Viability to predict driver intentions in unmarked roundabouts using vehicle and contextual information

As the demand for advanced driver assistance systems (ADAS) and autonomous driving technologies increases, accurate prediction of driver intentions in complex scenarios such as unmarked roundabouts becomes crucial. This study presents a predictive model based on Long Short-Term Memory (LSTM) neural networks that integrates kinematic variables, vehicle positioning, and road texture features to anticipate whether a driver intends to turn right, turn left, or continue straight when approaching a roundabout. Experimental evaluation compared classical and deep learning predictors including Logistic Regression, Decision Tree, CNN, RNN, Hybrid CNN-RNN, and LSTM, under two feature configurations. With 9 features, all models showed limited predictive capacity, with accuracies ranging from 0.37 to 0.57 and AUC values below 0.77. With 10 features, performance improved substantially: Logistic Regression achieved 0.82 accuracy (AUC 0.93), Decision Tree 0.97 (AUC 0.98), and deep networks 0.93–0.96 accuracy with AUC values above 0.99. The LSTM model reached 0.93 accuracy and 0.99 AUC, maintaining consistent performance under degraded conditions when only nine features were available. These results demonstrate the contribution of contextual road texture analysis to maneuver intention classification and highlight the potential of LSTM-based approaches for real-world ADAS deployment. Future work will extend validation to multiple drivers, diverse roundabouts, and simulation-based scenarios to enhance generalizability and robustness.