Robust Detection of Anomalous Driving Behavior

Abstract

Driving behavior is a major factor in traffic safety applications. Abnormal behavior, such as extremely aggressive or passive driving, can endanger both the driver and other traffic participants. Most driving behavior analysis approaches to date rely on classification, which requires labeled data for both normal driving and various types of anomalous behavior. We propose an approach that detects anomalous driving patterns based on outlier detection, which does not require such data. Apart from the required data set, existing approaches have difficulties dealing with changing behavior that overlaps with normal behavior (e.g., aggressive drivers still stop in traffic jams). We introduce a post-processing step that significantly improves results in this regard. The approach is evaluated using simulations based on the realistic LuST traffic scenario, and shows reliable detection of anomalous vehicles with low false positive rates.