Performance Analysis of YOLOv4 Tiny for Real-Time Traffic Sign Recognition on Edge AI Systems Under Adverse Conditions
doi.org/10.63339/FME20250301
Keywords:
Traffic Sign Detection, YOLOv4-Tiny, Deep Learning, Convolutional Neural Networks (CNN), Object Detection, Real-time Image Processing, Machine Learning, Computer VisionAbstract
Traffic sign detection and recognition are two important capabilities of intelligent transportation systems or autonomous vehicles that can be used to feature real-time traffic sign detection. This paper is a proposal for real time lightweight traffic sign detection system running on Raspberry Pi 4 Model B with employment of YOLOv4-Tiny for its object classification and detection. Training was conducted on German Traffic Sign Recognition Benchmark (GTSRB) dataset which has 50,000+ images of different traffic signs in total of 43 classes. Image augmentation, contrast normalization and adaptive thresholding techniques were used to enhance the detection performance with different illumination on the image. We train the model on Darknet for 100 epochs with a batch size of 64 and achieve 96.5% test set accuracy as well as a mAP of 94.2%. Using OpenVINO and TensorRT, we optimized the model to be used for real time inference on Raspberry Pi that was achieved with 45ms per frame latency. Performance evaluations revealed that the system could consistently determine and classify traffic signs in the daylight, low light, and occlusion cases with 2.3% false positives and 3.7% false negatives. Aspect input speed of our YOLO model in comparison with conventional machine learning methods is 32% faster (HOG SVM), 110% faster (haar cascades), 21% more accurate (HOG SVM), and 18% more accurate (haar cascades). We have deployed it on edge devices, such as Raspberry Pi, which makes it suitable to be used in real world applications of intelligent cars, road surveillance, and driver assistance systems. Future work toward decreasing power consumption, adding multi class objects tracking and increasing robustness to incorrect motion blur and bad weather will be addressed.
References
[1] A. Bochkovskiy, C. Y. Wang, and H. Y. M. Liao, "YOLOv4: Optimal Speed and Accuracy of Object Detection," arXiv preprint arXiv:2004.10934, 2020.
[2] M. H. Asad, S. Khaliq, M. H. Yousaf, M. O. Ullah, and A. Ahmad, "Pothole Detection Using Deep Learning: A Real‐Time and AI‐on‐the‐Edge Perspective," Adv. Civ. Eng., vol. 2022, p. 9221211, 2022.
[3] Y. Zhang, C. Song, and D. Zhang, "Deep Learning-Based Object Detection Improvement," IEEE Access, vol. 8, pp. 56607–56614, 2020.
[4] M. Algabri, H. Mathkour, M. A. Bencherif, et al., "Towards Deep Object Detection Techniques for Phoneme Recognition," IEEE Access, vol. 8, pp. 54663–54680, 2020.
[5] R. Girshick, J. Donahue, T. Darrell, and J. Malik, "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Columbus, OH, USA, Jun. 2014, pp. 125–138.
[6] R. Girshick, "Fast R-CNN," in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), Santiago, Chile, Dec. 2015, pp. 127–135.
[7] S. Ren, K. He, R. Girshick, and J. Sun, "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks," IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, pp. 1137–1149, 2017.
[8] J. Dai, Y. Li, K. He, and J. Su, "R-FCN: Object Detection via Region-Based Fully Convolutional Networks," in Proc. 30th Int. Conf. Neural Inf. Process. Syst. (NIPS), Barcelona, Spain, Dec. 2016, pp. 379–387
[9] S. H. Park, H. S. Yoon, and K. R. Park, "Faster R-CNN and Geometric Transformation-Based Detection of Driver’s Eyes Using Multiple Near-Infrared Camera Sensors," Sensors, vol. 19, no. 1, pp. 1–29, 2019.
[10] P. N. R. S. Vara, B. D'Souza, and K. Bhargava, "A Downscaled Faster-RCNN Framework for Signal Detection and Time-Frequency Localization in Wideband RF Systems," IEEE Trans. Wireless Commun., vol. 19, no. 7, pp. 4847–4862, 2020.
[11] L. Zhou, W. Min, D. Lin, et al., "Detecting Motion Blurred Vehicle Logo in IoV Using Filter-DeblurGAN and VL-YOLO," IEEE Trans. Veh. Technol., vol. 69, no. 4, pp. 3604–3614, 2020.
[12] H. Zhang, L. Qin, J. Li, et al., "Real-Time Detection Method for Small Traffic Signs Based on YOLOv3," IEEE Access, vol. 8, pp. 64145–64156, 2020.
[13] O. A. Egaji, G. Evans, M. G. Griffiths, and G. Islas, "Real-Time Machine Learning-Based Approach for Pothole Detection," Expert Syst. Appl., vol. 184, p. 115562, 2021.
[14] A. Faisal and S. Gargoum, "Cost-Effective LiDAR for Pothole Detection and Quantification Using a Low-Point-Density Approach," Autom. Constr., vol. 172, p. 106006, 2025.
[15] B. H. Kang and S. I. Choi, "Pothole Detection System Using 2D LiDAR and Camera," in Proc. 9th Int. Conf. Ubiquitous Future Netw. (ICUFN), Jul. 2017, pp. 744–746.
[16] D. Chen, N. Chen, X. Zhang, and Y. Guan, "Real-Time Road Pothole Mapping Based on Vibration Analysis in Smart City," IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 15, pp. 6972–6984, 2022.
[17] H. Xin, Y. Ye, X. Na, H. Hu, G. Wang, C. Wu, and S. Hu, "Sustainable Road Pothole Detection: A Crowdsourcing-Based Multi-Sensors Fusion Approach," Sustainability, vol. 15, no. 8, p. 6610, 2023.
[18] L. Li, J. Liu, J. Xing, Z. Liu, K. Lin, and B. Du, "Road Pothole Detection Based on Crowdsourced Data and Extended Mask R-CNN," IEEE Trans. Intell. Transp. Syst., 2024.
[19] P. Jiang, D. Ergu, F. Liu, Y. Cai, and B. Ma, "A Review of YOLO Algorithm Developments," Procedia Comput. Sci., vol. 199, pp. 1066–1073, 2022.
[20] A. Bochkovskiy, C. Y. Wang, and H. Y. M. Liao, "YOLOv4: Optimal Speed and Accuracy of Object Detection," arXiv preprint arXiv:2004.10934, 2020.
[21] C. Li, L. Li, H. Jiang, K. Weng, Y. Geng, L. Li, et al., "YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications," arXiv preprint arXiv:2209.02976, 2022.
