Analysis of the Potential of Divergence Metrics for Trajectory Predictions to Identify Prediction Errors

Modern vehicles are increasingly being equipped with Advanced Driver Assistance Systems (ADAS) and active safety features, such as Advanced Emergency Braking Systems (AEBS), designed to improve road safety. While AEBS has been shown to significantly reduce collision rates, testing and validation pose significant challenges. Extensive field tests are required to validate the system’s safety and reliability, which are resource-intensive and time-consuming. To address this, the TarA project (Testaufwandsreduktion AEB), conducted at the Institute of Automotive Engineering (FZD) at Technical University of Darmstadt in cooperation with and funded by the Continental Autonomous Mobility Germany GmbH, seeks to identify methodologies to reduce testing efforts. One of the explored methods is the Prediction Divergence Principle (PDP), which classifies predictions and system activations as correct or incorrect based on the prediction divergence. This principle includes various approaches, one of which is examined in this thesis, hypothesizing that anomalous trajectory divergences between predicted trajectories and the pseudo-ground truth path correlate with False Positive (FP) AEBS system activations. To evaluate the underlying hypothesis, this thesis aims to develop a system that provides an initial indication of whether the hypothesis holds or should be reconsidered. For this, appropriate divergence metrics, machine learning models, and anomaly detectors are researched to determine suitability for the given data structure. The selected methodology involves training a machine learning model on field test data that learns normal trajectory divergences. This model takes the initial situational parameters between the ego vehicle and the target object as input, with the trajectory divergences between predicted and pseudo-ground truth trajectories serving as the target values for learning. The model’s estimation errors form the basis for identifying instances where the actual divergence notably exceeds these estimations, signaling anomalous trajectory divergences. This method is applied to a labeled dataset, with the classification results analyzed for correlations with the provided labels. The main achievement of this research is the establishment of a modular and adaptive analytical framework capable of integrating different divergence metrics, situational parameters and machine learning algorithms. This framework supports future exploration, allowing the incorporation of new metrics and models seamlessly. Additionally, an implementation of a synthetic data generator provides a controlled environment for model verification, offering potential for expanding the analysis with more complex error patterns. Despite these contributions, the findings indicate that the initial hypothesis cannot be fully confirmed based on the selected divergence metrics and machine learning model used in this work. The analysis reveals limitations in the correlation between anomalous trajectory divergences and FP system activations. Nevertheless, the results point towards a potential new direction: anomalous trajectory divergences of the ego vehicle correlate with True Positive (TP) system activations. A preliminary analysis in this direction has been conducted, which demonstrates higher correlations in 2 out of 3 cases. This alternative hypothesis builds on the observation that driver interventions in hazardous situations lead to significant deviations from the predicted trajectories.