nuScenes dataset,the method achieves mean average precision(mAP)of 71.2%and nuScenes detection score(NDS)of 73.1%,outperforming BEVFusion by 3.3%and 2.1%,respectively,demonstrating its effectiveness.关键词
鸟瞰图/深度估计/3D目标检测/多模态融合/特征融合Key words
bird's-eye view/depth estimation/3D target detection/multimodal fusion/feature fusion分类
信息技术与安全科学引用本文复制引用
PENG Gang,LIU Xingyu,LIU Jie,GAO Qiang..结合深度梯度与对齐融合策略的目标检测方法[J].华中科技大学学报（自然科学版）,2025,53(11):15-20,27,7.基金项目
湖北省农机装备补短板核心技术应用攻关项目(HBSNYT202213) （HBSNYT202213）
湖北省自然科学基金资助项目(2019CFB526) （2019CFB526）
湖北省揭榜制重大科技项目(2021BEC008) （2021BEC008）
湖北省科技重大专项(2020ACA006). （2020ACA006）
