基于多尺度感知的密集人群计数网络

李恒超; 刘香莲; 刘鹏; 冯斌

doi:10.3969/j.issn.0258-2724.20220823

基于多尺度感知的密集人群计数网络

doi: 10.3969/j.issn.0258-2724.20220823

李恒超^{1, 2,},
刘香莲¹,
刘鹏¹,
冯斌^3, ,

1.
西南交通大学信息科学与技术学院，四川成都 611756
2.
西南交通大学综合交通大数据应用技术国家工程实验室，四川成都 611756
3.
西南交通大学体育学院，四川成都 611756

基金项目: 国家自然科学基金项目（62271418）；四川省自然科学基金项目（23NSFSC0058）

详细信息

作者简介:
李恒超（1978—），男，教授，博士，研究方向为智能遥感图像处理，E-mail：hcli@home.swjtu.edu.cn

通讯作者:
冯斌（1979—），男，讲师，硕士，研究方向为体育赛事组织管理，E-mail：feng197947@163.com

中图分类号: TP391.41
计量
- 文章访问数: 72
- HTML全文浏览量: 33
- PDF下载量: 10
- 被引次数: 0
出版历程
- 收稿日期: 2022-11-29
- 修回日期: 2023-03-08
- 网络出版日期: 2024-07-04

Dense Crowd Counting Network Based on Multi-Scale Perception

LI Hengchao^{1, 2
,},
LIU Xianglian¹,
LIU Peng¹,
FENG Bin^{3
, ,}

1.
School of Information Science and Technology, Southwest Jiaotong University, Chengdu 611756, China
2.
National Engineering Laboratory of Integrated Transportation Big Data Application Technology, Southwest Jiaotong University, Chengdu 611756, China
3.
School of Sports, Southwest Jiaotong University, Chengdu 611756, China

摘要

摘要:
针对密集人群场景存在的目标尺度多样、人群大规模变化等问题，提出一种基于多尺度感知的密集人群计数网络. 首先，考虑到小尺度目标在图像中占比较大，以VGG-16 （visual geometry group 2016）网络为基础，引入空洞卷积模块，以挖掘图像细节信息；其次，为充分利用目标多尺度信息，构建新的上下文感知模块，以提取不同尺度之间的对比特征；最后，考虑到目标尺度连续变化的特点，设计多尺度特征聚合模块，提高密集尺度采样范围与多尺度信息交互，从而提升模型性能. 实验结果显示：在ShangHai Tech （Part_A/Part_B）和UCF_CC_50数据集上，平均绝对误差（mean absolute error，MAE）分别降至62.5、6.9、156.5，均方根误差（root mean square error，RMSE）降至95.7、11.0、223.3；相较于对比模型最优方法，在UCF_QNRF数据集中MAE和RMSE分别降低1.1%和4.3%，NWPU数据集上分别降低8.7%和13.9%.
- 人群密度估计 /
- 多尺度聚合 /
- 空洞卷积 /
- 密度图
Abstract:
A dense crowd counting network based on multi-scale perception was proposed to solve the problems of diverse target scales and large-scale changes of crowds in dense crowd scenes. Firstly, since the small-scale targets account for a relatively large proportion of the images, a dilated convolution module was introduced based on the visual geometry group 2016 (VGG-16) network to mine the detailed information in the images. Then, by utilizing the multi-scale information of the target, a novel context-aware module was designed to extract the contrast features between different scales. Finally, In view of the continuous change of target scales, the multi-scale feature aggregation module was designed to improve the sampling range of dense scales, enhance the interaction of multi-scale information, and thus improve the model performance. The experimental results show that mean absolute error (MAE) decreases to 62.5, 6.9, and 156.5, and the root mean square error (RMSE) decreases to 95.7, 11.0, and 223.3 on ShangHai Tech (Part_A/ Part_B) and UCF_CC_50 datasets, respectively. Compared with the optimal method of comparison model, the MAE and RMSE are reduced by 1.1% and 4.3% on the UCF_QNRF dataset and by 8.7% and 13.9% on the NWPU dataset.
- crowd density estimation /
- multi-scale aggregation /
- dilated convolution /
- density map

HTML全文

图 1 基于多尺度感知的密集人群计数网络结构

Figure 1. Structure of dense crowd counting network based on multi-scale perception

下载: 全尺寸图片幻灯片

图 2 特征增强块结构

Figure 2. Structure of feature enhancement block

下载: 全尺寸图片幻灯片

图 3 部分可视化结果

Figure 3. Partial visualization results

下载: 全尺寸图片幻灯片

表 1 不同模型在Shanghai Tech、UCF_CC_50、UCF_QNRF、NWPU数据集上的对比结果

Table 1. Comparison results of different models on Shanghai Tech, UCF_CC_50, UCF_QNRF, and NWPU datasets

算法	Shanghai Tech Part_A		Shanghai Tech Part_B		UCF_ CC_50		UCF_ QNRF		NWPU
算法	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
MCNN^[7]	110.2	173.2	26.4	41.3	377.6	509.1	277.0	426.0	218.5	700.6
CSRNet^[10]	68.2	115.0	10.6	16.0	266.1	397.5	120.3	208.5	104.8	433.4
PDD-CNN^[29]	64.7	99.1	8.8	14.3	205.4	311.7	115.3	190.2
TEDNet^[18]	64.2	109.1	8.2	12.8	249.4	354.5	113.0	188.0
KDMG^[30]	63.8	99.2	7.8	12.7			99.5	173.0	100.5	415.5
BL^[31]	62.8	101.8	7.7	12.7	229.3	308.2	88.7	154.8	93.6	470.3
CAN^[20]	62.3	100.0	7.8	12.2	212.2	243.7	107.0	183.0	93.5	489.9
MCANet^[32]	60.1	100.2	6.8	11.0	181.3	258.6	100.8	185.9
SC2Net^[33]	58.9	97.7	6.9	11.4	209.4	286.3	98.5	174.5	89.7	348.9
MSPNet	62.5	95.7	6.9	11.0	156.5	223.3	87.7	148.2	81.9	300.3

下载: 导出CSV

表 2 CAM结构的消融实验

Table 2. Ablation experiment of CAM structure

方法	MAE	RMSE
本文 + PPM	63.6	105.4
本文 + CAM	62.5	95.7

下载: 导出CSV

表 3 模块结构的消融实验

Table 3. Ablation experiment of module structure

方法	MAE	RMSE
CAM	68.2	118.8
DCM	66.2	113.0
DCM+CAM	64.9	109.8
CAM+MSAM	65.5	111.4
DCM+MSAM	64.0	111.5
DCM+CAM+MSAM	62.5	95.7

下载: 导出CSV

表 4 FEB层选择消融实验

Table 4. Ablation experiment of number selection for FEB

FEB 层数/层	MAE	RMSE
0	64.9	109.8
2	63.5	103.4
4	62.5	95.7
6	67.1	115.8

下载: 导出CSV

参考文献(33)

[1]	FAN Z Z, ZHANG H, ZHANG Z, et al. A survey of crowd counting and density estimation based on convolutional neural network[J]. Neurocomputing, 2022, 472: 224-251. doi: 10.1016/j.neucom.2021.02.103
[2]	余鹰,朱慧琳,钱进,等. 基于深度学习的人群计数研究综述[J]. 计算机研究与发展,2021,58(12): 2724-2747. doi: 10.7544/issn1000-1239.2021.20200699 YU Ying, ZHU Huilin, QIAN Jin, et al. Survey on deep learning based crowd counting[J]. Journal of Computer Research and Development, 2021, 58(12): 2724-2747. doi: 10.7544/issn1000-1239.2021.20200699
[3]	TOPKAYA I S, ERDOGAN H, PORIKLI F. Counting people by clustering person detector outputs[C]//2014 11th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). Seoul: IEEE, 2014: 313-318.
[4]	RYAN D, DENMAN S, SRIDHARAN S, et al. An evaluation of crowd counting methods, features and regression models[J]. Computer Vision and Image Understanding, 2015, 130: 1-17. doi: 10.1016/j.cviu.2014.07.008
[5]	LEMPITSKY V, ZISSERMAN A. Learning to count objects in images[C]//Proceedings of the 23rd International Conference on Neural Information Processing Systems. New York: ACM, 2010: 1324-1332.
[6]	WANG C, ZHANG H, YANG L, et al. Deep people counting in extremely dense crowds[C]//Proceedings of the 23rd ACM international conference on Multimedia. Brisbane: ACM, 2015: 1299-1302.
[7]	ZHANG Y Y, ZHOU D S, CHEN S Q, et al. Single-image crowd counting via multi-column convolutional neural network[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas: IEEE, 2016: 589-597.
[8]	SINDAGI V A, PATEL V M. Generating high-quality crowd density maps using contextual pyramid CNNs[C]//2017 IEEE International Conference on Computer Vision (ICCV). Venice: IEEE, 2017: 1879-1888.
[9]	CAO X K, WANG Z P, ZHAO Y Y, et al. Scale aggregation network for accurate and efficient crowd counting[C]//European Conference on Computer Vision. Cham: Springer, 2018: 757-773.
[10]	LI Y H, ZHANG X F, CHEN D M. CSRNet: dilated convolutional neural networks for understanding the highly congested scenes[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 1091-1100.
[11]	SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]//3rd International Conference on Learning Representations. San Diego: [s.n.], 2015: 1-14.
[12]	YU F, KOLTUN V. Multi-scale context aggregation by dilated convolutions[C]//Proceedings of the 4th International Conference on Learning Representations. Puerto Rico: ICLR, 2016: 1-13.
[13]	田晟,张剑锋,张裕天,等. 基于扩张卷积金字塔网络的车道线检测算法[J]. 西南交通大学学报,2020,55(2): 386-392,416. doi: 10.3969/j.issn.0258-2724.20181026 TIAN Sheng, ZHANG Jianfeng, ZHANG Yutian, et al. Lane detection algorithm based on dilated convolution pyramid network[J]. Journal of Southwest Jiaotong University, 2020, 55(2): 386-392,416. doi: 10.3969/j.issn.0258-2724.20181026
[14]	左静,巴玉林. 基于多尺度融合的深度人群计数算法[J]. 激光与光电子学进展,2020,57(24): 315-323. ZUO Jing, BA Yulin. Population-depth counting algorithm based on multiscale fusion[J]. Laser & Optoelectronics Progress, 2020, 57(24): 315-323.
[15]	CHEN X Y, BIN Y R, SANG N, et al. Scale pyramid network for crowd counting[C]//2019 IEEE Winter Conference on Applications of Computer Vision (WACV). Waikoloa: IEEE, 2019: 1941-1950.
[16]	SINDAGI V, PATEL V. Multi-level bottom-top and top-bottom feature fusion for crowd counting[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul: IEEE, 2019: 1002-1012.
[17]	ZHOU Y, YANG J X, LI H R, et al. Adversarial learning for multiscale crowd counting under complex scenes[J]. IEEE Transactions on Cybernetics, 2021, 51(11): 5423-5432. doi: 10.1109/TCYB.2019.2956091
[18]	JIANG X L, XIAO Z H, ZHANG B C, et al. Crowd counting and density estimation by trellis encoder-decoder networks[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach: IEEE, 2019: 6126-6135.
[19]	TIAN Y K, LEI Y M, ZHANG J P, et al. PaDNet: pan-density crowd counting[J]. IEEE Transactions on Image Processing, 2020, 29: 2714-2727. doi: 10.1109/TIP.2019.2952083
[20]	LIU W Z, SALZMANN M, FUA P. Context-aware crowd counting[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach: IEEE, 2019: 5094-5103.
[21]	LI H, KONG W H, ZHANG S H. Effective crowd counting using multi-resolution context and image quality assessment-guided training[J]. Computer Vision and Image Understanding,2020,201:103065.1-103065.10.
[22]	IDREES H, SALEEMI I, SEIBERT C, et al. Multi-source multi-scale counting in extremely dense crowd images[C]//2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland: IEEE, 2013: 2547-2554.
[23]	IDREES H, TAYYAB M, ATHREY K, et al. Composition loss for counting, density map estimation and localization in dense crowds[C]//European Conference on Computer Vision. Cham: Springer, 2018: 544-559.
[24]	WANG Q, GAO J Y, LIN W, et al. NWPU-crowd: a large-scale benchmark for crowd counting and localization[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(6): 2141-2149. doi: 10.1109/TPAMI.2020.3013269
[25]	ZHAO H S, SHI J P, QI X J, et al. Pyramid scene parsing network[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu: IEEE, 2017: 6230-6239.
[26]	ZHANG Z L, ZHANG X Y, PENG C, et al. ExFuse: enhancing feature fusion for semantic segmentation[C]//European Conference on Computer Vision. Cham: Springer, 2018: 273-288.
[27]	王银,王立德,邱霁. 基于DenseNet结构的轨道暗光环境实时增强算法[J]. 西南交通大学学报,2022,57(6): 1349-1357. doi: 10.3969/j.issn.0258-2724.20210199 WANG Yin, WANG Lide, QIU Ji. Real-time enhancement algorithm based on DenseNet structure for railroad low-light environment[J]. Journal of Southwest Jiaotong University, 2022, 57(6): 1349-1357. doi: 10.3969/j.issn.0258-2724.20210199
[28]	DAI F, LIU H, MA Y K, et al. Dense scale network for crowd counting[C]//Proceedings of the 2021 International Conference on Multimedia Retrieval. Taipei: ACM, 2021: 64-72.
[29]	WANG W X, LIU Q L, WANG W. Pyramid-dilated deep convolutional neural network for crowd counting[J]. Applied Intelligence, 2022, 52(2): 1825-1837. doi: 10.1007/s10489-021-02537-6
[30]	WAN J, WANG Q Z, CHAN A B. Kernel-based density map generation for dense object counting[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(3): 1357-1370. doi: 10.1109/TPAMI.2020.3022878
[31]	MA Z H, WEI X, HONG X P, et al. Bayesian loss for crowd count estimation with point supervision[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul: IEEE, 2019: 6141-6150.
[32]	WANG X, LV R R, ZHAO Y, et al. Multi-scale context aggregation network with attention-guided for crowd counting[C]//2020 15th IEEE International Conference on Signal Processing (ICSP). Beijing: IEEE, 2020: 240-245.
[33]	LIANG L J, ZHAO H L, ZHOU F B, et al. SC2Net: scale-aware crowd counting network with pyramid dilated convolution[J]. Applied Intelligence, 2023, 53(5): 5146-5159.