Robust Prediction Method for Pedestrian Trajectories in Occluded Video Scenarios

Abstract

In smart surveillance and intelligent transportation systems, accurately predicting pedestrian movements, especially under occlusions, remains a significant challenge. Traditional surveillance methods, primarily based on fixed CCTV footage, often overlook occlusions, noise, and camera angle changes, leading to inaccuracies in pedestrian trajectory predictions. Addressing this gap, our research introduces a novel trajectory prediction system optimized for various walking and filming conditions. By leveraging the ETH/UCY dataset, we simulate scenarios where pedestrians are temporarily obscured, employing re-identification techniques upon their reappearance to ensure path continuity. Our system utilizes linear prediction to reconstruct missing paths, integrating these with the PECNET baseline model for future path forecasting. This methodology allows for the effective handling of occlusions, a common yet underaddressed issue in current studies. Performance evaluation using Average Displacement Error (ADE) and Final Displacement Error (FDE) metrics reveals our system achieves an ADE of 0.25 meters and an FDE of 0.50 meters, demonstrating comparable accuracy to predictions based on fully visible paths within the same dataset. These results highlight our approach's efficacy in complex environments, marking a significant step forward in occlusion-aware pedestrian movement analysis and prediction for enhanced surveillance system accuracy.