A real-time traffic flow prediction framework for Ho Chi Minh city using camera based vehicle counting and transfer learning

Abstract

Traffic congestion in Ho Chi Minh City creates an urgent need for real-time traffic flow forecasting systems to support traffic monitoring and management. However, traffic data collected from cameras are often limited, noisy, and difficult to directly exploit for forecasting models. This study proposes an end-to-end real-time traffic flow forecasting framework that combines camera-based vehicle counting with transfer learning for spatio-temporal forecasting. Specifically, the system employs YOLOv8 to detect and count vehicles from traffic camera images, and then aggregates the results into traffic flow time series. The STGCN model is pretrained on the METR-LA traffic sensor dataset and fine-tuned on real-world traffic camera data collected in Ho Chi Minh City. Preliminary results on data from 8 cameras show that the proposed approach achieves stable forecasting performance and demonstrates feasibility under limited local data conditions. This study provides a practical direction for urban traffic forecasting from camera-based data.