A comparative analysis of YOLOv5 and YOLOv7 object detecting models for speed-limit traffic-sign recognition

• Published in: Journal of physics. Conference series Vol. 2949; no. 1; pp. 12022 - 12031
• Main Authors: Quang, Lam Tran, Truong, Thinh Vo, Duong Ly, Phi, Nhat, Phi Lam, Dang, Long Tran
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.02.2025
• Subjects: Algorithms; Automatic vehicle identification systems; Cameras; detecting distance; Entrances; frame rate; Image quality; Inference; Jetson Nano; Machine learning; Object recognition; Performance evaluation; Real time; Residential areas; Traffic control; traffic sign recognition; Traffic signs; Traffic speed; Vehicle safety; YOLO model
• ISSN: 1742-6588; 1742-6596
• DOI: 10.1088/1742-6596/2949/1/012022

Traffic sign recognition is a key element in automatic driver assist systems and autonomous vehicles, significantly improving driver’s comfort and driving safety. The You Only Look Once (YOLO) family of deep-learning-based object-detection algorithm has been widely adopted due to their real-time capabilities and ease of implementation. In this study, we compared the performance of Small-YOLOv5 and Tiny-YOLOv7, two recent variants of the YOLO architecture, on various traffic sign detection benchmarks. Our evaluation focused on three aspects: accuracy, speed/inference time and performance/computational complexity. The two models of YOLOv5 and YOLOv7 were trained with the same data set which consists of 3206 images representative for 7 different speed-limit signs including 50, 60, 70, 80, 90, 100 and 120 km/h, 2 signs of residential area entrance and exit, and 1 sign of all-limit removal. These images were collected in various real-life environmental conditions such as daylight, nighttime, rain, and motion blur from a Raspberry Pi V2 8MP Camera with a resolution of 3280x2464 pixels and a focal length of 3.04 mm. Comparisons were conducted in different scenarios of image quality, distance from camera, and computer resources. In terms of image quality, performance decreased significantly as image quality dropped with YOLOv5, especially in low light, blurry images, and bad weather. Meanwhile, YOLOv7 maintained better performance in various environmental conditions. YOLOv7 was found to be more stable and reliable than YOLOv5 when image quality dropped. In terms of speed/processing time, when running on a system with high resources (such as Google Colab T4), YOLOv7 outperformed YOLOv5 in speed/inference time. On low-end resources (Jetson Nano), YOLOv7 was also slightly better than YOLOv5. Finally, YOLOv7 was able to maintain better performance than YOLOv5 when the subject was at different distances from the camera. The findings of this study suggested that the Tiny-YOLOv7, when combined with a high-level computing platform, could be beneficial for real-time traffic sign recognition.