Traffic Sign Detection Based on Channel Attention and Feature Enhancement

doi:10.12141/j.issn.1000-565X.220639

Abstract

Abstract:

Traffic signs on the road contain a large amount of semantic information about traffic rules, and rapid and accurate access to this information helps to achieve higher levels of assisted driving functions, thus improving vehicle’s safety performance. In view of the traffic signs are susceptible to external factors and the problems of high similarity between categories and small size, this paper made targeted improvements in data augmentation, feature extraction and feature enhancement based on YOLOv5s model. In the data augmentation part, color space transformation and geometric transformation matrix were used to simulate the possible color changes and shape changes of traffic signs in actual scenes, and the Mosaic algorithm and Copy-paste algorithm were used to improve the number of tiny traffic signs in the training set and the richness of the background. In the feature extraction part, a feature extraction module based on channel attention calibration was constructed to improve the model’s ability to discriminate similar features. In the feature enhancement part, the number of prediction branches and downsampling multiplier were optimized by fusing shallow features and deep features with a dual-path enhancement structure, so as to increase the detection accuracy of tiny traffic signs. In addition, the K-means++ algorithm was used to cluster the prior bounding box templates and construct the loss function based on the CIoU metric, thus reducing the difficulty of the prior bounding box regression. Experiments on the TT100K and CCTSDB dataset test show that the mAP@0.5 of the proposed model is 88.8% and 83.5% respectively, and the speed of the model is 120.5 f/s and 114.7 f/s respectively. Compared with the existing traffic sign detection models, the proposed model reaches the advanced level in both accuracy and speed. Comparison experiments for data augmentation algorithms, prediction branches, and channel attention module positions further demonstrate the effectiveness of the proposed specific optimization methods.

Key words: traffic sign detection, convolutional neural network, channel attention, feature enhancement

CLC Number:

TP391.41

LUO Yutao, GAO Qiang. Traffic Sign Detection Based on Channel Attention and Feature Enhancement[J]. Journal of South China University of Technology(Natural Science Edition), 2023, 51(12): 64-72.

Figures/Tables 15

Fig.1

Fig.2

Fig.3

Fig.4

Fig.5

Table 1

Fig.6

Fig.7

Table 2

Table 3

Fig.8

Table 4

Table 5

Table 6

Table 7

References 28

1	GIRSHICK R， DONAHUE J， DARRELL T，et al ．Rich feature hierarchies for accurate object detection and semantic segmentation［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Columbus：IEEE，2014：580-587.
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 Transactionson on Patern Analysis and Machine Inteligence，2017，39（6）：1137-1149.
3	GIRSHICK R. Fast R-CNN［C］∥Proceedings of the IEEE International Conference on Computer Vision. Santiago：IEEE，2015：1440-1448.
4	REDMON J， DIVVALA S， GIRSHICK R，et al ．You only look once：unified，real-time object detection［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition．Las Vegas：IEEE， 2016：779-788.
5	REDMON J， FARHADI A. Yolov3：an incremental improvement［J］．arXiv Preprint arXiv：，2018.
6	REDMON J， FARHADI A ．YOLO9000：better，faster，stronger［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu：IEEE，2017：7263-7271.
7	BOCHKOVSKIY A， WANG C Y， LIAO H Y M ．Yolov4：optimal speed and accuracy of object detection［J］．arXiv Preprint arXiv：，2020.
8	GE Z， LIU S， WANG F，et al ．Yolox：exceeding yolo series in 2021［J］．arXiv Preprint arXiv：，2021.
9	LIU W， ANGUELOV D， ERHAN D，et al ．SSD：single shot multibox detector［C］∥Proceedings of the 14th European Conference on Computer Vision. Amsterdam：Springer，2016：21-37.
10	王卜，何扬．基于改进YOLOv3的交通标志检测［J］．四川大学学报（自然科学版），2022，59（1）：57-67.
	WANG Bu， HE Yang ．Traffic sign detection based on improved YOLOv3［J］．Journal of Sichuan University（Natural Science Edition），2022，59（1）：57-67.
11	陈梦涛，余粟．基于改进YOLOV4模型的交通标志识别研究［J］．微电子学与计算机，2022，39（1）：17-25.
	CHEN Mengtao， YU Su ．Research on traffic sign recognition based on improved YOLOV4 model［J］．Microelectronics & Computer，2022，39（1）：17-25.
12	HOU Q B， ZHOU D Q， FENG J S ．Coordinate attention for efficient mobile network design［C］∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition．［S.l.］：IEEE，2021：13713-13722.
13	LIU S T， HUANG D， WANG Y H ．Learning spatial fusion for single-shot object detection［J］．arXiv Preprint arXiv：，2019.
14	LIU X， JIANG X K， HU H C，et al ．Traffic sign recognition algorithm based on improved YOLOv5s［C］∥Proceedings of the International Conference on Control，Automation and Information Sciences．Xi’an：IEEE，2021：980-985.
15	KINGMA D P， BA J ．Adam：a method for stochastic optimization［J］．arXiv Preprint arXiv：，2014.
16	SANDLER M， HOWARD A， ZHU M，et al ．Mobilenetv2：inverted residuals and linear bottlenecks［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition．Salt Lake City：IEEE，2018：4510-4520.
17	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 IEEE/CVF Conference on Computer Vision and Pattern Recognition．［S.l.］：IEEE，2020：390-391.
18	冯润泽，江昆，于伟光，等．基于两阶段分类算法的中国交通标志牌识别［J］．汽车工程，2022，44（3）：434-441，448.
	FENG Runze， JIANG Kun， YU Weiguang，et al ．Chinese traffic sign recognition based on two-stage classification algorithm［J］．Automotive Engineering，2022，44（3）：434- 441，448.
19	HE K M， ZHANG X Y， REN S Q，et al ．Deep residual learning for image recognition［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition．Las Vegas：IEEE，2016：770-778.
20	LIU S， QI L， QIN H F，et al ．Path aggregation network for instance segmentation［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition．Salt Lake City：IEEE，2018：8759-8768.
21	GHIASI G， CUI Y， SRINIVAS A，et al ．Simple copy-paste is a strong data augmentation method for instance segmentation［C］∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition．［S.l.］：IEEE，2021：2918-2928.
22	KRISHNA K， MURTY M N ．Genetic K-means algorithm［J］．IEEE Transactions on Systems，Man，and Cybernetics，Part B （Cybernetics），1999，29（3）：433-439.
23	ZHENG Z H， WANG P， LIU W，et al ．Distance-IoU loss：faster and better learning for bounding box regression［C］∥Proceedings of the AAAI Conference on Artificial Intelligence．New York：AAAI，2020，34（7）：12993-13000.
24	IOFFE S， SZEGEDY C ．Batch normalization：accelerating deep network training by reducing internal covariate shift［C］∥Proceedings of International Conference on Machine Learning．Lille：ACM，2015：448-456.
25	HU J， SHEN L， SUN G ．Squeeze-and-excitation networks［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition．Salt Lake City：IEEE，2018：7132-7141.
26	LIN T Y， DOLLÁR P， GIRSHICK R，et al ．Feature pyramid networks for object detection［C］∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition．Honolulu：IEEE，2017：2117-2125.
27	ZHANG H， CISSE M， DAUPHIN Y N，et al ．Mixup：beyond empirical risk minimization［J］．arXiv Preprint arXiv：，2017.
28	YUN S， HAN D， OH S J，et al ．Cutmix：regularization strategy to train strong classifiers with localizable features［C］∥Proceedings of the IEEE/CVF International Conference on Computer Vision．Seoul：IEEE，2019：6023-6032.

代号	输出特征图尺寸	先验框尺寸
H2	160×160	（17，18）、（23，25）、（35，37）
H3	80×80	（10，12）、（12，20）、（15，16）
H4	40×40	（4，5）、（6，7）、（8，9）

模型	精确率/%	召回率/%	mAP@0.5/%	速度/（f·s^-1）
王卜等^［10］	56.1	86.7	82.0	36.8
陈梦涛等^［11］			86.6	34.5
YOLOv3^［5］	44.5	82.4	70.9	48.5
YOLOv4^［7］	81.8	85.4	82.1	36.2
YOLOv5s	84.0	71.3	79.4	128.2
TSDNet	90.1	83.9	88.8	120.5

模型	精确率/%	召回率/%	mAP@0.5/%	速度/（f·s^-1）
SSD^［9］	86.4	55.0	56.6	4.9
YOLOv3^［5］	84.6	42.7	50.5	20.3
YOLOv4^［7］	76.2	52.5	51.7	16.6
YOLOv5s	90.8	69.2	76.3	123.5
TSDNet	90.7	76.8	83.5	114.7

数据增强优化	预测分支优化	C3-TCA	K-means++	CIoU	mAP@0.5/%	mAP/%	参数量	FLOPs/10⁹
					79.4	60.1	7 141 794	16.7
√					82.0	62.1	7 141 794	16.7
√	√				85.8	65.0	5 442 346	18.4
√	√	√			88.1	67.6	5 117 541	18.1
√	√	√	√		88.2	68.1	5 117 541	18.1
√	√	√	√	√	88.8	68.5	5 117 541	18.1

+H2	-H5	mAP@0.5/%	mAP/%	小目标mAP/%	中目标mAP/%	大目标mAP/%	参数量	FLOPs/10⁹
		82.0	62.1	43.2	70.5	81.3	7 141 794	16.7
√		82.7	62.5	48.0	69.0	80.7	7 293 590	19.9
√	√	85.8	65.0	53.2	70.8	80.8	5 442 346	18.4