双分流多尺度注意力与单层输出的密集行人检测算法

doi:10.13229/j.cnki.jdxbgxb.20250284

摘要/Abstract

摘要：

针对密集行人检测因遮挡和小尺度目标导致的漏检与误检问题，提出了一种基于双分流多尺度注意力与单层输出的密集行人检测算法DSS R-CNN。首先，设计了一种双分流多尺度注意力模块（DBMAB），将其嵌入CloFormer主干网络中以增强特征提取能力，使模型更关注重要信息，从而提升检测精度；其次，针对特征金字塔网络多层输出对Sparse R-CNN贡献有限的问题，提出了一种基于特征调制模块（FMB）的单层输出颈（SLON），在保持检测精度的同时提高了推理速度；最后，采用Focal-EIoU损失函数进一步优化模型训练。在CrowdHuman数据集上的实验结果表明，本文算法达到了90.4%的AP、96.2%的Recall和44.1%的MR-2，检测速度保持在46 f/s，优于现有密集行人场景下的检测算法。

关键词: 密集行人检测, Sparse R-CNN, CloFormer, 多尺度注意力, 单层输出颈

Abstract:

Aiming at the problem of missed detection and false detection caused by occlusion and small-scale targets in dense pedestrian detection， the paper proposes a dense pedestrian detection algorithm DSS R-CNN （Dual-branch Single-layer Sparse R-CNN） based on dual-branch multi-scale attention and single-layer output. Firstly， a Dual-Branch Multi-scale Attention Block （DBMAB） is designed and embedded into the CloFormer backbone network to enhance the feature extraction ability， so that the model can pay more attention to important information， so as to improve the detection accuracy. Secondly， aiming at the problem that the multi-layer output of feature pyramid network has limited contribution to Sparse R-CNN， a Single-layer Output Neck （SLON） based on Feature Modulation Block （FMB） is proposed. It improves the inference speed while maintaining the detection accuracy. Finally， Focal-EIoU loss function is used to further optimize the model training. Experimental results on the CrowdHuman dataset show that the proposed algorithm achieves 90.4% AP， 96.2% Recall and 44.1% MR-2， and the detection speed remains at 46 f/s， which is better than the existing pedestrian detection algorithms in complex scenes.

Key words: dense pedestrian detection, Sparse R-CNN, CloFormer, multi-scale attention, single-layer output neck

中图分类号:

TP391

金静,韩子阳,翟凤文. 双分流多尺度注意力与单层输出的密集行人检测算法[J]. 吉林大学学报(工学版), 2026, 56(9): 2476-2488.

Jing JIN,Zi-yang HAN,Feng-wen ZHAI. Dense pedestrian detection algorithm with dual-branch multi-scale attention and single-layer output[J]. Journal of Jilin University(Engineering and Technology Edition), 2026, 56(9): 2476-2488.

图/表 14

图1

图2

图3

图4

图5

图6

图7

表1

表2

表3

表4

图8

表5

图9

参考文献 33

[1]	齐向明, 刘晓暄, 王子键. 关键特征感知并行细粒度特征提取的密集行人检测[J]. 计算机工程与应用, 2026, 62(1): 297-306.
	Qi Xiang-ming, Liu Xiao-xuan, Wang Zi-jian. Dense pedestrian detection based on key feature perception and parallel fine-grained feature extraction[J]. Computer Engineering and Applications, 2026, 62(1): 297-306.
[2]	Ren S Q, He K M, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[3]	Cai Z W, Vasconcelos N. Cascade R-CNN: Delving into high quality object detection[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 6154-6162.
[4]	He K, Gkioxari G, Dollár P, et al. Mask R-CNN[C]∥Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Venice, Italy,2017: 2961-2969.
[5]	Sun P Z, Zhang R, Jiang Y F, et al. Sparse r-CNN: end-to-end object detection with learnable proposals[C]∥Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Nashville, USA, 2021, 14449-14458.
[6]	Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas,USA, 2016: 779-788.
[7]	Redmon J, Farhadi A. YOLO9000: better, faster, stronger[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 6517-6525.
[8]	Redmon J, Farhadi A. YOLOv3: an incremental improvement [EB/OL]. [2018-04-08]. .
[9]	Bochkovskiy A, Wang C Y, Liao H Y M. YOLOv4: optimal speed and accuracy of object detection[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020, 11:829-838.
[10]	Zhu X K, Lyu S C, Wang X, et al. TPH-YOLOv5: improved YOLOv5 based on transformer prediction head for object detection on drone-captured scenarios[C]∥2021 IEEE/CVF International Conference on Computer Vision Workshops, Montreal, Canada, 2021:2778-2788.
[11]	Wang C Y, Bochkovskiy A, Liao H Y M. YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]∥Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Vancouver, Canada, 2023: 7464-7475.
[12]	Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]∥Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands,2016,9905: 21-37.
[13]	Alexey D. An image is worth 16x16 words: Transformers for image recognition at scale[J]. Arxiv Preprint arxiv: , 2020.
[14]	Liu Z, Lin Y T, Cao Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]∥Proceedings of the IEEE/CVF International Conference on Computer Vision, Montreal, Canada, 2021:9992-10002.
[15]	Li Q M, Su Y J, Gao Y, et al. OAF-Net: an occlusion-aware anchor-free network for pedestrian detection in a crowd[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(11): 21291-21300.
[16]	Wu J T, Dai G J, Zhou W H, et al. Multi-scale feature fusion with attention mechanism for crowded road object detection[J]. Journal of Real-Time Image Processing, 2024, 21(2): 29.
[17]	柳长源, 张玉亮, 毕晓君. 基于多阶段提议稀疏区域卷积网络的城市交通目标检测[J]. 电子学报, 2023, 51(1): 26-31.
	Liu Chang-yuan, Zhang Yu-liang, Bi Xiao-jun. Urban traffic object detection based on multi-stage proposal sparse regional convolutional neural network[J]. Chinese Journal of Electronics, 2023, 51(1): 26-31.
[18]	Chen T, Chen J H, Gao T, et al. A novel scene-aware pedestrian detection in dense scenes[C]∥2024 27th International Conference on Computer Supported Cooperative Work in Desig, Tianjin, China, 2024: 1316-1321.
[19]	Park S K, Chung J H, Pae D S, et al. Fusion-attention network using dense scale-invariant feature transform flow image and point cloud for 3D pedestrian detection[J]. Multimed Tools and Applications,2025,84(14): 12813-12833.
[20]	Fan Q, Huang H B, Guan J Y, et al. Rethinking local perception in lightweight vision transformer[J]. arXiv, 2023. arXiv:.
[21]	Chollet F. Xception: deep learning with depthwise separable convolutions[C]∥Proceedings of the IEEE conference on computer vision and pattern recognition, Honolulu, USA, 2017:1800-1807.
[22]	Li J F, Wen Y, He L H. Scconv: spatial and channel reconstruction convolution for feature redundancy[C]∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, Canada,2023: 6153-6162.
[23]	Wang Y, Li Y S, Wang G, et al. Multi-scale attention network for single image super-resolution[C]∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2024: 5950-5960.
[24]	Zheng M J, Sun L, Dong J X, et al. SMFANet: A lightweight self-modulation feature aggregation network for efficient image super-resolution[C]∥European Conference on Computer Vision, Milan, Italy, 2024:359-375.
[25]	Zhang Y F, Ren W, Zhang Z, et al. Focal and efficient IOU loss for accurate bounding box regression[J]. Neurocomputing, 2022, 506: 146-157.
[26]	Shao S, Zhao Z J, Li B X, et al. Crowdhuman: a benchmark for detecting human in a crowd[J]. arXiv, 2018. arXiv:.
[27]	Tan M X, Pang R M, Le Q V. EfficientDet: scalable and efficient object detection[C]∥2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020:10778-10787.
[28]	He J B, Erfani S, Ma X J, et al. Alpha-IoU: a family of power intersection over union losses for bounding box regression[C]∥2021 Conference on Neural Information Processing Systems,San Diego,USA,2021: 20230-20242
[29]	Ge Z, Liu S T, Wang F, et al. YOLOX: exceeding YOLO series in 2021[J].arXiv, 2021. arXiv:.
[30]	Zhou X, Wang D, Krähenbühl P. Objects as points[J]. arXiv, 2019. arXiv:.
[31]	Varghese R, Sambath M. YOLOv8: a novel object detection algorithm with enhanced performance and robustness[C]∥2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems, Chennai, India, 2024: 1-6.
[32]	Khanam R, Hussain M. Yolov11: An overview of the key architectural enhancements[J].arXiv, 2024. arXiv:.
[33]	Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[C]∥Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 2017:2999-3007.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

算法	AP /%	MR-2 /%	Recall /%	帧率/ （f·s^-1）
Sparse RCNN+IoU	86.8	47.5	95.4	43
Sparse RCNN+GIoU loss	87.1	46.9	95.6	42
Sparse RCNN+DIoU loss	87.2	47.3	95.6	42
Sparse RCNN+CIoU loss（default）	87.1	47.0	95.6	43
Sparse RCNN+SIoU loss	87.3	46.5	95.6	42
Sparse RCNN+Alpha-IoU loss^［28］	87.4	46.3	95.7	43
Sparse RCNN+Focal-EIoU	87.6	46.1	95.8	43

算法	Image-size	AP /%	MR-2 /%	Recall /%	帧率/ （f·s^-1）
Faster R-CNN	640×640	72.4	51.2	88.5	29
Cascade R-CNN	640×640	74.6	44.2	87.2	36
YOLOX	640×640	86.5	46.3	95.9	79
CenterNet	640×640	87.0	45.7	96.1	48
YOLOv8	640×640	90.2	44.5	96.2	76
Sparse R-CNN（Baseline）	640×640	87.1	46.9	95.6	43
Swin-Transformer	640×640	90.1	44.8	95.9	51
文献［16］	640×640	87.4	45.4	95.6	44
YOLOv11	640×640	91.1	44.3	97.2	62
DSS R-CNN（Ours）	640×640	90.4	44.1	96.2	46

算法	AP	AP50	AP75	AP_l	AP_m	AP_s
RetinaNet^［33］	30.16	65.87	22.88	33.15	30.85	16.96
Faster R-CNN	38.78	73.12	35.53	40.79	38.71	29.05
Cascade R-CNN	42.64	77.63	42.69	41.56	41.97	31.13
Sparse R-CNN	45.37	80.19	45.07	48.43	43.85	32.86
DSS R-CNN（Ours）	47.46	81.38	47.26	49.62	44.93	36.21

Improved scheme				AP /%	MR-2 /%	Recall /%	帧率/ （f·s^-1）
Model	DBMAB-CloFormer	SLON	Focal-EIoU	AP /%	MR-2 /%	Recall /%	帧率/ （f·s^-1）
A	×	×	×	87.1	46.9	95.6	43
B	√	×	×	88.2	45.3	95.8	40
C	×	√	×	87.6	45.9	95.8	46
D	×	×	√	87.6	46.1	95.8	43
E	√	×	√	88.9	44.8	96.0	41
F	√	√	×	89.6	44.5	96.1	42
H	√	√	√	90.4	44.1	96.2	46