DoAssist: Domain Invariant Driving Anomaly Detection based on Spatiotemporal Driving Data

Driving anomaly detection is crucial for developing Advanced Driver Assistance Systems (ADAS). Traditional methods for driving anomaly detection rely on rule-based and domain adaptation to account for varying driving patterns, often based on human-labeled data. However, these methods may be susceptible to the annotation process and struggle with diverse driving environments. This paper addresses the importance of analyzing driving anomalies personalized to drivers within their specific context. We propose DoAssist, a novel method that leverages spatiotemporal multimodal data to detect driving anomalies in real-time by utilizing unsupervised learning models. DoAssist also explores the relationship between anomalies and different modalities to develop an alert-based system for drivers. DoAssist is validated on diverse datasets across two countries, achieving an average AUC-PR of 86.7%. Furthermore, we demonstrate that DoAssist is robust to variations in road type, traffic and weather and performs better than various other baselines.