Efficient Spatio-Temporal Out-of-Distribution Detection for Autonomous Systems

Learning-enabled components (LECs) are increasingly deployed in autonomous systems to provide adaptive and intelligent decision making in a complex dynamic environment. However, most LECs are typically data-driven, and they may exhibit erroneous predictions impacting system safety when there are some distribution shifts between real testing scenarios and training scenarios. In this paper, we introduce an efficient spatio-temporal Out-of-Distribution (OOD) detection scheme for time-series data to improve the detection performance and robustness over the state-of-the-art. Specifically, the optical flow method is employed to extract the spatio-temporal features in the high-dimensional time series data, and the extracted spatio-temporal optical flow features are encoded with the variational autoencoder to obtain input manifold for the OOD detection. Moreover, we proposed a martingale-based inductive conformal anomaly detection method to achieve lower bounded false alarms owing to the statistical foundation and robustness characteristic of martingale. The experimental evaluation demonstrates that our proposed approach significantly outperforms the existing methods in terms of detection capability and robustness.