Prior Knowledge Integration for Feasible and Interpretable Trajectory Prediction

Trajectory prediction is crucial for autonomous driving to navigate safely by anticipating the movements of surrounding road users. However, current deep learning models often lack alignment with human reasoning, leading to predictions that are physically infeasible or are not logical to a human. To address this challenge, recent research has incorporated prior expert knowledge, like the social force model for modeling interactions and kinematic models for physical realism.
However, these approaches do not evaluate the impact of the added expert knowledge on the network's interpretability and mainly focus on priors that either suit vehicles or pedestrians but do not generalize to all agent types. In contrast, we propose incorporating priors for interactions and kinematics of all agent classes -- specifically vehicles, pedestrians, and cyclists -- with class-specific layers to capture the differences in the rationale for each class. To improve interaction interpretability, we guide an agent-to-agent attention layer with rule-based interaction importance scores. We compare two prior models for calculating these scores and two integration methods for guiding attention. ... mehrTo ensure kinematic feasibility, we let the network predict control inputs, from which kinematic models then compute trajectories. We benchmark our proposed modifications on the Argoverse 2 Motion Forecasting dataset, using the state-of-the-art deep neural network HPTR as our baseline. Our experiments demonstrate that we improve interpretability without degrading prediction accuracy, finding a correlation between incorrect predictions and a divergence of the model's attention from our interaction prior. Even though incorporating the kinematic models causes a slight decrease in accuracy, they more than compensate by eliminating infeasible motions present in both the dataset and particularly in the baseline model's predictions, where up to 88 % of predictions are impossible in reality. Thus, our approach ensures that predicted trajectories are both interpretable and kinematically feasible, contributing to safer and more reliable trajectory prediction.