PROACTIVE RECOGNITION OF RISKY DRIVING BEHAVIOR OF VEHICLES USING DEEP LEARNING

Today, proactive road safety management is essential to reduce the number of accidents, particularly in dense traffic environments. In this context, this study introduces a deep learning framework for predicting vehicle trajectories and identifying risky behaviors from natural traffic data. The developed system combines Yolo and ByteTrack for detection and multi-object tracking and an Lstm seq2seq model to predict future trajectories from short observation windows. From these predictions, three dynamic indicatorsspeed, acceleration, and steering angleare extracted, and a threshold-based mechanism classifies maneuvers as normal or risky. The performance is evaluated using the Ade, Fde, Auc and Eer metrics over different periods of the Ngsim-US101 dataset. The results show moderate prediction accuracy, with higher Ade/Fde errors during periods of dense traffic and a notable improvement when traffic density decreases. The system achieves an AUC of 0.74 and an EER of 0.316, indicating a reasonable ability to distinguish between normal and risky behaviors, while highlighting room for improvement. Analyses also confirm the major influence of traffic density on prediction stability and classification performance.Overall, this study demonstrates the feasibility and potential of combining detection, tracking, trajectory prediction, and dynamic analysis for proactive recognition of risky behaviors in real-world conditions. These results pave the way for future improvements, particularly by enhancing the thresholding mechanism using adaptive or learning-based strategies.