Utilizing Machine Learning Multi-Output Regression Methods to Predict Turning Movements at Intersections

Traffic signal timing design at intersections requires data on the turning movements (TMs) of vehicles that travel through the intersections. Turning movements at traffic intersections refer to the directional choices that drivers of the vehicles make—left turns, through movements, right turns, and U-turns—as they navigate the intersection. Traditional methods of TM data collection are costly, time-consuming, and do not deliver real-time data. Alternative techniques such as video image processing enable the near-real-time collection of approach volumes and turning movements. However, these methods require expensive, specialized equipment at the intersection. As a result, manual counting has remained the predominant method for collecting TM data at intersections. In this current study, the authors developed a robust method for collecting TM data. The TM counts, approach volume, intersection type, and lane configuration of 400 intersections around Ada County, Idaho, USA, were analyzed to develop an accurate and reliable turning-movement estimation model. A total of 2400 hours of TM counts were used. Three different machine learning multi-output regression methods were applied to analyze the relationship between the approach volumes and the corresponding turning movements, intersection type, and lane configuration. The results indicated that the developed model had a remarkable capability for accurately forecasting TMs.