Lane-Changing Maneuver Prediction With Urban Arterial Traffic Data From Drones

This study addresses the challenge of predicting lane-change behavior in complex urban arterial environments, where interactions are heterogeneous and lane changing events are inherently rare. We propose Lane Change Direction and Timing Network (LC-DiTiNet), a Transformer-based framework that jointly predicts lane change direction and timing using an event-based formulation. To effectively handle the strong class imbalance between lane-keeping and lane changing events, we introduce a two-stage training strategy with focal loss, which stabilizes learning and improves performance on rare maneuvers. In addition, we propose a novel evaluation metric, the Driver Maneuver Prediction Index (DMPI), which integrates spatial (direction) and temporal (timing) accuracy through a normalized soft-DTW formulation, enabling a holistic assessment of maneuver prediction. Extensive experiments on drone-based urban traffic data demonstrate that the proposed model achieves robust performance across multiple prediction horizons, outperforming existing baselines particularly in long-horizon and minority-class scenarios. Furthermore, the model is computationally efficient, with approximately 2.02 million parameters and an inference time of 5.7 ms per vehicle, making it suitable for real-time intelligent transportation system (ITS) applications. These results highlight the effectiveness of jointly modeling maneuver direction and timing for capturing complex driver behavior in urban traffic.