Research On Fatigue State Recognition Technology Of Train Driver

With the rapid development of China’s railway transportation system,railway operation safety has been brought to people’s attention.Fatigue driving is one of the major causes leading to railway accidents,the fatigue status detection of driver,however,is a weak part in the railway operation safety,hence it is of great significance to conduct effective detection on driver’s fatigue status.In this paper,train driver serves as the object of study,research is carried out on driver’s fatigue status detection via deep learning and automatic machine learning by adopting computer vision,and the approach come up with in this paper is proved effective.The main study of this paper is shown as follows:1.In most facial landmark detection approaches based on deep learning currently,only the features mapped in the last convolutional layer of CNN is applied to facial parts detection,ignoring the local features extracted by the front end of CNN.To solve this problem,a landmark and head pose detection approach based on multi-scale and multi-tasking CNN is proposed in this paper.By fusing the features of different scales extracted from different layers of CNN,the features obtained can better express the information of human face in the image,then multi-tasking learning method is employed to return landmark and head pose.The experiment reveals that the approach has high precision and better robustness,lay a foundation for future fatigue detection of train driver.2.Against the context that current fatigue detection algorithm has single characteristics,poor robustness,and manual algorithm selection and hyper-parameter configuration are commonly involved,a multi-feature fusion fatigue detection approach based on TPOT is studied in this paper.First,common characteristics of driver fatigue based on the vision are studied,then those applicable to train driver are selected,finally,driver’s fatigue characteristics are quantified and extracted according to landmark and head pose.In later modeling indicating the relationship between characteristics and fatigue,automatic machine learning TPOT is applied.It is able to generate random tree structure process in which the algorithm of each node can be matched randomly and evolved via genetic programming,obtaining an ideal machine learning process.It is proved in the experiment that the approach has good effect on fatigue detection.3.As conventional fatigue detection method based on feature extraction is susceptible to driving environment and individual differences of drivers,a multi-modal fusion fatigue detection approach is proposed in this paper.Modal data from driver’s landmark and facial image are input to two fatigue detection networks corresponding to landmark and image respectively.For landmark-based network,the coordinate of key parts is processed as 1D data and sent to LSTM network for modeling;for image-based network,static feature extraction is carried out on several facial areas relying on strong feature extraction ability of CNN,then sent to LSTM for time series modeling,at last,decision-making level of two networks are fused by introducing multi-modal fusion to further improve network effect.The approach is proved to have good fatigue detection effect by the experiment.