An improved IR target tracking algorithm based on Mean Shift is proposed combined the mean-shift based gradient matched searching strategy with the feature-classification based tracking algorithm.An improved target representing model is set up by taking the likelihood ratio of gray level features of a target and a local background as a weighted value of the original kernel histogram of target area.The expression of mean-shift vector in this target model is deduced,when Bhattacharyya coefficients are regarded as the similarity measures.Meanwhile, the criterion of model updating based on tracking complexity estimation under target occlusion is presented.The experimental result indicates that the algorithm can improve the shift weight of target pixel gray level and can suppress the background interference, therefore the tracking performance of the low contrast IR target is robust and the average Bhattacharyya coefficients can keep above 0.97 in a correct tracking case.