At present, particle filter is widely used in underwater physical field auxiliary navigation. Aiming at the defect of particle degradation caused by long-term particle filtering, an improved particle filtering method is proposed in this paper. The innovation of this method is that the navigation positioning error and terrain measurement error are estimated in real time in the filtering, and each particle resampling strategy is adjusted according to the estimation result. The newborn particles are randomly distributed in the pool of error while reflecting the probability density of the carrier being located in the pool of error. Using the algorithm for underwater terrain-assisted inertial navigation simulation, compared with the traditional filtering method, the positional error can be quickly reduced, meanwhile the method can avoid the filter divergence caused by particle degradation, and the final filtering accuracy is improved. By this method, the positioning error accumulated by long-term inertial navigation is reduced from 900 m to 50 m within 500 s. This method will have broad application prospects in the field of underwater multi-physics assisted navigation.