基于改进YOLOv7的交通信号灯检测

doi:10.16182/j.issn1004731x.joss.23-1562

摘要/Abstract

摘要：

针对通用物体检测算法在信号灯检测方面存在着识别精度较低的问题，提出一种专门针对交通信号灯检测任务的改进的YOLOv7算法。去掉20×20检测尺度，添加160×160的检测尺度，在使模型轻量化的同时增加浅层特征；结合BiFormer中提出的BRA(bi-level routing attention)与坐标轴注意力，针对交通信号灯位置特点提出了ABRA(axially-guided BRA)；针对IoU指标对物体尺寸敏感的问题，引入NWD(normalized wasserstein distance)度量，改进物体定位损失与置信度损失。实验结果表明：改进YOLOv7算法的mAP值达到了97.7%，比原始YOLOv7提高了11.4%，检测速度提升了90帧/s，计算复杂度降低了4.5%。

关键词: 交通信号灯检测, 注意力机制, BiFormer, NWD(normalized wasserstein distance), YOLOv7

Abstract:

An improved YOLOv7 is proposed to address the problem of low recognition accuracy in general object detection algorithms for traffic signal detection. The algorithm removes the 20×20 detection scale and adds a 160×160 detection scale to increase shallow features while making the model lightweight. It combines the bi-level routing attention (BRA) proposed in BiFormer with axial attention, and innovatively proposes axially-guided BRA (ABRA). This module is specifically designed for the characteristics of traffic signal positions. To address the issue of object size sensitivity to the IoU metric, the normalized wasserstein distance (NWD) measurement is introduced to improve object location loss and objectness loss. Experimental results show that on the S²TLD dataset, the improved YOLOv7 algorithm achieves a mAP value of 97.7%, which is an improvement of 11.4% over the original YOLOv7. The detection speed is increased by 90 frames/s, and the computational complexity is reduced by 4.5%.

Key words: traffic signal detection, attention mechanism, BiFormer, normalized wasserstein distance (NWD), YOLOv7

中图分类号:

TP183

郑岚月,张玉洁 . 基于改进YOLOv7的交通信号灯检测[J]. 系统仿真学报, 2025, 37(4): 993-1007.

Zheng Lanyue,Zhang Yujie . Traffic Signal Detection Based on Improved YOLOv7[J]. Journal of System Simulation, 2025, 37(4): 993-1007.

图/表 20

图1

图2

图3

图4

图5

图6

图7

图8

图9

图10

表1

图11

表2

表3

表4

图12

表5

表6

图13

表7

参考文献 29

1	方彬楠, 赵天舒. 解锁AI 特斯拉万亿美元市值可期[N]. 北京商报, 2023-09-13(008).
2	马艳. 我国自动驾驶汽车产业加速起跑[N]. 中国工业报, 2023-10-20(003).
3	孙迎春, 潘树国, 赵涛, 等. 基于优化YOLOv3算法的交通灯检测[J]. 光学学报, 2020, 40(12): 137-145.
	Sun Yingchun, Pan Shuguo, Zhao Tao, et al. Traffic Light Detection Based on Optimized YOLOv3 Algorithm[J]. Acta Optica Sinica, 2020, 40(12): 137-145.
4	Che Mingliang, Che Mingjun, Chao Zhenhua, et al. Traffic Light Recognition for Real Scenes Based on Image Processing and Deep Learning[J]. Computing and Informatics, 2020, 39(3): 439-463.
5	蔡自兴, 谷明琴. 城市环境中交通信号灯准确识别与状态估计[J]. 控制与决策, 2014, 29(1): 163-167.
	Cai Zixing, Gu Mingqin. Accuracy Recognition and State Estimation for Traffic Light in Urban Environment[J]. Control and Decision, 2014, 29(1): 163-167.
6	Kim H K, Park J H, Jung H Y. Vision Based Traffic Light Detection and Recognition Methods for Daytime LED Traffic Light[J]. IEMEK Journal of Embedded Systems and Applications, 2014, 9(3): 145-150.
7	李宗鑫, 秦勃, 王梦倩. 基于时空关系模型的交通信号灯的实时检测与识别[J]. 计算机科学, 2018, 45(6): 314-319.
	Li Zongxin, Qin Bo, Wang Mengqian. Real-time Detection and Recognition of Traffic Light Based on Time-space Model[J]. Computer Science, 2018, 45(6): 314-319.
8	Zong Wenhao, Chen Qijun. Traffic Light Detection Based on Multi-feature Segmentation and Online Selecting Scheme[C]//2014 IEEE International Conference on Systems, Man, and Cybernetics (SMC). Piscataway: IEEE, 2014: 204-209.
9	Shi Zhenwei, Zou Zhengxia, Zhang Changshui. Real-time Traffic Light Detection with Adaptive Background Suppression Filter[J]. IEEE Transactions on Intelligent Transportation Systems, 2016, 17(3): 690-700.
10	Ying Jie, Tian Jin, Lei Lei. Traffic Light Detection Based on Similar Shapes Searching for Visually Impaired Person[C]//2015 Sixth International Conference on Intelligent Control and Information Processing (ICICIP). Piscataway: IEEE, 2015: 376-380.
11	Gokul R, Nirmal A, Bharath K M, et al. A Comparative Study Between State-of-the-art Object Detectors for Traffic Light Detection[C]//2020 International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE). Piscataway: IEEE, 2020: 1-6.
12	Abraham Alvin, Purwanto Djoko, Kusuma Hendra. Traffic Lights and Traffic Signs Detection System Using Modified You Only Look Once[C]//2021 International Seminar on Intelligent Technology and Its Applications (ISITIA). Piscataway: IEEE, 2021: 141-146.
13	Zhao Ying, Feng Yiyuan, Wang Yueqiang, et al. Study on Detection and Recognition of Traffic Lights Based on Improved YOLOv4[J]. Sensors, 2022, 22(20): 7787.
14	钱伍, 王国中, 李国平. 改进YOLOv5的交通灯实时检测鲁棒算法[J]. 计算机科学与探索, 2022, 16(1): 231-241.
	Qian Wu, Wang Guozhong, Li Guoping. Improved YOLOv5 Traffic Light Real-time Detection Robust Algorithm[J]. Journal of Frontiers of Computer Science & Technology, 2022, 16(1): 231-241.
15	Yu F, Chen Haofeng, Wang Xin, et al. BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 2633-2642.
16	Greer R, Gopalkrishnan A, Landgren J, et al. Robust Traffic Light Detection Using Salience-sensitive Loss: Computational Framework and Evaluations[C]//2023 IEEE Intelligent Vehicles Symposium (IV). Piscataway: IEEE, 2023: 1-7.
17	Zhu Xizhou, Su Weijie, Lu Lewei, et al. Deformable DETR: Deformable Transformers for End-to-end Object Detection[C]//ICLR 2021. New York: ICLR, 2021: 3448.
18	Wang C Y, Bochkovskiy A, Liao Hongyuan. YOLOv7: Trainable Bag-of-freebies Sets New State-of-the-art for Real-time Object Detectors[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 7464-7475.
19	Zhu Lei, Wang Xinjiang, Ke Zhanghan, et al. BiFormer: Vision Transformer with Bi-level Routing Attention[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 10323-10333.
20	Zheng Zhaohui, Wang Ping, Ren Dongwei, et al. Enhancing Geometric Factors in Model Learning and Inference for Object Detection and Instance Segmentation[J]. IEEE Transactions on Cybernetics, 2022, 52(8): 8574-8586.
21	Liu Shu, Qi Lu, Qin Haifang, et al. Path Aggregation Network for Instance Segmentation[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 8759-8768.
22	Zhang Xiangping, Fan Honghui, Zhu Hongjin, et al. Improvement of YOLOV5 Model Based on the Structure of Multiscale Domain Adaptive Network for Crowdscape[C]//2021 IEEE 7th International Conference on Cloud Computing and Intelligent Systems (CCIS). Piscataway: IEEE, 2021: 171-175.
23	Wang C Y, Liao Hongyuan, Wu Y H, et al. CSPNet: A New Backbone that Can Enhance Learning Capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Piscataway: IEEE, 2020: 1571-1580.
24	Yang Xue, Yan Junchi, Liao Wenlong, et al. SCRDet++: Detecting Small, Cluttered and Rotated Objects via Instance-level Feature Denoising and Rotation Loss Smoothing[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(2): 2384-2399.
25	Ren Sucheng, Zhou Daquan, He Shengfeng, et al. Shunted Self-attention via Multi-scale Token Aggregation[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 10843-10852.
26	Wang Jinwang, Xu Chang, Yang Wen, et al. A Normalized Gaussian Wasserstein Distance for Tiny Object Detection[EB/OL]. (2022-06-14) [2023-11-12]. .
27	Vaswani A, Shazeer N, Parmar N, et al. Attention Is All You Need[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2017: 6000-6010.
28	Liu Peilin, Xie Zhaoyang, Li Taijun. Traffic Light Detection Based on Improved YOLOV5[C]//2023 8th International Conference on Intelligent Computing and Signal Processing (ICSP). Piscataway: IEEE, 2023: 1891-1895.
29	Terven Juan, Cordova-Esparza Diana. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS[EB/OL]. [2023-11-23]. .

类别	各类别平均精确度/%
红	95.5
绿	95.5
关	77.9
黄	76.2
总体	86.3

方法	平均精确度/ %	帧率/ (帧/s)	浮点运算量/G
YOLOv7_P1234	94.3	312	116.2
YOLOv7_P123	95.2	333	97.9

k值	平均精确度/%	浮点运算量/G
1	95.6	98.4
2	96.8	98.5
3	90.9	98.5
4	95.6	98.6
5	96.0	98.6

注意力模块	平均精确度/%
Vanilla Attention^[27]	93.9
BRA^[19]	88.1
ABRA	96.8

损失函数	平均精确度/%
原始损失函数	96.8
改进的损失函数	97.7