As one of the background estimation algorithms for Infrared (IR) point target detection, Least Squares (LS) method has a poor performance to the complex nonlinear background. A nonlinear version of the least squares algorithm, called Kernel Least Squares (KLS) is deduced by using Kernel Methods (KMs). Furthermore, the exponential weighted form of KLS, called KEWLS, is deduced. KEWLS is more adaptive to dynamic nonlinear system’s time-series prediction. A kernel-based IR target detection algorithm is proposed, image background is estimated by KEWLS nonlinear regression, and then target is detected by self-adaptive threshold detection in the difference image. It is shown by nonlinear function regression and sequence IR images detection experiments that the kernel methods improve the performance of nonlinear function regression and IR background estimation.