Predicting Unintentional Traffic Lane Departures Using Neural Networks

ABSTRACT PREDICTION UNINTENTIONAL LANE DEPARTURES BASED ON NEURAL NETWORKS by JAMAA AMBARAK May 2018 Advisor: Dr. Hao Ying Major: Electrical Engineering Degree: Doctor of Philosophy Advanced driving assistant systems can potentially reduce unintentional lane departure traffic accidents by predicting/detecting driving situations and alerting drivers to avoid or mitigate traffic accidents. These systems could potentially reduce or mitigate the severity of crashes by providing a warning to assist the driver with considerable time to steer the vehicle from departing the lane boundary. In this dissertation, we explored the effectiveness of the three-layer perceptron neural network and the dynamical recurrent neural network in predicting unintentional lane departures. To verify the effectiveness of the networks, the recurrent neural network demonstrated high performance compared with the conventional static perceptron neural network for unintentional lane departure predictions. To the best of our knowledge, this method is innovative and no comparison has been attempted for this purpose in literature. Furthermore, the networks provide comparative prediction accuracy and robustly outperform a time to lane crossing method. Moreover, we developed a learning-based training scheme to address the number of false positive prediction errors and enhance the predicting performance of the networks. The target of the learning-based training scheme is to determine to what extent the number of false positives can be reduced. Experiment data driver were generated by VIRTTEX, a hydraulically powered 6-degrees-of-freedom moving base driving simulator at Ford Motor Company. The experimental data represented 16 drowsy drivers who drove a simulated 2000 Volvo S80 (three hours per driver), which consisted of a total of 3,508 lane departure occurrences. Two-thirds of the lane departures were randomly selected to generate training examples for the networks and the remaining were used for testing. The number of hidden neurons, as well as the input vehicle variables, were optimized experimentally through the training process. The networks were optimized experimentally through the training process and then used to predict lane departure by processing the entire driving time series of the 16 drivers one by one after all the training data was removed from the time series. The networks made a prediction at each sampling moment of the time series and there were over 6.3 million prediction decisions. The experimental results demonstrate that the learning-based training scheme can be an effective solution to reduce the number of false positives and improve the forecasting performance with minimum prediction errors