A Hybrid RL-GNN Approach for Precise Pedestrian Trajectory Prediction in Autonomous Navigation

With the rapid advancement of reinforcement learning and intelligent transport system, the adoption of autonomous vehicles has been increased in real world. Dynamic pedestrian environments faces significant challenges in accurately predicting pedestrian trajectories and ensuring safety in autonomous vehicles in places of crosswalks, sidewalks, and roundabouts. To address this challenges, this paper presents a hybrid work of integrating reinforcement learning (RL) and Graph neural network(GNN) to enhances the accuracy and adaptability of trajectory prediction by efficiently modeling pedestrian interactions and dynamically adjusting predictions in real time. The Model was evaluated using state of the art dataset ETH, UCY and nuScenes featuring diverse dynamic objects in the real world. Results from experiments conducted on benchmark datasets indicate our GNN-RL-based pedestrian trajectory prediction model outperform existing machine learning and Transformer-based models, with an Average Displacement Error (ADE) of 0.6-0.8m and an Final Displacement Error (FDE) of 1.2-1.8m. Furthermore, our technique decreases accident rates by up to 50%, improving pedestrian safety during autonomous navigation.