TPK: Trustworthy Trajectory Prediction Integrating Prior Knowledge For Interpretability and Kinematic Feasibility

Abstract: Trajectory prediction is crucial for autonomous driving, enabling vehicles to navigate safely by anticipating the movements of surrounding road users. However, current deep learning models often lack trustworthiness as their predictions can be physically infeasible and illogical to humans. To make predictions more trustworthy, recent research has incorporated prior knowledge, like the social force model for modeling interactions and kinematic models for physical realism. However, these approaches focus on priors that suit either vehicles or pedestrians and do not generalize to traffic with mixed agent classes. We propose incorporating interaction and kinematic priors of all agent classes--vehicles, pedestrians, and cyclists with class-specific interaction layers to capture agent behavioral differences. To improve the interpretability of the agent interactions, we introduce DG-SFM, a rule-based interaction importance score that guides the interaction layer. To ensure physically feasible predictions, we proposed suitable kinematic models for all agent classes with a novel pedestrian kinematic model. We benchmark our approach on the Argoverse 2 dataset, using the state-of-the-art transformer HPTR as our baseline. Experiments demonstrate that our method improves interaction interpretability, revealing a correlation between incorrect predictions and divergence from our interaction prior. Even though incorporating the kinematic models causes a slight decrease in accuracy, they eliminate infeasible trajectories found in the dataset and the baseline model. Thus, our approach fosters trust in trajectory prediction as its interaction reasoning is interpretable, and its predictions adhere to physics.

Trajectory Prediction: Feasibility and Interpretability through Prior Knowledge Integration

In the field of autonomous driving, trajectory prediction is vital to ensure safe navigation by anticipating the movement of surrounding road users. The paper, "Prior Knowledge Integration for Feasible and Interpretable Trajectory Prediction," offers a novel approach to enhance the trustworthiness of prediction systems by integrating prior knowledge for interpretability and kinematic feasibility, presenting the Trustworthy Trajectory Prediction (TPK) framework.

Integration of Prior Knowledge

Traditionally, neural networks have been instrumental in trajectory prediction due to their ability to learn complex patterns from extensive datasets. However, these models often result in predictions lacking physical feasibility and interpretability, failing to sufficiently align with human reasoning. This gap in trustworthiness hinders the practical deployment of trajectory prediction models in autonomous driving systems.

The proposed approach in the paper addresses these challenges by incorporating prior knowledge models into deep learning systems. It integrates interaction and kinematic models suitable for various agent classes, including vehicles, pedestrians, and cyclists, using class-specific interaction layers. This allows for predictions that are not only physically feasible but also interpretable.

Novel Contributions

1. DG-SFM Interaction Prior: To model interactions and guide the agent-to-agent layer, the Directed-Gradient Social Force Model (DG-SFM) is introduced. It assigns rule-based interaction importance scores to neighboring agents, based on directional potentials and movement gradients, enhancing the alignment with intuitive human reasoning in traffic scenarios.
2. Kinematic Models for Feasibility: The paper introduces kinematic models suitable for ensuring physical realism across different agent types. A notable innovative addition is the pedestrian kinematic model, aligning pedestrian motion with real-world constraints.

Experimental Analysis and Results

The paper benchmarks its approach against state-of-the-art models like the Argoverse 2 dataset, employing the transformer model HPTR as a baseline. It demonstrates improvements in interaction interpretability, showing a correlation between mispredictions and deviation from interaction priors. While incorporating these models slightly decreases accuracy, they crucially eliminate the physically infeasible trajectories present in both the dataset and baseline predictions.

Implications and Future Directions

The proposed framework boosts trust in trajectory predictions by ensuring they adhere to physical laws while presenting interpretable interaction reasoning. Although adding priors results in a minor decrease in prediction accuracy, the benefits in terms of eliminating infeasible predictions outweigh this drawback.

Further exploration can focus on refining the integration methods for prior knowledge to enhance both accuracy and computational efficiency. Additionally, research can investigate extending these kinematic and interaction models to accommodate a broader range of agent types and scenarios, contributing to building more adaptable and robust autonomous systems.

The findings from this paper emphasize the importance of bridging the gap between raw predictions and their alignment with human intuition and physical realism, paving the way for more reliable trajectory prediction in the context of autonomous driving.