An accurate method is proposed for horizon detection according to the imaging characteristics of infrared images. Unlike most previous sea-sky-line detection algorithms, the method employs linear Harr features to gain edges after infrared image de-noising, thus to fix the position of the sky and the quasi horizon. Then, the position of the horizon is detected roughly via analyzing the energy of the margin. Finally, the accurate and complete horizon is acquired by means of the numerical analysis techniques. The experiment results show that the method can detect the horizon effectively in different scenes, and it is robust even if the situation is complex when the sky is overcast with cloud and mist.