The normal image features are easily impacted by factors,such as varia nt illumination,head motion and wearing glasses,which lead to the low accuracy in various eye state recognition algorithms. To improve recognition accuracy,various infrared eye image features are analyz ed in this paper and the following features are found to be propitious when used in eye state classi fiers:the pseudo-zernike feature is invariant for rotation and scale changing but sensiti ve to variant illumination;the complexity feature is simple and effective but re quires a high-precision contour extraction;the HOG feature is robust to variant illumination but cannot tolerat e a wide range of head motion.Based on these analyses,an eye state recognition algorithm based on mul ti-feature decision fusion is proposed.Firstly,corresponding optimal support vector machine (SVM) mo dels for these three features are built; then weights for decision of three classifiers are calculated through the autom atic weight learning algorithm;lastly,the final recognition results are obtained through th e decision fusion of three classifiers′ recognition results.The robustness of eye state recognition algor ithm is improved by combining the performance characteristics of different features.Experimental res ults show that the accuracy of eye state recognition algorithm can achieve 91.9% in this paper,whi ch overcomes the impacts of variant illumination and head motion.