Trajformer: Trajectory Prediction with Local Self-Attentive Contexts for Autonomous Driving

Abstract

Effective feature-extraction is critical to models' contextual understanding, particularly for applications to robotics and autonomous driving, such as multimodal trajectory prediction. However, state-of-the-art generative methods face limitations in representing the scene context, leading to predictions of inadmissible futures. We alleviate these limitations through the use of self-attention, which enables better control over representing the agent's social context; we propose a local feature-extraction pipeline that produces more salient information downstream, with improved parameter efficiency. We show improvements on standard metrics (minADE, minFDE, DAO, DAC) over various baselines on the Argoverse dataset. We release our code at: https://github.com/Manojbhat09/Trajformer

Accepted: Machine Learning for Autonomous Driving @ NeurIPS 2020

References in corpus (5)

• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Argoverse: 3D Tracking and Forecasting with Rich Maps
• Learn to Combine Modalities in Multimodal Deep Learning
• DESIRE: Distant Future Prediction in Dynamic Scenes with Interacting Agents
• Learning Lane Graph Representations for Motion Forecasting

Cited by in corpus (3)

• Facial Expression Recognition with Visual Transformers and Attentional Selective Fusion
• Trajectory Prediction for Autonomous Driving: Progress, Limitations, and Future Directions
• Scene Transformer: A unified architecture for predicting multiple agent trajectories