Improved yolo v5 with balanced feature pyramid and attention module for traffic sign detection

 
SSD300 [10] Faster RCNN [9] YOLOv3 
YOLOv4 YOLOv5 
Ours 
mAP@.5 63.71% 79.1% 
82.4% 86.8% 
87.8% 
89.7% 
In table 2, our method gets the best results on the TT100K dataset. Therefore, we can 
conclude that our method is effective on the dataset, and has made progress compared with 
the original method. 
5 Conclusion 
Fig. 5. Detection results in TT100K dataset. The detection results of small targets are marked with red 
boxes. 
In this paper, we proposed an improved YOLO v5 to solve the problems existing in 
traffic sign detection. Aiming at traffic sign in small size, we chose the TT100K dataset. By 
comparing it with state-of-the-art methods mentioned above, our methods are better in 
accuracy. In the future, we will keep attempting to modify our model, so as to more suitable 
for automatic driving.
References 
1.
Redmon J, Divvala S, Girshick R, et al. You Only Look Once: Unified, Real-Time 
Object Detection[J]. Computer Vision & Pattern Recognition, 2016.
2.
Liu W, Anguelov D, Erhan D , et al. SSD: Single Shot MultiBox Detector[J]. European 
Conference on Computer Vision, 2016.
3.
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. 2014: 580-587. 
MATEC Web of Conferences 355, 03023 (2022) 
ICPCM2021
https://doi.org/10.1051/matecconf/202235503023
6

---

4.
Felzenszwalb P F, Mcallester D A, Ramanan D . A discriminatively trained, 
multiscale, deformable part model[C]// 2008 IEEE Conference on Computer Vision 
and Pattern Recognition. IEEE, 2008.
5.
Girshick R. Fast r-cnn[C]//Proceedings of the IEEE international conference on 
computer vision. 2015: 1440-1448. 
6.
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 & Machine 
Intelligence, 2017, 39(6):1137-1149.
7.
He K, Zhang X, Ren S, et al. Deep residual learning for image 
recognition[C]//Proceedings of the IEEE conference on computer vision and pattern 
recognition. 2016: 770-778.
8.
Zhe Z, Liang D, Zhang S, et al. Traffic-Sign Detection and Classification in the 
Wild[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition 
(CVPR). IEEE, 2016.
9.
Yao Z, Song X, Zhao L, et al. Realtime method for traffic sign detection and 
recognition based on YOLOv3tiny with multiscale feature extraction[J]. Proceedings 
of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, 
2021, 235(7): 1978-1991. 
10.
Pan W, Liu B, Chen Y, et al. Traffic sign detection and recognition based on YOLO 
v3[J]. Transducer and Microsystem Technologies, 2019.
11.
Pang J, Chen K, Shi J, et al. Libra R-CNN: Towards Balanced Learning for Object 
Detection[C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern 
Recognition (CVPR). IEEE, 2020. 
MATEC Web of Conferences 355, 03023 (2022) 
ICPCM2021
https://doi.org/10.1051/matecconf/202235503023
7