基于密集连接卷积神经网络的道路车辆检测与识别算法

doi:10.11772/j.issn.1001-9081.2021030384

《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (3): 883-889.DOI: 10.11772/j.issn.1001-9081.2021030384

• 人工智能 • 上一篇

基于密集连接卷积神经网络的道路车辆检测与识别算法

邓天民, 冒国韬(), 周臻浩, 段志坚

重庆交通大学交通运输学院，重庆 400074

收稿日期:2021-03-15 修回日期:2021-06-22 接受日期:2021-06-23 发布日期:2022-04-09 出版日期:2022-03-10
通讯作者: 冒国韬
作者简介:邓天民（1979—），男，四川阆中人，副教授，博士，主要研究方向：交通大数据、交通环境感知
周臻浩（1995—），男，浙江宁波人，硕士研究生，主要研究方向：交通信息与控制、交通环境感知
段志坚（1997—），男，湖南株洲人，硕士研究生，主要研究方向：深度学习、交通环境感知。
基金资助:
国家重点研发计划项目(SQ2020YFF0418521);中央引导地方科技发展专项(CSTC2020JSCX?DXWTB0003);川渝联合实施重点研发项目(CSTC2020JSCX?CYLHX0007)

Road vehicle detection and recognition algorithm based on densely connected convolutional neural network

Tianmin DENG, Guotao MAO(), Zhenhao ZHOU, Zhijian DUAN

School of Traffic and Transportation，Chongqing Jiaotong University，Chongqing 400074，China

Received:2021-03-15 Revised:2021-06-22 Accepted:2021-06-23 Online:2022-04-09 Published:2022-03-10
Contact: Guotao MAO
About author:DENG Tianmin， born in 1979， Ph. D.， associate professor. His research interests include traffic big data， traffic environment perception.
ZHOU Zhenhao， born in 1995， M. S. . candidate. His research interests include traffic information and control， traffic environment perception.
DUAN Zhijian， born in 1997， M. S. candidate. His research interests include deep learning， traffic environment perception.
Supported by:
National Key Research and Development Program of China(SQ2020YFF0418521);Special Project of Central Government Guiding Local Science and Technology Development(CSTC2020JSCX-DXWTB0003);Sichuan and Chongqing Jointly Implemented Key Research and Development Project(CSTC2020JSCX-CYLHX0007)

摘要/Abstract

摘要：

针对现有道路车辆检测识别算法中存在的检测精度不高、实时性差以及小目标车辆漏检等问题，提出一种基于密集连接卷积神经网络的道路车辆检测与识别算法。首先，基于YOLOv4网络框架，通过采用密集连接的深度残差网络结构，加强特征提取阶段的特征复用，实现对浅层复杂度较低的特征的利用；然后，在多尺度特征融合网络引入跳跃连接结构，强化网络的特征信息融合和表征能力，以降低车辆漏检率；最后，采用维度聚类算法重新计算先验框尺寸，并按照合理的策略分配给不同检测尺度。实验结果表明，该算法在KITTI数据集上获得了98.21%的检测精度和48.05 frame/s的检测速度，对于BDD100K数据集中复杂恶劣环境中的车辆也有较好的检测效果，在满足实时检测要求的同时有效提升检测精度。

关键词: 智能交通, 道路车辆检测, YOLOv4, 密集连接网络, 多尺度特征融合

Abstract:

Regarding to the problems of low detection accuracy， poor real-time performance， and missed detection of small target vehicles in existing road vehicle detection and recognition algorithms， a road vehicle detection and recognition algorithm based on densely connected convolutional neural networks was proposed. Firstly， Based on YOLOv4 （You Only Look Once version 4） network framework， by adopting the densely connected deep residual network structure， the feature reuse in the feature extraction stage was strengthened to realize the use of features with lower complexity on shallow layers. Then， a jump connection structure was integrated to the multi-scale feature fusion network to strengthen the feature information fusion and expression capability of the network， which reduced the missed detection rate of vehicles. Finally， the dimensional clustering algorithm was used to recalculate the anchor sizes， which were allocated to different detection scales according to a reasonable strategy. Experimental results show that the proposed algorithm achieves the detection accuracy of 98.21% and the detection speed of 48.05 frame/s on KITTI dataset， and it also has a good detection effect for vehicles in the complex and harsh environment of Berkeley DeepDrive （BDD100K） dataset， ensuring required real-time performance and effective accuracy improvement.

Key words: intelligent transportation, road vehicle detection, You Only Look Once version 4 (YOLOv4), densely connected network, multiscale feature fusion

中图分类号:

TP391.4

邓天民, 冒国韬, 周臻浩, 段志坚. 基于密集连接卷积神经网络的道路车辆检测与识别算法[J]. 计算机应用, 2022, 42(3): 883-889.

Tianmin DENG, Guotao MAO, Zhenhao ZHOU, Zhijian DUAN. Road vehicle detection and recognition algorithm based on densely connected convolutional neural network[J]. Journal of Computer Applications, 2022, 42(3): 883-889.

图/表 17

图1 残差模块结构

Fig. 1 Residual block structure

图2 DenseNet密集连接结构

Fig. 2 Dense connected structure of DenseNet

图3 DRNet结构

Fig. 3 Structure of DRNet

图4 三种不同的CSPBlock

Fig. 4 Three types of CSPBlock

图5 CSPDRNet53结构

Fig. 5 Structure of CSPDRNet53

图6 PANet结构

Fig. 6 Structure of PANet

图7 引入跳跃连接结构的特征融合网络

Fig. 7 Feature fusion network introduced by skip connection structure

图8 提出的DR-YOLOv4算法结构

Fig. 8 Structure of proposed DR-YOLOv4 algorithm

图9 部分样本示意图

Fig. 9 Schematic diagram of partial samples

表1 不同特征提取网络对模型性能影响对比

Tab. 1 Influence comparison of different feature extraction networks on model performance

网络结构	mAP/%	检测速度/（frame·s^-1）
YOLOv4	97.41	45.72
YOLOv4+密集连接	97.89	40.35
YOLOv4+跨层连接	97.35	46.41
YOLOv4+深度残差	96.44	55.92
DR-YOLOv4特征提取网络	96.93	51.39

表2 不同特征融合网络对模型性能影响对比

Tab. 2 Influence comparison of different feature fusion networks on model performance

网络结构	mAP/%	检测速度/（frame·s^-1）
YOLOv4	97.41	45.72
YOLOv4+跳跃连接	97.47	45.39
YOLOv4+多尺度特征检测	97.98	39.87
DR-YOLOv4特征融合网络	98.09	39.68

表3 不同先验框分配策略对模型性能影响对比

Tab. 3 Influence comparison of different anchor allocation strategies on model performance

13×13 特征图	26×26 特征图	52×52 特征图	104×104 特征图	mAP/%	检测速度/ （frame·s^-1）
3	3	3	3	98.09	39.68
2	4	3	3	98.25	41.83
2	3	4	3	98.27	42.51
4	2	2	4	98.01	39.14
2	4	4	2	98.19	40.63
1	3	4	4	98.13	39.64
4	4	2	2	97.96	37.92

表4 不同实验方案对模型性能影响分析

Tab. 4 Analysis of influence of different experimental schemes on model performance

实验序号	特征提取网络改进	特征融合网络改进	先验框聚类与分配	各类别AP/%			mAP/%	检测速度/ （frame·s^-1）
实验序号	特征提取网络改进	特征融合网络改进	先验框聚类与分配	Car	Van	Truck	mAP/%	检测速度/ （frame·s^-1）
A	○	○	○	95.68	97.72	98.83	97.41	45.72
B	●	○	○	95.22	97.23	98.35	96.93	51.39
C	○	●	○	97.09	98.23	98.94	98.09	39.68
D	○	○	●	96.60	98.37	98.92	97.96	48.54
E	●	●	○	96.69	97.76	98.49	97.65	46.17
F	○	●	●	98.08	98.91	99.09	98.69	41.88
G	●	○	●	96.17	97.89	98.45	97.50	53.92
H	●	●	●	97.48	98.51	98.65	98.21	48.05

图10 本文算法平均损失收敛曲线

Fig. 10 Average loss convergence curve of proposed algorithm

表5 不同目标检测模型在KITTI数据集上效果对比

Tab. 5 Comparison of effects of different target detection models on KITTI dataset

模型	各类别AP/%			mAP/%	检测速度/（frames·s^-1）
模型	Car	Van	Truck	mAP/%	检测速度/（frames·s^-1）
Faster R-CNN	78.25	72.93	80.09	77.09	12.47
SSD300	81.59	74.82	87.36	81.26	56.71
SSD512	84.81	70.72	82.98	79.50	26.24
YOLOv3	90.53	92.39	94.87	92.60	47.91
YOLOv4	95.68	97.72	98.83	97.41	45.72
DR-YOLOv4	97.48	98.51	98.65	98.21	48.05

表6 模型改进前后在BDD100K数据集上测试结果

Tab. 6 Test results on BDD100K dataset before and after model improvement

模型

各类别AP/%

mAP/%

检测速度/

（frames·s^-1）

图11 部分检测结果示意图

Fig. 11 Partial schematic diagram of test results

参考文献 22

1	张新钰，高洪波，赵建辉，等.基于深度学习的自动驾驶技术综述［J］.清华大学学报（自然科学版）， 2018， 58（4）：438-444. 10.16511/j.cnki.qhdxxb.2018.21.010
	ZHANG X Y， GAO H B， ZHAO J H， et al. Overview of deep learning intelligent driving methods ［J］. Journal of Tsinghua University （Science and Technology）， 2018， 58（4）：438-444. 10.16511/j.cnki.qhdxxb.2018.21.010
2	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks ［J］. Communication of the ACM, 2017, 60(6): 84-90. 10.1145/3065386
3	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation [C]// Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2014: 580-587. 10.1109/cvpr.2014.81
4	GIRSHICK R. Fast R-CNN [C]// Proceedings of the 2015 IEEE International Conference on Computer Vision. Piscataway: IEEE, 2015: 1440-1448. 10.1109/iccv.2015.169
5	REN S, HE K, 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, 2015, 39(6):1137-1149.
6	NGUYEN H. Improving faster R-CNN framework for fast vehicle detection ［J］. Mathematical Problems in Engineering, 2019,2019: 1-11. 10.1155/2019/3808064
7	杨薇, 王洪元, 张继, 等. 一种基于Faster-RCNN的车辆实时检测改进算法［J］. 南京大学学报(自然科学版), 2019, 55(2): 231-237. 10.13232/j.cnki.jnju.2019.02.008
	YANG W, WANG H Y, ZHANG J, et al. An improved vehicle real-time detection algorithm based on Faster-RCNN [J]. Journal of Nanjing University (Natural Science), 2019, 55(2): 231-237. 10.13232/j.cnki.jnju.2019.02.008
8	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection [C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016: 779-788. 10.1109/cvpr.2016.91
9	李珣, 刘瑶, 李鹏飞,等. 基Darknet框架下YOLOv2算法的车辆多目标检测方法[J]. 交通运输工程学报, 2018, 18(6): 146-162. 10.1155/2018/7075814
	LI X, LIU Y, LI P F, et al. Vehicle multitarget detection method based on YOLOv2 algorithm under Darknet framework [J]. Journal of Traffic and Transportation Engineering, 2018, 18(6): 146-162. 10.1155/2018/7075814
10	REDMON J, FARHADI A. YOLOv3: an incremental improvement [EB/OL]. [2019-12-15]. . 10.1145/3429945
11	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: optimal speed and accuracy of object detection [EB/OL]. [2020-04-23]. . 10.1109/cvpr46437.2021.01283
12	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector [C]// Proceedings of the 2016 European Conference on Computer Vision. Cham: Springer, 2016: 21-37. 10.1007/978-3-319-46448-0_2
13	WANG X, WANG S, CAO J, et al. Data-driven based tiny-YOLOv3 method for front vehicle detection inducing SPP-net [J]. IEEE Access, 2020, 8: 110227-110236. 10.1109/access.2020.3001279
14	赵文清, 周震东, 翟永杰. 基于反卷积和特征融合的SSD小目标检测算法[J]. 智能系统学报, 2020, 15(2): 310-316. 10.11992/tis.201905035
	ZHAO W Q, ZHOU Z D, ZHAI Y J. SSD small target detection algorithm based on deconvolution and feature fusion [J]. CAAI Transactions on Intelligent Systems, 2020,15(2):310-316. 10.11992/tis.201905035
15	WANG C Y, LIAO H Y M, WU Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN [C]// Proceedings of the 2020 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2020: 390-391. 10.1109/cvprw50498.2020.00203
16	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition [C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016: 770-778. 10.1109/cvpr.2016.90
17	HUANG G, LIU Z, LAURENS V D M, et al. Densely connected convolutional networks [C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 4700-4708. 10.1109/cvpr.2017.243
18	LIU S, QI L, QIN H, et al. Path aggregation network for instance segmentation [C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 8759-8768. 10.1109/cvpr.2018.00913
19	ARTHUR D, VASSILVITSKII S. K-Means++: the advantages of careful seeding [C]// Proceedings of the 8th Annual ACM-SIAM Symposium on Discrete algorithms. Stroudsburg, PA: Association for Computational Linguistics, 2007: 1027-1035.
20	REDMON J. Darknet: open source neural networks in C [EB/OL]. [2019-07-15]. .
21	GEIGER A, LENZ P, URTASUN R. Are we ready for autonomous driving? The KITTI vision benchmark suite [C]// Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2012: 3354-3361. 10.1109/cvpr.2012.6248074
22	YU F, CHEN H, WANG X, et al. BDD 100K: a diverse driving video database with scalable annotation tooling [EB/OL]. [2020-01-15]. .

[1]	谭兆一, 陈白帆. 基于坐标逆映射的增强型车辆三维全景影像[J]. 计算机应用, 2021, 41(4): 1165-1171.
[2]	徐婉晴, 王保栋, 黄艺美, 李金屏. 基于人体骨骼关键点的吸烟行为检测算法[J]. 《计算机应用》唯一官方网站, 2021, 41(12): 3602-3607.
[3]	楼豪杰, 郑元林, 廖开阳, 雷浩, 李佳. 基于Siamese-YOLOv4的印刷品缺陷目标检测[J]. 《计算机应用》唯一官方网站, 2021, 41(11): 3206-3212.
[4]	高建, 毛莺池, 李志涛. 基于高斯混合时间序列模型的轨迹预测[J]. 计算机应用, 2019, 39(8): 2261-2270.
[5]	陈龙杰, 张钰, 张玉梅, 吴晓军. 基于多注意力多尺度特征融合的图像描述生成算法[J]. 计算机应用, 2019, 39(2): 354-359.
[6]	陈珂. 用于视频中车速自动检测的摄像机自动标定方法[J]. 计算机应用, 2017, 37(8): 2307-2312.
[7]	徐晓伟, 杜一, 周园春. 基于多源出行数据的居民行为模式分析方法[J]. 计算机应用, 2017, 37(8): 2362-2367.
[8]	郑林江, 刘旭, 易兵. 考虑时空特性的动态权重实时地图匹配算法[J]. 计算机应用, 2017, 37(8): 2381-2386.
[9]	戚欣, 梁伟涛, 马勇. 基于出租车轨迹数据的最优路径规划方法[J]. 计算机应用, 2017, 37(7): 2106-2113.
[10]	董峻妃, 郑伯川, 杨泽静. 基于卷积神经网络的车牌字符识别[J]. 计算机应用, 2017, 37(7): 2014-2018.
[11]	张军, 胡震波, 朱新山. 基于AdaBoost分类器的实时交通事故预测[J]. 计算机应用, 2017, 37(1): 284-288.
[12]	甘玲, 李瑞. 基于自适应虚拟线圈的多车道车流量检测算法[J]. 计算机应用, 2016, 36(12): 3511-3514.
[13]	寇兰, 杨立娜, 刘科征, 胡敏, 毛一丁. 基于城市公共交通移动模型的协作延迟容忍网络路由策略[J]. 计算机应用, 2016, 36(11): 3021-3027.
[14]	任鹏飞, 秦贵和, 董劲男, 李滨, 郑啸天. 具有交通规则约束的改进Dijkstra算法[J]. 计算机应用, 2015, 35(9): 2503-2507.
[15]	曹波, 韩燕波, 王桂玲. 基于车牌识别大数据的伴随车辆组发现方法[J]. 计算机应用, 2015, 35(11): 3203-3207.

基于密集连接卷积神经网络的道路车辆检测与识别算法

Road vehicle detection and recognition algorithm based on densely connected convolutional neural network

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 17

参考文献 22

相关文章 15

编辑推荐

Metrics