Concerns regarding the safety of partial automated vehicles (AVs) remain prevalent, especially in complex roadway environments. AVs incorporate Advanced Driving Assistance Systems to perform the dynamic driving task; however, unexpected disengagements of the systems still occur due to various contextual and infrastructural factors. Among these, two-lane rural roads—with their geometric design and operational challenges—represent a critical setting. Notably, a high number of disengagements are concentrated on horizontal curves.
This study analyzes naturalistic disengagement data from two SAE Level 2 AVs operating on different segments of two-lane rural roads. The horizontal curves were characterized across geometrical and operational variables (radii, curvature change rate (CCR), curve direction, speed or visibility). These variables served as inputs to an artificial neural network (ANN) model designed to predict disengagement occurrences.
The ANN, a feedforward multilayer perceptron with one hidden layer, was trained using backpropagation. Performance was validated with K-fold cross-validation, and accuracy assessed via cross-entropy loss and confusion matrices. A Monte Carlo-style simulation tested robustness by generating multiple confusion matrices from randomized data partitions to evaluate classification stability. The results highlight CCR and lane width as key predictive factors. The calibrated ANN demonstrated robust classification (accuracy = 87.8 %, sensitivity = 92.7 %, specificity = 85.9 %) in identifying curve segments with a higher likelihood of disengagement.
This study provides road administrations a new neural network derived empirical formula to identify potential AV disengagement zones. By identifying risk-prone areas, authorities can consider targeted measures—such as enhanced signage or driver alerts— to support safer and more efficient automated driving in rural settings.
扫码关注我们
求助内容:
应助结果提醒方式:
