Machine learning for intelligent transportation systems

Focusing on prediction, control, and perception tasks in Intelligent Transportation Systems (ITS), this study proposes a network-based traffic forecasting method integrating graph diffusion convolutions and gated recurrent units. It also outlines an adaptive signal control and deep detection-tracking monitoring process based on reinforcement learning. Evaluations were conducted using publicly available benchmarks (METR-LA) and 2024 domestic expressway corridor sensor/video data, employing metrics such as MAE, RMSE, and MAPE. Results demonstrate: The constructed graph-recurrent model reduces MAE by approximately 20% compared to LSTM and 35% compared to seasonal ARIMA for 15/30/60-minute forecasts, with more pronounced advantages at longer time horizons. In four-way intersection micro-simulations, the deep reinforcement learning strategy reduces average delay by about 15%. Under daytime conditions, YOLOv5 achieves vehicle detection accuracy >95% and recall ~90%. Thus, the learning paradigm integrating road network structure and temporal dynamics more accurately captures congestion formation and dissipation processes, providing a data foundation and methodological support for prediction-driven signal optimization and operational monitoring.