多通路高分辨率网络与注意力机制融合的车辆检测模型

doi:10.3969/j.issn.1674-8484.2022.01.012

汽车安全与节能学报 ›› 2022, Vol. 13 ›› Issue (1): 122-130.DOI: 10.3969/j.issn.1674-8484.2022.01.012

多通路高分辨率网络与注意力机制融合的车辆检测模型

丁飞¹^,²(), 米冠宇¹, 童恩³, 张楠¹, 暴建民¹, 张登银¹^,²

1.南京邮电大学江苏省宽带无线通信和物联网重点实验室,南京 210003,中国
2.南京邮电大学通信与网络技术国家工程研究中心,江苏南京 210003,中国
3.中国移动通信集团江苏有限公司,南京 210029,中国

收稿日期:2021-09-15 修回日期:2021-12-20 出版日期:2022-03-31 发布日期:2022-04-02
作者简介:丁飞（1981—）,男（汉族）,江苏,副教授。E-mail：dingfei@njupt.edu.cn。
基金资助:
国家自然科学基金(61872423);工业和信息化部产业技术基础公共服务平台项目(2019-00892-3-1);工业和信息化部通信软科学研究项目(2019-R-26);江苏省“六大人才高峰”高层次人才资助项目(DZXX-008)

Multi-channel high-resolution network and attention mechanism fusion for vehicle detection model

DING Fei¹^,²(), MI Guanyu¹, TONG En³, ZHANG Nan¹, BAO Jianmin¹, ZHANG Dengyin¹^,²

1. Jiangsu Key Laboratory of Broadband Wireless Communication and Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
2. National Local Joint Engineering Research Center for Communication and Network Technology, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
3. China Mobile Group Jiangsu Company Ltd., Nanjing 210029, China

Received:2021-09-15 Revised:2021-12-20 Online:2022-03-31 Published:2022-04-02

摘要/Abstract

摘要：

为改善道路交通监测和保证智能网联交通系统的安全、可靠与稳定运行,提出了在路侧边缘平台中基于多通路高分辨率网络与注意力机制融合的车辆检测模型MCHRANet。该模型采用多通路的高分辨率网络的结构设计,保留高分辨率特征并保障识别准确率。融入注意力机制的特征融合方法,通过特征连接权重自学习实现多尺度特征的深度融合。各通路网络采用跳跃连接促进跨层特征融合,加速网络收敛,并利用公开数据集对车辆检测性能进行评估并验证。结果表明：所提模型的车辆检测性能优于3个传统模型,改进后的网络识别平均精度均值（mAP）指标接近95%,且对于不同场景下的检测具有良好的鲁棒性。

关键词: 车辆检测, 高分辨率网络, 注意力机制, 特征融合, 跳跃连接

Abstract:

A vehicle detection model named MCHRANet based on the multi-channel high-resolution network and attention mechanism was proposed to improve the safe, reliable, and stable operation of road traffic monitoring and intelligent connected transportation system (ICTS) in the roadside edge platform. The model adopted multi-channel high-resolution network structure design to retain high-resolution features and ensured detection robustness. The feature fusion method incorporating the attention mechanism realized the deep fusion of multi-scale features through the self-learning of feature connection weights. Each channel network adopted jump connections to promote cross-layer feature fusion and accelerate network convergence. A public data set was used to evaluate the vehicle detection performances in different scenarios. The results show that the vehicle detection performance of the proposed model is better than the three traditional models, the improved network recognition mean average precision (AP) index is close to 95%, and it has good robustness for detection in different scenarios.

Key words: vehicle detection, high-resolution network, attention mechanism, feature fusion, jump connection

中图分类号:

TP391.41

丁飞, 米冠宇, 童恩, 张楠, 暴建民, 张登银. 多通路高分辨率网络与注意力机制融合的车辆检测模型[J]. 汽车安全与节能学报, 2022, 13(1): 122-130.

DING Fei, MI Guanyu, TONG En, ZHANG Nan, BAO Jianmin, ZHANG Dengyin. Multi-channel high-resolution network and attention mechanism fusion for vehicle detection model[J]. Journal of Automotive Safety and Energy, 2022, 13(1): 122-130.

图/表 17

图1 通用网络结构

图2 特征融合网络结构

图3 跳跃连接示意图

图4 MCHRANet模型结构与应用架构

类别	实例数量
类别	训练集	测试集
轿车	49 940	12 630
卡车	3 043	776
总量	52 983	13 406

表1 实验数据集信息

图5 MCHRANet模型在不同场景下的车辆检测结果

图6 目标检测准确率指标对比图

图7 目标检测召回率指标对比图

图8 目标检测F度量值指标对比图

图9 不同检测模型的AP性能对比图

图10 中间网络输出特征融合对比图

图11 后端网络输出特征融合对比图

图12 不同场景下的数据损伤图

图13 不同网络架构下的检测鲁棒性对比图

图14 收敛性能对比图

图15 不同检测模型的Pre-Rec曲线对比图

图16 不同检测模型的mAP性能对比图

参考文献 24

[1]	LIU Kaiqi, WANG Jianqiang. Fast dynamic vehicle detection in road scenarios based on pose estimation with convex-hull model[J]. Sensors, 2019,19(14):1-18.
[2]	Divya T K, Pablo M P, Lucio M, et al. Self-awareness in intelligent vehicles: Feature based dynamic Bayesian models for abnormality detection[J]. Robot Auto Syst, 2020,134:1-16.
[3]	LIAN Yanqi, ZHANG Guoqing, Lee J, et al. Review on big data applications in safety research of intelligent transportation systems and connected/automated vehicles[J]. Accid Analy Prevent, 2020,146:1-13.
[4]	中国智能网联汽车产业创新联盟. 车路云一体化融合控制系统白皮书[R/OL]. [2020-10-06], http://www.chuangze.cn/third_down.asp?txtid=2805.
	China Industry Innovation Alliance for the Intelligent and Connected Vehicles. White paper on system of coordinated control by vehicle-road-cloud integration[R/OL]. [2020-10-06], http://www.chuangze.cn/third_down.asp?txtid=2805. (in Chinese)
[5]	Chetouane A, Mabrouk S, Jemili I, et al. A comparative study of vehicle detection methods in a video sequence[C]// Proceed Distributed Computing Emerging Smart Networks (DiCESN), Commun Comput Info Sci. Springer, Cham, 2020,1130:37-53.
[6]	Anandhalli M, Baligar V P. A novel approach in real-time vehicle detection and tracking using Raspberry Pi[J]. Alexandria Engineering J, 2017,57(3):1597-1607.
[7]	ZHANG Yunsheng, ZHAO Chihang, SHI Wen, et al. Vehicles detection for illumination changes urban traffic scenes employing adaptive local texture feature background model[J]. IET Intel Transport Syst, 2018,12(10):1283-1290.
[8]	Saha R, Debi T, Arefin M S. Developing a framework for vehicle detection, tracking and classification in traffic video surveillance[C]// Proceed Intel Comput Optim (ICO), Advan Intel Syst Comput. Springer, Cham, 2021,1324:326-341.
[9]	WANG Zhang, LIU Chunsheng, CHANG Faliang, et al. Multi-scale and occlusion aware network for vehicle detection and segmentation on UAV aerial images[J]. Remote Sensing, 2020,12(11):1-22.
[10]	Cheon M, Lee W, Yoon C, et al. Vision-based vehicle detection system with consideration of the detecting location[J]. IEEE Trans Intel Transport Syst, 2012,13(3):1243-1252.
[11]	Elkerdawi S M, Sayed R, Elhelw M. Real-time vehicle detection and tracking using Haar-like features and compressive tracking[C]// Proceed First Iberian Robotics Conf (ROBOT2013). Springer, Cham, 2014,252:381-390.
[12]	MU Kenan, HUI Fei, ZHAO Xiangmo. Multiple vehicle detection and tracking in highway traffic surveillance video based on sift feature matching[J]. J Info Proc Syst, 2016,12(2):183-195.
[13]	Marcomini L A, Cunha A L. A comparison between background modelling methods for vehicle segmentation in highway traffic videos[J]. arXiv preprint arXiv: 1810.02835, 2018: 1-12.
[14]	杨敏, 裴明涛, 王永杰, 等. 一种基于运动目标检测的视觉车辆跟踪方法[J]. 北京理工大学学报, 2014,34(4):370-375.
	YANG Min, PEI Mingtao, WANG Yongjie, et al. Video-based vehicle tracking based on moving object detection[J]. Trans Beijing Instit Tech, 2014,34(4):370-375. (in Chinese)
[15]	Shehata M Abo-Al-Ez R Zaghlool F, et al. Vehicles detection based on background modeling[J]. Int’l J Engineer Trends Tech, 2019,66(2):92-95.
[16]	Charouh Z, Ghogho M, Guennoun Z. Improved background subtraction-based moving vehicle detection by optimizing morphological operations using machine learning[C]// Proceed IEEE Int’l Symp Innov Intel Syst Appli (INISTA). Sofia, Bulgaria, 2019: 1-6.
[17]	徐谦, 李颖, 王刚. 基于深度学习的行人和车辆检测[J]. 吉林大学学报(工学版), 2019,49(5):1661-1667.
	XU Qian, LI Ying, WANG Gang. Pedestrian-vehicle detection based on deep learning[J]. J Jilin Univ (Engi Tech Edit), 2019,49(5):1661-1667. (in Chinese)
[18]	Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// Proceed IEEE Conf Comput Vision Pattern Recog (CVPR). Columbus, OH, USA, 2014: 580-587.
[19]	Girshick R. Fast R-CNN[C]// Proceed IEEE Int’l Conf Comput Vision (ICCV). Santiago, Chile, 2015: 1440-1448.
[20]	REN Shaoqing, HE Kaiming, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Trans Pattern Anal Mach Intel, 2017,39(6):1137-1149.
[21]	Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]// Proceed IEEE Conf Comput Vision Pattern Recog (CVPR). Las Vegas, NV, USA, 2016: 779-788.
[22]	Tahir H, Khan M S, Tariq M O. Performance analysis and comparison of faster R-CNN, mask R-CNN and ResNet50 for the detection and counting of vehicles[C]// Proceed Int’l Conf Comput, Commun, Intel Syst (ICCCIS). Greater Noida, India, 2021: 587-594.
[23]	WANG Jingdong, SUN Ke, CHENG Tianheng. et al. Deep High-Resolution Representation Learning for Visual Recognition[J]. IEEE Trans Patt Anal Mach Intel, 2021,43(10):3349-3364.
[24]	Buyval A, Gabdullin A, Mustafin R, et al. Real time vehicle and pedestrian tracking for didi udacity self-driving car challenge[C]// Proceed IEEE Int’l Conf Robot Auto (ICRA). Brisbane, QLD, Australia, 2018: 2064-2069.

多通路高分辨率网络与注意力机制融合的车辆检测模型

Multi-channel high-resolution network and attention mechanism fusion for vehicle detection model

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 17

参考文献 24

相关文章 4

编辑推荐

Metrics

本文评价

期刊信息

在线期刊

作者中心

审稿中心

联系我们

[1]	张平, 迟志诚, 陈一凡, 惠飞. 用于自动驾驶车辆的融合注意力机制多目标跟踪算法[J]. 汽车安全与节能学报, 2021, 12(4): 516-521.
[2]	焦鑫, 杨伟东, 刘全周, 李占旗, 贾鹏飞. 用于 ADAS 实时目标车辆检测的改进 SSD 算法[J]. 汽车安全与节能学报, 2020, 11(3): 337-344.
[3]	姜义成，李凡 . 基于深度可分离卷积和多级特征金字塔网络的行人检测[J]. JASE, 2020, 11(1): 94-101.
[4]	金立生，程蕾，成波. 基于毫米波雷达和机器视觉的夜间前方车辆检测[J]. JASE, 2016, 07(02): 167-174.