Trajectory prediction of moving objects in traffic for autonomous vehicles using a deep learning-based end-to-end system

Autonomous vehicle technologies play a critical role in enhancing traffic safety and improving transportation efficiency. In this thesis, a new method called End-to-End Route, which predicts the trajectories of moving objects in traffic in an end-to-end manner—including object detection and tracking processes—is proposed. Additionally, a deep learning-based model named ROUTE, developed for trajectory prediction, is introduced. In the study, moving objects in traffic were detected using data obtained from LiDAR sensors. These objects were then tracked across consecutive point cloud frames, with unique IDs assigned and motion trajectories constructed. For object detection and tracking stages, open-source and pre-trained deep learning models were utilized. Through the 3D bounding boxes obtained during detection, the position, velocity, acceleration, and orientation of each moving object were extracted. These features were then used to determine the temporal tracks and identities of the objects. For the trajectory prediction stage, the ROTA model was developed by integrating a heterogeneous graph structure with spatio-temporal deep learning models. This model aims to accurately predict the possible future trajectories of traffic participants by learning their interactions and motion dynamics. The proposed method and model have demonstrated successful results in predicting the trajectories of various types of moving objects, including vehicles, pedestrians, and cyclists. For 8-second future trajectory predictions, the method achieved an average 1.64 m Average Displacement Error (ADE) and 3.91 m Final Displacement Error (FDE), corresponding to improvements of 12.5% and 18% respectively compared to existing methods. These findings indicate that the proposed method and model offer a strong alternative in terms of accuracy and effectiveness.