CNN-GRU Based Pedestrian Trajectory Prediction for Smart Cities

Pedestrian trajectory prediction is a critical component of smart city applications, including autonomous driving, crowd monitoring, and urban planning. Accurately forecasting pedestrian movements in dynamic and crowded environments is challenging due to factors such as unpredictable motion patterns, occlusions, missed detections, and identity switches. In this paper, we explore deep learning-based approaches for multi-frame pedestrian trajectory prediction using Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and a hybrid Convolutional Neural Network–GRU (CNN-GRU) model. Although the CNN-GRU architecture is a known combination, our contribution lies in its practical deployment on detection- and tracking-based trajectories, generated using BoT-SORT tracking over YOLOv11 detections, and its comprehensive evaluation across three challenging benchmark datasets: MOT16, MOT17, and MOT20. The models take past trajectory sequences as input and predict future pedestrian positions, with performance measured using Average Displacement Error (ADE) and Final Displacement Error (FDE). Experimental results demonstrate that the CNN-GRU model outperforms standalone recurrent models by effectively capturing spatial and temporal dependencies while remaining computationally efficient and suitable for real-time applications. This study provides valuable insights into practical deep learning-based pedestrian trajectory prediction, emphasizing robustness under realistic tracking conditions and its potential impact on pedestrian safety and traffic management in smart city environments.