Local causal dynamic integrated global mode guidance transformer network for pedestrian trajectory prediction.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous vehicles require accurate pedestrian trajectory prediction for safe navigation.
• Existing methods struggle with complex spatiotemporal dynamics and multi-modal future behaviors.
• Real-time performance is a critical constraint for autonomous driving systems.

Purpose of the Study:

• To introduce the Local-Global Collaborative Transformer Network (LGCMT) for improved pedestrian trajectory prediction.
• To address challenges in spatiotemporal dynamics, multi-modal prediction, and real-time performance.
• To develop a model balancing accuracy, multi-modality, and operational speed.

Main Methods:

• Developed a Local-Global Collaborative Encoder with Sparse Causal Temporal Attention (SCT-MSA) for local dynamics and Cosine Similarity Attention for global patterns.
• Implemented a parallel Non-Autoregressive (NAR) decoder guided by a motion pattern library for diverse trajectory generation.
• Evaluated the LGCMT model on standard benchmarks (ETH/UCY) and a large-scale dataset (Stanford Drone Dataset).

Main Results:

• LGCMT demonstrated robust performance on ETH/UCY and SDD datasets.
• Achieved significant improvements in Average Displacement Error (ADE) and Final Displacement Error (FDE) compared to the TUTR baseline.
• Showcased exceptional inference efficiency, suitable for real-time applications.

Conclusions:

• The LGCMT model effectively integrates local and global context for accurate pedestrian trajectory prediction.
• The proposed NAR decoder efficiently generates diverse and relevant future trajectory candidates.
• LGCMT offers a potent solution for real-time autonomous driving systems, balancing prediction accuracy and computational speed.