摘要
夜间低光照环境下的行人检测面临误检率高、漏检显著及识别精度不足等挑战。为此,本文提出一种基于改进RT-DETR的检测算法,通过多模块协同设计实现低光照下的精准检测。算法在特征金字塔顶层嵌入空域变换器(FDT)模块,采用双阶段注意力机制增强微弱特征提取与全局上下文建模能力;在颈部网络部署动态采样(DySample)模块,通过动态可学习的空间重采样机制实现多尺度特征对齐与小目标检测增强;并以DRBC3模块作为特征提取核心,融合多膨胀率卷积与重参数化技术,构建多尺度感受野以提升对模糊与遮挡目标的细节捕捉能力。在LLVIP数据集上的实验表明,本算法在参数量下降的同时,mAP0.5、准确率(Precision)和召回率(Recall)分别提升1.39%、2.21%和3%,推理速度也显著提高。NightSurveillance和Nightowls数据集上的泛化实验进一步验证了其优越性能。本文算法在保障实时性的前提下,有效提高了检测精度并降低了漏检率,具备良好的鲁棒性与实用性。
Pedestrian detection in low-light night-time environments faces challenges including high false positive rates,significant false negatives,and insufficient recognition accuracy.To address this,this paper proposes a detection algorithm based on an improved RT-DETR,achieving precise detection under low illumination through multi-module collaborative design.The algorithm embeds an FDT module at the top layer of the feature pyramid,employing a two-stage attention mechanism to enhance weak feature extraction and global context modelling capabilities.A DySample module is deployed in the neck network,employing a dynamically learnable spatial resampling mechanism to achieve multi-scale feature alignment and small object detection enhancement.Furthermore,the DRBC3 module serves as the feature extraction core,integrating multi-expansion-rate convolutions and re-parameterisation techniques to construct multi-scale receptive fields,thereby enhancing the capture of details in blurred and occluded objects.Experiments on the LLVIP dataset demonstrate that this algorithm achieves a 1.39%increase in mAP0.5,a 2.21%rise in Precision,and a 3%improvement in Recall,while simultaneously reducing the number of parameters.Inference speed is also significantly enhanced.Generalisation experiments on the NightSurveillance and Nightowls datasets further validate its superior performance.The algorithm effectively improves detection accuracy and reduces false negatives while maintaining real-time capability,exhibiting robust and practical performance.
作者
陆燕
李富
祁铭瑞
杨心梦
LU Yan;LI Fu;QI Mingrui;YANG Xinmeng(School of Computer,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Internet of Things Engineering,Wuxi University,Wuxi 214105,China;Key Laboratory of Embedded System and Service Computing Ministry of Education,Tongji University,Shanghai 201804,China)
出处
《液晶与显示》
北大核心
2025年第12期1881-1893,共13页
Chinese Journal of Liquid Crystals and Displays
基金
国家自然科学基金(No.42305158)
江苏省产学研合作资助项目(No.BY20240827)。
