Domain Invariant Driving Behaviour Prediction Based on Autoencoder Anomaly Detection
Sugandh Pargal, Sandip Chakraborty, Shohreh Deldari, Salil S. Kanhere · 2024
Driving behavior prediction has become increasingly popular and challenging in developing advanced driver assistance systems (ADAS) systems. The main challenge lies with the varying domains and surroundings. Current ADAS systems fail to predict when a driver drives in a different environment, often leading to accidents. To mitigate these challenges, one possible approach is to understand the changes observed in the regular driving behavior of the driver. This paper presents a model for predicting driving behavior across different domains using autoencoder anomaly detection and, thereby, defines a metric to compute the driving behavior score. The experimental setup includes a pilot study for feature selection and a semi-controlled experiment for data collection across multiple countries. The findings highlight the significant role of in-vehicle driving features and physiological features of the driver in predicting domain-invariant driving behavior based on contextual information.