School of Railway Intelligent Engineering, Dalian Jiaotong University, Dalian 116052, China
Show less
文章历史+
Received
Published
Revised
2025-06-10
2026-05-18
2025-11-19
Online Date
2026-08-06
PDF (12904K)
摘要
为了解决复杂场景下密集行人检测精度低的问题,通过改进RT-DETR算法设计了密集行人检测算法CHD-DETR(crowd-human-detect-DETR)。首先,设计了一个部分分离PSNET(pyramid scene parsing network)模块改进主干网络,提升网络对小目标特征的提取能力,降低模型的计算量。其次,设计了一个全新的超图聚合HAM(hybrid aggregation module)模块,捕捉高阶和低阶特征关系,结合该模块和金字塔空间池化的思想,设计出平均聚合AHNET(average hybrid network)网络作为多尺度特征融合部分,减少细节信息丢失;最后,设计了一个卷积高效层聚合CELA-Block(convolution ELAN block)模块替换RepC3模块,增强模型对局部特征的提取,提高模型的检测精度。CHD-DETR模型在INRIA+Kaist数据集和CrowdHuman数据集上进行了消融实验,并且与主流算法进行了对比实验。相比RT-DETR基线模型,CHD-DETR模型在INRIA+Kaist数据集上mAP50、mAP50:95和查准率提高了1.7%、3.3%和5.5%,在CrowdHuman数据集上mAP50、mAP50:95和召回率提高了1.0%、1.1%和1.4%,计算量下降了10.3 GFLOPs。实验结果表明,所提出的CHD-DETR模型在复杂场景密集行人检测中具有较好的性能。
Abstract
In order to solve the problem of low accuracy of dense pedestrian detection in complex scenarios, the dense pedestrian detection algorithm CHD-DETR (crowd-human-detect-DETR) was developed by improving the RT-DETR algorithm. Firstly, a partially separated PSNET(pyramid scene parsing network) module was designed to improve the backbone network. This improves the network’s ability to extract small target features, and reduce the computational cost of the model. Secondly, a novel hypergraph aggregation module named HAM (hybrid aggregation module) was designed to capture the high-order and low-order feature relationships. By integrating this module with the concept of pyramid spatial pooling, an average hybrid network named AHNET(average hybrid network) was designed as the multi-scale feature fusion, effectively reducing the loss of detailed information. Finally, a convolutional high-efficiency layer aggregation CELA-Block (convolution ELAN block) module was designed to replace the RepC3 module, enhancing the extraction of local features and improving detection accuracy. The CHD-DETR model was evaluated through ablation experiments on the INRIA+Kaist and CrowdHuman datasets, and comparative experiments were conducted with mainstream algorithms. Compared with the RT-DETR baseline model, CHD-DETR achieved improvements of 1.7%, 3.3%, and 5.5% in mAP50, mAP50:95, and precision, respectively, on the INRIA+Kaist dataset. On the CrowdHuman dataset, it achieved increases of 1.0%, 1.1%, and 1.4% in mAP50, mAP50:95, and recall, respectively, while reducing computational cost by 10.3 GFLOPs. Experimental results show that the proposed CHD-DETR model exhibits good performance in dense pedestrian detection under complex scenarios.
WangHong, HanChen, YuanBoyang, et al. A crowded pedestrian detection algorithm based on improved YOLOv5[J]. Science Technology and Engineering, 2023, 23(27): 11730-11738.
LiuChengxiao, HeTao, JingJiabao. A partially occluded pedestrian detection algorithm based on binocular vision[J]. Science Technology and Engineering, 2024, 24(13): 5465-5472.
[7]
ChuX, ZhengA, ZhangX, et al. Detection in crowded scenes: one proposal, multiple predictions[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York:IEEE, 2020: 12214-12223.
YinZeyu, YangBo, ChenJinling, et al. A lightweight small object detection algorithm based on STD-DETR[J]. Laser & Optoelectronics Progress, 2025, 62(8): 146-156.
SongTianze, CaoCongjun, HeJiaqi, et al. Research on dense pedestrian detection algorithm based on improved DETR[J]. Computer Engineering, 2026. DOI: 10.19678/j.issn.1000-3428.0070106.
FanRong, MaXiaolu. An improved detr algorithm for crowded pedestrian detection[J]. Computer Engineering and Applications, 2023, 59(19): 159-165.
[18]
PengH, ChenS. FedsNet: the real-time network for pedestrian detection based on RT-DETR[J]. Journal of Real-Time Image Processing, 2024, 142 (21). DOI: 10.1007/s11554-024-01523-8.
[19]
ZhaoY, LüW, XuS, et al. Detrs beat YOLOs on real-time object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York:IEEE, 2024: 16965-16974.
[20]
LiuM, DanJ, LuZ, et al. CM-UNet: hybrid CNN-Mamba UNet for remote sensing image semantic segmentation[J]. arXiv Preprint arXiv:2405.10530, 2024.
[21]
HanK, WangY, TianQ, et al. Ghostnet: more features from cheap operations[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York:IEEE, 2020: 1580-1589.
[22]
BaiS, ZhangF, TorrP H S. Hypergraph convolution and hypergraph attention[J]. Pattern Recognition, 2024, 110: 107637.
[23]
FengY, HuangJ G, DuS, et al. Hyper-YOLO: when visual object detection meets hypergraph computation[C]//IEEE Transactions on Pattern Analysis and Machine Intelligence. Online:IEEE, 2025: 2388-2401.
[24]
ZhaoH, ShiJ, QiX, et al. Pyramid scene parsing network[C]//Proceedings of the IEEE Conference On Computer Vision And Pattern Recognition. New York: IEEE, 2017: 2881-2890.
[25]
ZhangX, ZengH, GuoS, et al. Efficent longe-range attention network for image super-resolution[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 649-667.
[26]
WanC, YuH, LiZ, et al. Swift parameter-free attention network for efficient super-resolution[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2024: 6246-6256.
[27]
ZhuX, HuH, LinS, et al. Deformable convnets V2: More deformable, better results[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York:IEEE, 2019: 9308-9316.
[28]
ShaoS, ZhaoZ, LiB, et al. Crowdhuman: a benchmark for detecting human in a crowd[J]. arXiv Preprint arXiv: 1805.00123, 2018.
[29]
HwangS, ParkJ, KimN, et al. Multispectral pedestrian detection: benchmark dataset and baseline[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. New York:IEEE, 2015: 1037-1045.
WangWeifeng, ShaoLinqian, HuangJianxin, et al. A dual-modal fusion pedestrian detection method based on improved YOLOv5[J]. Journal of Jilin University (Engineering and Technology Edition), 2025. DOI: 10.13229/j.cnki.jdxbgxb.20241288.
[32]
MehtaS, RastegariM. MobileViT: light-weight, general-purpose, and mobile-friendly vision transformer[C]//International Conference on Learning Representations (ICLR). Online: ICLR, 2022: 2178.
JiaoTianwen, TianXiuyun. Research on pedestrian detection system based on YOLOv5[J]. Modern Information Technology, 2025, 9(2): 33-38.
[37]
LuoS, LiH, LiuJ. Research on Pedestrian Detection Based on Improved SVDD Algorithm[C]//IEEE 6th Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC). New York:IEEE, 2024: 126-130.
[38]
HaoS, LiJ, SunX, et al. MDFOaNet: a novel multi-modal pedestrian detection network based on multi-scale image dynamic feature optimization and attention mapping[C]//IEEE Transactions on Intelligent Transportation Systems. New York: IEEE, 2025: 268-2822.
WangLili, FanPanpan, ZhangShiyu. Improved YOLOv8n algorithm for dense pedestrian scenes[J]. Application of Electronic Technique, 2025, 51(2): 15-20.
[41]
SongY, QianP, ZhangK, et al. An improved multi-scale fusion and small object enhancement method for efficient pedestrian detection in dense scenes[J]. Multimedia Systems, 2025, 31(2): 1-16.
ZhouYaluo, HanNing, YinBiaobiao, et al. Research on MCAP-YOLO and Bytetrack pedestrian tracking algorithms for dense scenes[J]. Mini-Micro Computer Systems, 2025. DOI: 10.20009/j.cnki.21-1106/TP.2025-0151.