Research On Identification Of Fatigue Driving Behavior

Fatigue driving is one of the major contributing factors for traffic crashes. The driver under fatigue would have impairments in driving performance, and the vehicle's speed variation and lane position on the road would change. All these changes, called changes of driver behavior, could indicate that the driver is fatigue. This study aims at examining the changes of driver behavior under the influence of fatigue and designing the identification algorithm for driving fatigue.Firstly, in order to reveal the driver behavior feature under fatigue, some long-time driving experiments have been carried out with the use of driving simulator. The driving simulator has a five-channel, and a 150-degree forward view. A validate study had been effectuated in order to verify the driving simulator, and relative validities had been established. The road scenario developed in this study is a national expressway in ring shape. All traffic signs and road alignments were designed in terms of the requirements of JTGTB05-2004(Evaluation Guideline of Road Safety Design in China.). Driving behavior data, i.e., speed, steering angle, etc., were collected on time at 8Hz.14 participants were recruited to drive on the simulator, and the corresponding data were recorded. The experiments were carried out in four steps, and each participate was required to drive in the experiments scenario for more than 7 circles, and the data related to 5 road sections in different road alignments were employed for analysis. Drivers'statuses in this study were classified into two levels: normal and fatigue. In terms of the simple reaction performance, the driver driving in the first 4-circles were deemed as normal, while the driver driving in the following 3 circles were considered to be under fatigue.Secondly, the driver behavior data including speed, acceleration, steering angle, steering angular speed on longitudinally and laterally driving direction were analyzed using signal processing method on time and frequency domain. A total of twelve factors were found to be different between the two status, i.e., standard deviation of speed, absolute average, standard deviation and range of acceleration, energy of acceleration, wavelet energy of scale cdl of acceleration, wavelet entropy of acceleration, absolute average and standard deviation of steering angle, energy of steering angle, wavelet energy of scale cdl and ca5 of steering angle, wavelet entropy of steering angle, absolute average and standard deviation of steering angular speed, energy of steering angular speed and wavelet energy of scale cdl of steering angular speed. An ANOVA method was used to identify the significant differences between the above factors on different road alignments, and seven factors were not found to be significantly different, i.e., standard deviation of speed, wavelet energy of scale cdl and wavelet entropy of acceleration, wavelet energy of scale ca5 and wavelet entropy of steering angle, standard deviation and energy of steering angular speed. These seven factors were then aggregated into feature vectors of identification.Thirdly, on the basis of the feature vectors, the driver fatigue features were abstracted by principal component analysis (PCA). Fuzzy cluster method (FCM) and three artificial neural network (ANN) models, BP-ANN, RBF-ANN, PNN, were used to identify driver's status. The results indicated thatâ‘ the accuracy of fuzzy cluster analysis based on the feature of the first component was better than the use of other feature grouping, and the accuracy rate is over 80%;â‘¡the BP-ANN model had the best performance in identification among all the methods, however the speed for training model was relatively slow, and the model would be sunk into the local minimum which resulted in misconvergence of the model. As such, it could not satisfy the requirement of real time;â‘¢the convergence speed of RBF-ANN was satisfied; however, its accuracy was only 70% which is not in a satisfied condition. The PNN which take into account the competition unit to the output layer of RBF-ANN could overcome the defect of RBF in classification, and the identification accuracy of PNN could amount to 90%. This study suggested that if the PNN could meet the requirement of real time, it would be the best method for driver fatigue detection, but in the restricted real time condition, the FCM was also a alternative choice.Finally, a proof of identification test had been done. The accuracy rate of FCM was 81% while the PNN was 92%. The accuracies of both the methods were accepted.