Evaluating Edge-Based YOLOv5-ByteTrack Frameworks for Real-Time Traffic Signal Control
2025 IEEE International Conference on Big Data-BigData, Macau, Çin, 8 - 11 Aralık 2025, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/bigdata66926.2025.11494521
- Basıldığı Şehir: Macau
- Basıldığı Ülke: Çin
- Orta Doğu Teknik Üniversitesi Adresli: Evet
Özet
Real-time traffic signal control requires the ability to process high-volume video streams with low latency and high reliability, especially in dense urban environments. This study presents an edge-based adaptive intersection management framework that combines YOLOv5 Large for vehicle detection, ByteTrack for multi-object tracking, and a dual ROI method for lane-level density and directional flow estimation. The system operates entirely on an NVIDIA Jetson Orin Nano, where a TensorRT optimized pipeline achieves 26 to 30 milliseconds end-to-end latency while handling continuous 17 FPS fisheye video that amounts to approximately 183 GB per day. Four YOLOv5 variants (Nano, Small, Medium, Large) were evaluated under identical conditions to determine the most effective model for embedded deployment, with YOLOv5 Large achieving the highest mAP at 0.5 (0.78). A matched A/B field trial at an intersection in Etimesgut, Ankara showed a 31 percent reduction in average vehicle delay compared to a fixed-time baseline, along with improvements in queue length and green-time utilization. The framework also produces standardized mobility indicators such as ROI-based density counts, directional flows, trajectory fragments, and cycle-level performance metrics that can be integrated into municipal data platforms and digital twin simulations. The results demonstrate that edge-deployed video analytics can serve both as a real-time control mechanism and as a scalable data source for smart city mobility systems.