Traffic Data Evaluation for Automated Driving Handover Scenarios

At the current stage of automated vehicle development, the control handover from the system to a human driver (and back) is inevitable. It is essential to distinguish between situations in which the handover is possible and in which it could be dangerous and is therefore highly undesirable. We evaluated traffic situations based on two modalities: own vehicle state and traffic objects. To assess the former, supervised machine learning was applied, reaching an accuracy of 80.3% and specificity of 77.8% with Multilayer perceptron Classification. Traffic objects data were subject to different clustering techniques. The final grouping was done according to manually elaborated rules, resulting in a range of situation complexity scores. Improving the discriminative power of vehicle state classification, including driver's state and weather information, and predicting situation complexity are to be addressed in future research.