Feedforward Mutual-Information Anomaly Detection: Application to Autonomous Vehicles

This paper describes a mutual-information (MI)-based approach that exploits a dynamics model to quantify and detect anomalies for applications such as autonomous vehicles. First, the MI is utilized to quantify the level of uncertainty associated with the driving behaviors of a vehicle. The MI approach handles novel anomalies without the need for data-intensive training; and the metric readily applies to multivariate datasets for improved robustness compared to, e.g., monitoring vehicle tracking error. Second, to further improve the response time of anomaly detection, current and past measurements are combined with a predictive component that utilizes the vehicle dynamics model. This approach compensates for the lag in the anomaly detection process compared to strictly using current and past measurements. Finally, three different MI-based strategies are described and compared experimentally: anomaly detection using MI with (1) current and past measurements (reaction), (2) current and future information (prediction), and (3) a combination of past and future information (reaction–prediction) with three different time windows. The experiments demonstrate quantification and detection of anomalies in three driving situations: (1) veering off the road, (2) driving on the wrong side of the road, and (3) swerving within a lane. Results show that by anticipating the movements of the vehicle, the quality and response time of the anomaly detection are more favorable for decision-making while not raising false alarms compared to just using current and past measurements.