Trajectory prediction for autonomous driving using deep learning approach

• Main Author: Du, Weifang
• Other Authors: Lyu Chen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Mechanical engineering
• Online Access: https://hdl.handle.net/10356/177789
• Institution: Nanyang Technological University

This project presents a detailed exploration of trajectory prediction models in autonomous driving, a critical component for enhancing the safety and efficiency of autonomous vehicles. The primary aim of this research is to develop a trajectory prediction algorithm using advanced deep learning techniques and evaluate its performance against conventional models. The study begins with a systematic literature review that establishes the foundation of trajectory prediction challenges and existing methodologies, including physical, maneuver-based, and interaction-aware models. To address the identified challenges, a new trajectory prediction framework is introduced, employing various deep learning architectures such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Graph Neural Networks (GNNs). The project uses multiple model architectures and assesses their performance in accurately predicting vehicle trajectories. The experimental setup is based on the Argoverse 1 Motion Forecasting Dataset, which provides high-quality, real-world data from various urban driving environments. The results demonstrate that the developed models, particularly the Graph and map-based models using attention mechanisms (GATConv) and LSTM encoders, significantly outperform traditional trajectory prediction approaches. The study not only highlights the effectiveness of using graph structures and deep learning techniques in capturing complex spatial and temporal patterns but also discusses the scalability and computational efficiency of these models. Furthermore, adjustments in the radius of identification for surrounding vehicles provide insights into the optimal setup for real-world applications, balancing contextual data intake and prediction accuracy.