Publication: Pedestrian detection and trajectory prediction in complex urban environment using deep learning methods for autonomous vehicle system.
Loading...
Date
2025-08
Authors
Zhang, Feifei
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Pedestrian safety in complex urban environments is a major challenge for intelligent transportation. Many current methods fail to address occlusions, changing
scenes, and real-time demands. This study proposes a three-stage deep learning framework for pedestrian detection, tracking, and trajectory prediction. The main novelty is the integration of dynamic scene modeling and probabilistic reasoning with lightweight networks. For detection, a new feature extraction network improves small object detection in crowded scenes. The structure uses EDLAN and VoVGSCPS, which increase accuracy and reduce computation. Experiments show the F1 score and mAP@50 improve by 6.8%, and model size is 67% smaller on the Mix dataset. For tracking, data association is improved with an optimized Kalman filter and camera motion compensation. ByteTrack, enhanced with YOLOv8s, increases Order Tracking Accuracy (HOTA) and Multiple Object Tracking Precision (MOTP) by 5.7% and 7.1% on MOT17. The framework performs well in low-light and high-density scenes. For trajectory prediction, the Dynamic Scene Information Module (DSIM) combines Transformer and Long Short-Term Memory (LSTM) networks. The Mixture Density Network (MDN) provides multimodal prediction. On the DUT dataset, Average Displacement Error (ADE) and Final Displacement Error (FDE) decrease by 60% and 57.7%. The framework supports real-time processing at 133 FPS on Towncentre, which offers practical value for smart city planning and autonomous driving.