There are many sources of error in airborne photoelectric pod, one of which is friction torque. When the optical axis is at low speed or rotating back and forth, its friction torque will change back and forth between static friction and sliding friction, which is difficult to predict. Using identification method to establish friction model and compensate it is a method to improve the accuracy of two-axis and two-frame optical axis. This paper explains the modeling process of friction torque in detail and compares the fitting effects of three models. Coulomb+viscosity model lacks the description of static friction. In the static friction+coulomb+viscosity model, the static friction to dynamic friction is a jump, so it is difficult to estimate and compensate the friction. The LuGre model not only has the relevant description of static friction, but also has a good fitting property because its curve changes slowly, which is more close to the characteristics of the contact surface changing back and forth between static friction and dynamic friction at low speed.