When the adaptive centroid tracking algorithm is used to track an extended target, object missing may appear. For this reason, a synthesis predictor which consists of a three-point linear predictor and a five-point square predictor is proposed and added to the adaptive centroid tracking algorithm to predict the location parameters of the tracked target. And the histogram distribution function is computed by estimating the proportion of background in the object window. Then, target pixels can be separated from background image by employing Bayias decision to classify target pixels and background pixels of image in the window, and the centroid coordinate values can be obtained. The improved algorithm can be used to realize the tracking from the small target to the extended target availably with high stability, reliability and precision, which will not be affected basically by the target changing of scale and rotation. The experiments also show that by using the improved algorithm, the problem of target missing that appears when using the original algorithm to track an extended target can be overcome, thus the availability range of the adaptive centroid tracking algorithm is extended, and its tracking precision is improved. The algorithm can provides a useful reference for improving attacking precision of missile guiding at last stage.