检索规则说明:AND代表“并且”;OR代表“或者”;NOT代表“不包含”;(注意必须大写,运算符两边需空一格)
检 索 范 例 :范例一: (K=图书馆学 OR K=情报学) AND A=范并思 范例二:J=计算机应用与软件 AND (U=C++ OR U=Basic) NOT M=Visual
作 者:张千龙 曾岳[1] ZHANG Qianlong;ZENG Yue(Jinling Institute of Technology,Nanjing 211169,China)
机构地区:[1]金陵科技学院软件工程学院,江苏南京211169
出 处:《金陵科技学院学报》2025年第1期8-14,共7页Journal of Jinling Institute of Technology
基 金:国家自然科学基金青年项目(42401454)。
摘 要:针对密集遮挡场景下行人检测的实时性需求与特征退化问题,提出了RG-YOLO(reinforced group you only look once)轻量化检测框架,通过三重创新实现精度与效率的协同优化:1)设计GSConv异构卷积核组,通过跨组跳连机制在保持85%通道交互能力的前提下降低了71.2%的计算负载。2)构建VoV-GSCSP动态分组模块,依据特征图尺度自适应调整分组策略,提升遮挡目标18.7%的特征区分度。3)开发级联式CBAM-SPPF注意力机制,通过双路径融合增强遮挡区域的上下文建模能力在WiderPerson和CrowdHuman数据集上的对比实验结果表明:在输入像素分辨率为640×640的情况下,本模型的mAP@0.5值达到83.6%,较YOLOv5s提升了9.3%;同时将计算量压缩至2.8 GFLOPs,在Jetson Nano嵌入式平台实现了27.3 fps的实时检测。消融实验结果表明:GSConv在密集遮挡子集上使误检率降低了21.4%。To address the real-time requirements and feature degradation issues in pedestrian detection in dense occlusion scenarios,we propose RG-YOLO(reinforced group you only look once),a lightweight detection framework that achieves synergistic optimization of accuracy and efficiency through three innovations:1)Designing GSConv heterogeneous convolutional kernel groups,which reduce computational load by 71.2%while maintaining 85%channel interaction capability through cross-group skip connections;2)Constructing VoV-GSCSP dynamic grouping modules,which adaptively adjust grouping strategies based on feature map scales,thereby enhancing the feature discriminability of occluded targets by 18.7%;3)Developing a cascaded CBAM-SPPF attention mechanism,which strengthens contextual modeling capabilities in occluded regions through dual-path fusion.Comparative experiments on the WiderPerson and CrowdHuman datasets demonstrate that,with an input resolution of 640×640,the mAP@0.5 value of proposed model achieves 83.6%,outperforming YOLOv5s by 9.3%.Simultaneously,the computational complexity is reduced to 2.8 GFLOPs,achieving real-time detection at 27.3 fps on Jetson Nano embedded platform.Ablation experiments further show that GSConv reduces the false detection rate by 21.4%on dense occlusion subsets.
关 键 词:行人检测 CA注意力模块 GSConv RG-YOLO 高密度遮挡
分 类 号:TP391[自动化与计算机技术—计算机应用技术]
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在链接到云南高校图书馆文献保障联盟下载...
云南高校图书馆联盟文献共享服务平台 版权所有©
您的IP:216.73.216.33