Abstract
Trajectory prediction is an important task that involves modeling the indeterminate nature of traffic actors to forecast future trajectories given the observed trajectory sequences. However, current methods confine themselves to presumed data manifolds, assuming that trajectories strictly adhere to these manifolds, resulting in overly simplified predictions. To this end, we propose a novel approach called SSWNP (Self-Supervised Waypoint Noise Prediction). In our approach, we first create clean and noise-augmented views of past observed trajectories across the spatial domain of waypoints. We then compel the trajectory prediction model to maintain spatial consistency between predictions from these two views, in addition to the trajectory prediction task. Introducing the noise-augmented view mitigates the model's reliance on a narrow interpretation of the data manifold, enabling it to learn more plausible and diverse representations. We also predict the noise present in the two views of past observed trajectories as an auxiliary self-supervised task, enhancing the model's understanding of the underlying representation and future predictions. Empirical evidence demonstrates that the incorporation of SSWNP into the model learning process significantly improves performance, even in noisy environments, when compared to baseline methods. Our approach can complement existing trajectory prediction methods. To showcase the effectiveness of our approach, we conducted extensive experiments on three datasets: NBA Sports VU, ETH-UCY, and TrajNet++, with experimental results highlighting the substantial improvement achieved in trajectory prediction tasks.
1. Introduction
Accurately forecasting the future movement of agents—such as pedestrians or vehicles—is a cornerstone of autonomous navigation and safety systems. Despite the success of deep learning in this domain, existing models often fall into the "average trap." By learning primarily from the most frequent paths in training data, they develop a narrow interpretation of the data manifold. This leads to overly simplified, uniform predictions that fail to capture the inherent uncertainty and diversity of real-world motion.
Furthermore, many state-of-the-art models exhibit brittleness when faced with imperfect data. Sensor noise or minor irregularities in observed past trajectories can cause significant degradation in prediction accuracy. To address these challenges, we introduce SSWNP (Self-Supervised Waypoint Noise Prediction), a plug-and-play training strategy designed to enhance both the diversity and robustness of trajectory forecasting models.
2. Methodology
The core philosophy of SSWNP is to utilize noise as a supervisory signal rather than treating it merely as a nuisance. The architecture integrates two primary self-supervised modules into existing trajectory prediction frameworks.
2.1 Spatial Consistency Module
The system generates two distinct "views" of the observed past trajectory: a Clean View (the original sequence) and a Noise-Augmented View (where waypoints are perturbed by Gaussian noise). The model is trained to ensure that the future trajectories predicted from both views remain spatially consistent with each other and the ground truth. This consistency constraint encourages the model to extract invariant motion intent, effectively filtering out irrelevant spatial jitter.
2.2 Noise Prediction Module
As an auxiliary pretext task, the model is tasked with explicitly predicting the noise vector added to the augmented view. For the feature extractor to succeed at this task, it must develop a granular understanding of the difference between valid trajectory patterns and random noise. This leads to richer internal representations that better capture the underlying dynamics of the scene.
3. Experiments and Results
We evaluated SSWNP across three diverse benchmarks: NBA Sports VU (complex player interactions), ETH-UCY (social pedestrian dynamics), and TrajNet++ (large-scale agent-agent interactions).
3.1 Quantitative Performance
The integration of SSWNP yielded substantial improvements across all datasets. Notably, the Final Displacement Error (FDE), which measures accuracy at the end of the predicted horizon, saw the most significant gains:
- NBA Dataset: Achieved a 22.7% improvement in Average Displacement Error (ADE) and a 38.9% improvement in FDE over the GroupNet baseline.
- TrajNet++: Improved ADE by 33.8% and FDE by 36.4% compared to the AutoBot baseline.
3.2 Robustness to Noise
A key strength of SSWNP is its resilience to input corruption. In experiments where input data was intentionally degraded with noise, SSWNP-trained models maintained high accuracy while baseline models suffered significant performance drops. On the TrajNet++ benchmark, this translated to a 76.8% relative performance gain in noisy conditions.
4. Discussion and Conclusion
SSWNP demonstrates that self-supervised noise prediction is a powerful tool for regularizing trajectory forecasting models. By forcing the model to break away from a narrow data manifold, it achieves state-of-the-art accuracy and exceptional robustness.
While effective, the approach requires careful hyperparameter tuning for the noise factor and loss weights, which vary by agent type. Future work will explore automated tuning and more complex noise models—such as simulating sensor dropouts—to further enhance the practical applicability of this approach in real-world autonomous systems.
Summary updated