TrajDiff: Trajectory Prediction With Diffusion Probabilistic Models.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Diffusion probabilistic models (DPMs) have shown significant success in computer vision tasks.
• Accurate agent future trajectory prediction is crucial for applications like autonomous driving and robotics.

Purpose of the Study:

• To introduce TrajDiff, a novel model for agent future trajectory prediction using conditional diffusion probabilistic models.
• To improve the accuracy and efficiency of trajectory prediction by mapping the task to a latent heatmap space.

Main Methods:

• TrajDiff employs a U-Net architecture trained with a denoising objective.
• The model maps trajectory prediction to a latent heatmap space, enabling soft cluster center learning.
• A novel residual block with a mutual attention mechanism captures agent-environment interactions.

Main Results:

• TrajDiff achieves state-of-the-art performance on benchmark datasets (Stanford Drone, ETH, UCY).
• The model demonstrates considerable accuracy gains compared to existing methods.
• Significant reduction in computational requirements was observed.

Conclusions:

• TrajDiff offers a powerful and efficient approach for agent future trajectory prediction.
• The heatmap-based latent space and attention mechanism contribute to generating physically and socially acceptable trajectories.
• The model represents a significant advancement in the field of trajectory prediction.