Study On Driver Fatigue Alertness Based On Bayesian Network

Driver fatigue is the major reason of severe traffic accidents, which in recent years has become the focus of traffic safety. Current research focuses on visual monitoring based on driver’s state of mind,which uses visual techniques to monitor the driver’s eyes, mouth and head characteristics without contacting with the driver’s body, and the features it detects are intuitive, but this method is limited to a specific conditions and only studies one specific feature, also the driver’s individual differences is not considered in this method; in addition, fatigue is not directly observable and it can only be inferred from some certain characteristics, in current there is no acknowledged evaluation criterion for the fatigue, it is hard to choose the accurate threshold by single feature, so a fatigue model which systematically account for various features is necessary. Warning accuracy can be improved by integrating multiple features to suit different personal circumstances, avoiding judicial errors due to only depending on a single feature to monitor.Therefore, it will become the future trend of research in this area toimprove the reliability of driver’s fatigue state monitoring based on multi-features fusion algorithm.Driver fatigue state monitoring is studied in this paper based on Bayesian network theories, which are as follows:1. Various causes and features of driver fatigue are systematically analyzed to build driver’s fatigue state assessment model, then the topology of this model is built using causal relationship based on the above variables as the model nodes. For some nodes in this model, their parameters can be provided by experts’ experience. The sample databases of this model are built through experiments, and the parameters of the remaining nodes are got by using the maximum likelihood estimation algorithm of parameter learning.2. In order to obtain the probability of driver’s fatigue variable state with the given evidence, driver’s fatigue state assessment model is inferred based on the junction tree algorithm. Firstly, translating this model into a junction tree. then, initializing the junction tree. Finally, calculating the probability through the message passing defined in the junction tree.3. Experiments were carried out to verify the accuracy of driver fatigue state assessment model, the correlation between the driver fatigue variables and each parameter was analyzed based on the experiment data. Finally, we make some analysis based on the network we have built, which shows that it can effectively solve the false alarm problems caused by false detection of the parameters with a single feature and it has a better robustness. At the same time, PERCLOS is proved to be an important show feature in driver fatigue assessment, it can provides an important theoretical reference for the research about driver fatigue device in the future.