In this paper,a new fatigue detection method based on cascade convolutional neur al network (CNN) structure is proposed.Firstly,for learning the eye features,the first-level network is trained to classify the eye and non-ey e regions.The eye region is extracted by the feature maps of the first-level network.Then,the eye images are sent to the nex t network for the eye feature point detection.The eye opening degree is computed to estimate the eye states o f the subject and construct the model of fatigue detection.Finally,whether fatigue driving or not is judged by t he time series of eye states in multi-consecutive frames.At the detection error of 5%,the average detection acc uracy of eye feature points is 93.10% and the highest detection accuracy of a single point is 97.14%.The experi mental results of our method show that the detected eye states based on the proposed method are obviously dif ferent in awake and fatigue states,which proves the proposed method is effective and has a better application prospect.