Abstract
Predicting pedestrian movement is critical for the safety of autonomous systems, yet it remains challenging due to complex social interactions and the scarcity of labeled training data. This paper proposes a novel trajectory forecasting framework that integrates a Social Reconstructor and a dynamic Pseudo-trajectory Generator. By leveraging masked trajectory reconstruction and a dedicated social loss function to prevent predicted collisions, the model produces more accurate and socially compliant forecasts. Evaluations on standard benchmarks like ETH/UCY and the Stanford Drone Dataset demonstrate that the proposed approach achieves state-of-the-art performance and generates higher-quality prediction distributions.
Introduction
For autonomous vehicles and robots to navigate safely, they must anticipate where pedestrians are going. This is not merely a physics problem; humans are social agents who constantly adjust their paths to avoid collisions or maintain proximity to others. Modeling these social dynamics is notoriously difficult. Furthermore, deep learning models for trajectory forecasting are data-hungry, yet collecting and labeling real-world trajectory data is expensive. While computer vision relies heavily on data augmentation, standard techniques like rotation or cropping do not translate well to time-series trajectory data. This work addresses both the data scarcity issue and the challenge of modeling realistic social behaviors simultaneously.
Methods
The proposed architecture is built around a Conditional Variational Autoencoder (CVAE) and transformer-based attention mechanisms. The framework introduces three key components to improve robustness and social compliance:
- Social Reconstructor: Alongside the main forecasting module, a parallel module learns to reconstruct past trajectories where specific timesteps have been randomly masked. This forces the model to learn stronger, more robust internal representations of pedestrian movement.
- Dynamic Pseudo-Trajectory Generation: Instead of static data augmentation, the model monitors training loss to identify "difficult" samples. It uses the Reconstructor to generate new, challenging pseudo-trajectories based on these hard examples, feeding them back into the training loop dynamically.
- Social Loss Function: A novel loss term explicitly penalizes the model if it predicts future pedestrian locations that are closer to each other than a specific minimum distance threshold ($\epsilon$). This physically grounds the predictions, ensuring the model doesn't forecast two people occupying the same space.
Key Results
The model was evaluated on standard pedestrian trajectory forecasting benchmarks, including the ETH/UCY collection and the Stanford Drone Dataset (SDD).
- State-of-the-Art Performance: On the SDD dataset, the model sets a new state-of-the-art with an Average Distance Error (ADE) of 0.33m and a Final Distance Error (FDE) of 0.57m. It also shows highly competitive results across the ETH/UCY datasets.
- Distribution Quality: Using Kernel Density Estimate-based Negative Log Likelihood (KDE-NLL), the authors demonstrate that the model produces more accurate and tightly clustered predictions than baselines, achieving an average KDE of 0.586 on ETH/UCY.
- Collision Avoidance: Ablation studies proved the effectiveness of the Social Loss. Removing it significantly increased the percentage of overlapping pedestrians in the predictions, highlighting its role in ensuring social compliance.
Discussion and Conclusion
The research highlights that dynamic augmentation based on training loss fluctuations is more effective than static pre-training augmentation for trajectory data. Furthermore, forcing a model to "fill in the blanks" of masked past trajectories creates internal representations that are highly beneficial for future prediction. While the Social Loss improves stability and compliance, it is sensitive to the chosen distance threshold ($\epsilon$). Future work may address perspective issues in non-birdseye view recordings, where optical overlaps can occur between pedestrians who are not actually colliding. Overall, the proposed method offers a robust framework for more reliable autonomous navigation in human-populated environments.
Summary updated