Search papers, labs, and topics across Lattice.
This paper addresses the problem of out-of-distribution (OOD) monitoring for trajectory prediction in safety-critical Cyber-Physical Systems by formulating it as a quickest changepoint detection (QCD) problem. They model the in-distribution (ID) prediction error evolution using a Hidden Markov Model (HMM) and extend the cumulative Maximum Mean Discrepancy (MMD) approach to detect OOD events without needing explicit knowledge of the post-change distribution. Experiments on real-world driving datasets demonstrate reduced detection delay and robustness to heavy-tailed errors while maintaining provable guarantees on detection delay and false alarm rates.
By modeling trajectory prediction errors as a Hidden Markov Model, this work achieves faster and more robust out-of-distribution detection in driving scenarios, even with heavy-tailed errors and unknown distributional shifts.
In safety-critical Cyber-Physical Systems (CPS), accurate trajectory prediction provides vital guidance for downstream planning and control, yet although deep learning models achieve high-fidelity forecasts on validation data, their reliability degrades under out-of-distribution (OOD) scenarios caused by environmental uncertainty or rare traffic behaviors in real-world deployment; detecting such OOD events is challenging due to evolving traffic conditions and changing interaction patterns, while safety-critical applications demand formal guarantees on detection delay and false-alarm rates, motivating us-following recent work [1]-to formulate OOD monitoring for trajectory prediction as a quickest changepoint detection (QCD) problem that offers a principled statistical framework with established theory; we further observe that the real-world evolution of prediction errors under in-distribution (ID) conditions can be effectively modeled by a Hidden Markov Model (HMM), and by leveraging this structure we extend the cumulative Maximum Mean Discrepancy approach to enable detection without requiring explicit knowledge of the post-change distribution while still admitting provable guarantees on delay and false alarms, with experiments on three real-world driving datasets demonstrating reduced detection delay and robustness to heavy-tailed errors and unknown post-change conditions.
