Vehicle egocentric object trajectory tracking system

With the rise of intelligent transportation systems, accurate pedestrian trajectory prediction is critical for autonomous driving safety. Traditional methods struggle with complex, dynamic environments and bird's-eye view datasets, limiting their applicability to vehicle-centric scenarios. This paper proposes the Multi-Feature Fusion YNet (MFF-YNet), an egocentric pedestrian trajectory prediction model that integrates multimodal features—head orientation, pedestrian behavior, status, and vehicle distance—using a novel fusion mechanism within the YNet architecture. Trained on a self-annotated Taiwan Pedestrian Road Use Dataset capturing localized hazardous scenarios, MFF-YNet significantly outperforms baselines in average displacement error (ADE) and final displacement error (FDE), achieving robust predictions for safety-critical applications.