Abstract: Standard Kalman filter algorithm assumes system mathematical model and statistical noise characteristics；it easily leads to errors and even divergence when the assumptive and actual models do not match.Proposed Kalman filter adaptive algorithm based on reverse prediction，by comparing the original state normalized innovation square and reverse predicted state normalized innovation square，corrects predicted state on-line by noise model adjustment when the normalized innovation square ratio is greater than the threshold.Radar target tracking simulation results show that the algorithm can improve filtering accuracy and robustness when target maneuvers and noise increases.