基于多粒度关系推理的自动驾驶域自适应视觉目标检测算法  

Domain Adaptive Visual Object Detection for Autonomous Driving Based on Multi-granularity Relation Reasoning

作　　者：索锦辉 王晓伟[1,2] 蒋沛文 丁驰 高铭 边有钢 Suo Jinhui;Wang Xiaowei;Jiang Peiwen;Ding Chi;Gao Ming;Bian Yougang(Hunan University,State Key Laboratory of Advanced Design and Manufacturing Technology for Vehicle,Changsha 410082;Wuxi Intelligent Control Research Institute of Hunan University,Wuxi 214072;School of Vehicle and Mobility,Tsinghua University,Beijing 100084)

机构地区：[1]湖南大学,整车先进设计制造技术全国重点实验室,长沙410082 [2]湖南大学无锡智能控制研究院,无锡214072 [3]清华大学车辆与运载工程学院,北京100084

出　　处：《汽车工程》2025年第2期201-210,共10页Automotive Engineering

基　　金：国家重点研发计划项目(2022YFB2503402);长三角科技创新共同体联合攻关计划项目(2023CSJGG0801);国家自然科学基金青年科学基金项目(52102461);湖南省青年科技创新人才项目(2022RC1033);湖南省自然科学基金项目(2023JJ40155)资助。

摘　　要：现有域自适应视觉目标检测算法大多基于两阶段检测器设计,且未能利用图像空间中不同元素之间的语义拓扑关系,导致次优的跨域适应性能。为此,本文提出一种基于多粒度关系推理的域自适应视觉目标检测算法。首先,提出粗粒度图块关系推理模块,使用粗粒度图块图结构来捕获前景和背景之间的拓扑关系,对前景区域进行跨域适配。然后,设计细粒度语义关系推理模块,推理细粒度语义图结构来增强跨域多类别语义依赖关系。最后,提出粒度诱导的特征对齐模块,根据节点的亲和性调节特征对齐的权重,提升检测模型面对场景整体变化时的适应性。多个自动驾驶跨域场景上的实验结果验证了所提算法的鲁棒性和实时性。Most of the existing domain adaptive visual object detection algorithms are based on two-stage detector design and fail to exploit the semantic topological relationship between different elements in the image space,resulting in suboptimal cross-domain adaptation performance.Therefore,in this paper a domain adaptive vi-sual object detection algorithm based on multi-granularity relationship reasoning is proposed.Firstly,a coarse-grained patch relationship reasoning module is proposed,which uses the coarse-grained patch graph structure to capture the topological relationship between the foreground and background and perform cross-domain adaptation on the foreground area.Then,a fine-grained semantic relationship reasoning module is designed to reason about the fine-grained semantic graph structure to enhance cross-domain multi-category semantic dependencies.Finally,a granularity-induced feature alignment module is proposed to adjust the weight of feature alignment according to the affinity of the nodes,thereby improving the adaptability of the detection model when facing overall scene changes.The experimental results on multiple cross-domain scenarios of autonomous driving verify the robustness and real-time performance of the proposed algorithm.

关 键 词：自动驾驶 视觉目标检测 域自适应 图推理 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U463.6[自动化与计算机技术—计算机科学与技术]