Abstract
Human trajectory forecasting in crowds, at its core, is a sequence prediction problem with specific challenges of capturing inter-sequence dependencies (social interactions) and consequently predicting socially-compliant multimodal distributions. In recent years, neural network-based methods have been shown to outperform hand-crafted methods on distance-based metrics. However, these data-driven methods still suffer from one crucial limitation: lack of interpretability. To overcome this limitation, we leverage the power of discrete choice models to learn interpretable rule-based intents, and subsequently utilise the expressibility of neural networks to model scene-specific residual. Extensive experimentation on the interaction-centric benchmark TrajNet++ demonstrates the effectiveness of our proposed architecture to explain its predictions without compromising the accuracy.
Introduction
Human trajectory forecasting is a critical challenge in AI, often presenting a trade-off between interpretability and accuracy. Hand-crafted models like Discrete Choice Models (DCMs) are highly transparent but often fail to capture complex, long-term dependencies. Conversely, deep learning models like LSTMs are highly accurate but act as "black boxes," providing little explanation for their predictions. This paper introduces Social Anchors, a hybrid framework that bridges this gap by disentangling high-level social "intents" from low-level motion "residuals." By combining the interpretability of DCMs with the high accuracy of Deep Learning, the model achieves state-of-the-art performance on crowd navigation benchmarks while providing a clear, rule-based rationale for every prediction.
Methodology
The proposed architecture treats trajectory forecasting as a choice among discrete "social anchors" refined by neural network corrections.
The framework consists of four main components:
- Social Anchors: The output space around a pedestrian is discretized into a radial grid of possible future intents (anchors), defined by speed and direction.
- Discrete Choice Modeling (DCM): A set of hand-crafted, rule-based functions assigns probabilities to these anchors. These rules include Keep Direction (momentum), Avoid Occupancy (density avoidance), Leader-Follower (path following), and Collision Avoidance (proactive encounter avoidance).
- Neural Network Refinement: An LSTM-based encoder-decoder calculates a scene-specific residual. This small correction accounts for subtle social nuances that rigid rules might miss, turning a discrete anchor into a continuous coordinate.
- Direction Normalization: The scene is rotated at every time-step so the pedestrian of interest is always facing "right," making the model rotation-invariant and simplifying the learning task.
Experimental Results
The model was evaluated on the TrajNet++ benchmark, using both synthetic and real-world data.
- Accuracy: The SAnchor model achieved an Average Displacement Error (ADE) of 0.62m, performing at par with or better than complex generative models like SGAN and CVAE.
- Safety: The model significantly outperformed baselines in Collision-I (prediction collisions), reducing the rate to 4.2% (compared to 5.5% for S-LSTM and 5.9% for SGAN).
Interpretability: Visualizing the "Why"
A key contribution of this work is the ability to generate "activation maps" that explain the model's rationale. Because the DCM uses explicit rules, the probability assigned to each anchor can be decomposed into its constituent social factors.
As shown in the figure above, if the model predicts a turn, researchers can inspect these maps to see if the decision was motivated by collision avoidance, following a leader, or simply maintaining momentum. This transparency is vital for safety-critical applications like autonomous driving or robot navigation.
Conclusion and Future Work
The Social Anchors framework demonstrates that accuracy and interpretability are not mutually exclusive. By combining rule-based logic with neural network expressibility, the model provides a transparent decision-making process. Future work will focus on dynamic anchors that adapt to environmental features and extending these rules to mixed-traffic environments involving cyclists and vehicles.
Summary updated