Abstract: The influence of the presupposed variance of the measured error on the distribution of weighed coefficient in multi-sensor fusion is analyzed. A new improved multi-sensor weighting and filtering algorithm which is the self-learning of the variance of the measured error is presented. This new algorithm can not only sufficiently utilize renewed information each time from sensor to optimize the variance of the measured error step by step, but also reasonably distributes weighted coefficients to improve the state estimation. Stimulation shows this algorithm can improve significantly the efficiency of maneuvering target tracking.