基于RT-DETR的轻量化车辆目标检测算法OA
Lightweight vehicle object detection algorithm based on RT-DETR
针对自动驾驶场景的硬件限制以及多尺度和遮挡现象导致检测性能不佳的问题,本文提出了一种用于车辆检测任务的轻量级目标检测算法RT-DETR-light.首先,提出利用CG Block模块改进骨干网络卷积模块,并基于此构建了轻量级特征提取网络CGResNet,实现了推理速度与检测精度的平衡.在特征融合阶段,引入双向特征金字塔网络BiFPN,通过双向信息传递实现精度的提升.最后,针对车辆目标检测任务中小目标与遮挡场景下定位精度不足的问题,设计一个改进的损失函数EPGIoU,通过多约束协同设计优化极端场景梯度稳定性.实验结果表明,本文算法在UA-DETRAC数据集上的mAP@0.5与精确率分别达到了75.0%、74.5%,相较于基线算法,参数量与计算量降低了26.4%与18.0%,检测速度提升了1.4个百分点.在BDD100K-Sub数据集上的跨数据集评估进一步验证了其泛化能力.本文提出的检测算法在检测精度、轻量化与推理速度上取得了显著优势,具备良好的泛化能力,为自动驾驶场景中实时车辆检测与边缘设备部署提供了更优的解决方案.
To tackle degraded vehicle detection performance caused by hardware constraints,multi-scale objects,and occlusions in autonomous driving,this paper proposes RT-DETR-light,a lightweight detection algorithm.First,we design a CG Block to enhance the backbone network,forming the lightweight feature extractor CGResNet,which balances speed and accuracy.A bidirectional feature pyramid network BiFPN is then introduced for feature fusion to improve precision via bidirectional information flow.Furthermore,an enhanced loss function,EPGIoU,is proposed to improve localization accuracy for small and occluded vehicles by stabilizing gradient optimization via multi-constraint collaboration.Experiments on the UA-DETRAC dataset show a mAP@0.5 of 75.0%and a precision of 74.5%.Compared to the baseline,it reduces parameters and computation by 26.4%and 18.0%,respectively,while improving detection speed by 1.4 percentage points.Cross-dataset evaluation on BDD100K-Sub confirms its strong generalization ability.The proposed algorithm offers superior accuracy,lightweight design,and inference speed,providing an effective solution for real-time vehicle detection and edge device deployment.
张子轶;马丽;吕帅;朱中宁
南京信息工程大学 计算机学院,江苏 南京 210044||无锡学院 物联网工程学院,江苏 无锡 214105无锡学院 物联网工程学院,江苏 无锡 214105无锡学院 物联网工程学院,江苏 无锡 214105无锡学院 物联网工程学院,江苏 无锡 214105
信息技术与安全科学
深度学习RT-DETR算法轻量化车辆目标检测
deep learningRT-DETR algorithmlightweightvehicle object detection
《液晶与显示》 2026 (3)
373-387,15
评论