Abstract: During the sowing and harvesting season, the agricultural machinery drivers often work day and night regardless of fatigue, making them extremely prone to safety accidents. At present, the detection methods for fatigue driving are mainly divided into two types: contact-type（collecting agricultural machinery driver body characteristic information） and non-contact type（photographing agricultural machinery driver face to detect fatigue state）. Aiming at the problems of inconvenience and inaccuracy of detection in the two methods, this paper proposes a face detection method using Haar-like feature and AdaBoost algorithm to train multiple weak classifiers. Based on the convolution neural network model, the face of the agricultural machinery driver is divided into the probability value of closed eyes, normal, and non-face. Fatigue can be detected by the PERCLOS method. The simulation results on the embedded platform show that the accuracy of the AdaBoost + CNN + PERCLOS method can reach 95.1%.