Scenario-based Evaluation of Prediction Models for Automated Vehicles

On Adversarial Robustness of Trajectory Prediction for Autonomous Vehicles

“…The process involves capturing a sequence of observed states for each agent at a fixed time interval ∆t and generating the projected future trajectory for each agent. Some models produce a single predicted trajectory, while others generate a set of feasible trajectories along with associated confidence levels [17].…”

“…To assess the robustness of a trajectory prediction model in specific settings, the test data can be modified [17]. We distinguish between two general strategies of evaluating robustness through data modification (Figure 1).…”

“…It includes modifying the training data or augmenting it with perturbations (Figure 1). These perturbations can be incorporated into the historical states X [7], [16], [18], [23], [38] and optionally into the future states Y [14], [24], [39], denoted as X and Ỹ . The objective is for the trained model to exhibit comparable behavior for perturbed and benign examples [40].…”

“…1) Scenario importance sampling: To operate safely, an automated vehicle must anticipate the evolving environment around it. To achieve this, it is crucial to determine the most suitable prediction models for various situations [17]. Addressing this challenge involves employing scenario extraction, which requires a substantial amount of data to identify common and uncommon scenarios [35].…”

“…As previously mentioned, an method to evaluate and improve model robustness against unseen situations involves collecting more data that includes the missing safety-critical scenarios, which is costly, time-consuming, and lacks scalability [38]. The necessity for additional data arises due to the prevalent issue of current prediction models being frequently overfitted on limited patterns within datasets and not capturing the long-tailed distribution of driving scenarios and behaviors [53].…”