Abstract
SILA (Similarity-based Incremental Learning Algorithm) is a novel framework for predicting pedestrian trajectories that learns continuously as new data arrives. By fusing similar "motion primitives" and utilizing a common coordinate frame, SILA achieves a 20x speedup in training time and a 66% reduction in model size compared to standard batch-learning methods, while maintaining or exceeding state-of-the-art accuracy.
Introduction
Predicting pedestrian movement is critical for autonomous vehicle safety. Traditional "batch learning" models require massive static datasets and full retraining to incorporate new information, leading to significant scalability, privacy, and generalization issues. SILA addresses these by updating the model on-the-fly without requiring access to past raw data.
Methods
Motion Primitives and Transitions
SILA models movement using Motion Primitives—atomic building blocks like "walking straight" or "turning right." Transitions between these primitives are modeled using Gaussian Processes (GPs).
Incremental Learning and Model Fusion
The algorithm follows a three-step pipeline to manage model growth and incorporate new behaviors:
- Incremental Learning: A temporary local model ($M'$) is learned from new data batches.
- Similarity Mapping: New primitives are compared to existing ones using a "coherence" metric. Primitives with high similarity are linked in a Similarity Graph.
- Model Fusion: Based on the graph, similar behaviors are merged. Redundant behaviors are fused, while novel ones are added to the dictionary, preventing "model bloat."
To ensure generalization across different environments, SILA uses a common coordinate frame based on curbside distances. This allows the model to recognize similar behaviors (e.g., a right turn) regardless of specific intersection geometry.
Results
SILA was evaluated on custom intersection datasets in Boston, Cambridge, and Mcity, as well as standard ETH and UCY benchmarks.
- Efficiency: Training on mini-batches took ~0.05 seconds, a 20x speedup over batch learning.
- Accuracy: SILA showed a 5% improvement in Average Displacement Error (ADE) over state-of-the-art batch models like SGAN and GAT.
- Scalability: The fusion mechanism resulted in a model 3x smaller than naive incremental approaches.
Conclusion
SILA provides a privacy-preserving, scalable solution for trajectory prediction. Future work will extend the framework to multi-agent learning, accounting for real-time interactions between pedestrians and other moving agents.
Summary updated