We present a multi-stage classification approach to detect targets in widely varying thermal imagery. A multi-level spatial-temporal median filter is utilized to extract the background frame, with which the background clutters are suppressed by using the principal component analysis technique. A spatially related fuzzy adaptive resonance theory (ART) neural network is then applied to identify the local regions-of-interest. Within each region, another fuzzy ART neural network is utilized to detect the targets. Experimental results demonstrate that the proposed approach is capable of detecting infrared moving targets effectively for F1 measurement up to 96.3%.